chat: cut v3.5.0-rc1 release candidate (#3218)

Signed-off-by: Jared Van Bortel <jared@nomic.ai>

.circleci/config.yml

@@ -1,7 +1,7 @@
 version: 2.1
 setup: true
 orbs:
-  path-filtering: circleci/path-filtering@0.0.1
+  path-filtering: circleci/path-filtering@1.1.0
 
 workflows:
   version: 2.1

.circleci/continue_config.yml

@@ -19,6 +19,13 @@ parameters:
     default: false
 
 jobs:
+  # work around CircleCI-Public/path-filtering-orb#20
+  noop:
+    docker:
+      - image: cimg/base:current
+    steps:
+      - run: "true"
+
   build-offline-chat-installer-macos:
     macos:
       xcode: 15.4.0
@@ -37,11 +44,11 @@ jobs:
           command: softwareupdate --install-rosetta --agree-to-license  # needed for QtIFW
       - run:
           name: Install dependencies
-          command: brew install ccache
+          command: brew install ccache wget
       - run:
           name: Installing Qt
           command: |
-            curl -o qt-unified-macOS-x64-4.6.0-online.dmg https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg
+            wget "https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg"
             hdiutil attach qt-unified-macOS-x64-4.6.0-online.dmg
             /Volumes/qt-unified-macOS-x64-4.6.0-online/qt-unified-macOS-x64-4.6.0-online.app/Contents/MacOS/qt-unified-macOS-x64-4.6.0-online --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
             hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
@@ -175,11 +182,11 @@ jobs:
           command: softwareupdate --install-rosetta --agree-to-license  # needed for QtIFW
       - run:
           name: Install dependencies
-          command: brew install ccache
+          command: brew install ccache wget
       - run:
           name: Installing Qt
           command: |
-            curl -o qt-unified-macOS-x64-4.6.0-online.dmg https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg
+            wget "https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg"
             hdiutil attach qt-unified-macOS-x64-4.6.0-online.dmg
             /Volumes/qt-unified-macOS-x64-4.6.0-online/qt-unified-macOS-x64-4.6.0-online.app/Contents/MacOS/qt-unified-macOS-x64-4.6.0-online --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
             hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
@@ -312,9 +319,9 @@ jobs:
       - run:
           name: Setup Linux and Dependencies
           command: |
-            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
-            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            wget -qO- "https://packages.lunarg.com/lunarg-signing-key-pub.asc" | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
+            wget -qO- "https://packages.lunarg.com/vulkan/1.3.290/lunarg-vulkan-1.3.290-jammy.list" | sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.3.290-jammy.list
+            wget "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb"
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
               bison build-essential ccache cuda-compiler-11-8 flex g++-12 gperf libcublas-dev-11-8 libfontconfig1
@@ -329,7 +336,7 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            wget https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run
+            wget "https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run"
             chmod +x qt-unified-linux-x64-4.6.0-online.run
             ./qt-unified-linux-x64-4.6.0-online.run --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
       - run:
@@ -388,9 +395,9 @@ jobs:
       - run:
           name: Setup Linux and Dependencies
           command: |
-            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
-            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            wget -qO- "https://packages.lunarg.com/lunarg-signing-key-pub.asc" | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
+            wget -qO- "https://packages.lunarg.com/vulkan/1.3.290/lunarg-vulkan-1.3.290-jammy.list" | sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.3.290-jammy.list
+            wget "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb"
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
               bison build-essential ccache cuda-compiler-11-8 flex g++-12 gperf libcublas-dev-11-8 libfontconfig1
@@ -405,7 +412,7 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            wget https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run
+            wget "https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run"
             chmod +x qt-unified-linux-x64-4.6.0-online.run
             ./qt-unified-linux-x64-4.6.0-online.run --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
       - run:
@@ -466,21 +473,21 @@ jobs:
             - ccache-gpt4all-win-amd64-
       - run:
           name: Install dependencies
-          command: choco install -y ccache
+          command: choco install -y ccache wget
       - run:
           name: Installing Qt
           command: |
-            Invoke-WebRequest -Uri https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe -OutFile qt-unified-windows-x64-4.6.0-online.exe
+            wget.exe "https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe"
             & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
       - run:
           name: Install VulkanSDK
           command: |
-            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            wget.exe "https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe"
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
       - run:
           name: Install CUDA Toolkit
           command: |
-            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe -OutFile cuda_11.8.0_windows_network.exe
+            wget.exe "https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe"
             .\cuda_11.8.0_windows_network.exe -s cudart_11.8 nvcc_11.8 cublas_11.8 cublas_dev_11.8
       - run:
           name: "Install Dotnet 8"
@@ -488,7 +495,7 @@ jobs:
             mkdir dotnet
             cd dotnet
             $dotnet_url="https://download.visualstudio.microsoft.com/download/pr/5af098e1-e433-4fda-84af-3f54fd27c108/6bd1c6e48e64e64871957289023ca590/dotnet-sdk-8.0.302-win-x64.zip"
-            Invoke-WebRequest -Uri $dotnet_url -Outfile dotnet-sdk-8.0.302-win-x64.zip
+            wget.exe "$dotnet_url"
             Expand-Archive -LiteralPath .\dotnet-sdk-8.0.302-win-x64.zip
             $Env:DOTNET_ROOT="$($(Get-Location).Path)\dotnet-sdk-8.0.302-win-x64"
             $Env:PATH="$Env:DOTNET_ROOT;$Env:PATH"
@@ -549,13 +556,16 @@ jobs:
       - checkout
       - attach_workspace:
           at: build
+      - run:
+          name: Install dependencies
+          command: choco install -y wget
       - run:
           name: "Install Dotnet 8 && Azure Sign Tool"
           command: |
             mkdir dotnet
             cd dotnet
             $dotnet_url="https://download.visualstudio.microsoft.com/download/pr/5af098e1-e433-4fda-84af-3f54fd27c108/6bd1c6e48e64e64871957289023ca590/dotnet-sdk-8.0.302-win-x64.zip"
-            Invoke-WebRequest -Uri $dotnet_url -Outfile dotnet-sdk-8.0.302-win-x64.zip
+            wget.exe "$dotnet_url"
             Expand-Archive -LiteralPath .\dotnet-sdk-8.0.302-win-x64.zip
             $Env:DOTNET_ROOT="$($(Get-Location).Path)\dotnet-sdk-8.0.302-win-x64"
             $Env:PATH="$Env:DOTNET_ROOT;$Env:PATH"
@@ -587,21 +597,21 @@ jobs:
             - ccache-gpt4all-win-amd64-
       - run:
           name: Install dependencies
-          command: choco install -y ccache
+          command: choco install -y ccache wget
       - run:
           name: Installing Qt
           command: |
-            Invoke-WebRequest -Uri https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe -OutFile qt-unified-windows-x64-4.6.0-online.exe
+            wget.exe "https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe"
             & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
       - run:
           name: Install VulkanSDK
           command: |
-            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            wget.exe "https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe"
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
       - run:
           name: Install CUDA Toolkit
           command: |
-            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe -OutFile cuda_11.8.0_windows_network.exe
+            wget.exe "https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe"
             .\cuda_11.8.0_windows_network.exe -s cudart_11.8 nvcc_11.8 cublas_11.8 cublas_dev_11.8
       - run:
           name: "Install Dotnet 8"
@@ -609,7 +619,7 @@ jobs:
             mkdir dotnet
             cd dotnet
             $dotnet_url="https://download.visualstudio.microsoft.com/download/pr/5af098e1-e433-4fda-84af-3f54fd27c108/6bd1c6e48e64e64871957289023ca590/dotnet-sdk-8.0.302-win-x64.zip"
-            Invoke-WebRequest -Uri $dotnet_url -Outfile dotnet-sdk-8.0.302-win-x64.zip
+            wget.exe "$dotnet_url"
             Expand-Archive -LiteralPath .\dotnet-sdk-8.0.302-win-x64.zip
             $Env:DOTNET_ROOT="$($(Get-Location).Path)\dotnet-sdk-8.0.302-win-x64"
             $Env:PATH="$Env:DOTNET_ROOT;$Env:PATH"
@@ -677,13 +687,16 @@ jobs:
       - checkout
       - attach_workspace:
           at: build
+      - run:
+          name: Install dependencies
+          command: choco install -y wget
       - run:
           name: "Install Dotnet 8"
           command: |
             mkdir dotnet
             cd dotnet
             $dotnet_url="https://download.visualstudio.microsoft.com/download/pr/5af098e1-e433-4fda-84af-3f54fd27c108/6bd1c6e48e64e64871957289023ca590/dotnet-sdk-8.0.302-win-x64.zip"
-            Invoke-WebRequest -Uri $dotnet_url -Outfile dotnet-sdk-8.0.302-win-x64.zip
+            wget.exe "$dotnet_url"
             Expand-Archive -LiteralPath .\dotnet-sdk-8.0.302-win-x64.zip
             $Env:DOTNET_ROOT="$($(Get-Location).Path)\dotnet-sdk-8.0.302-win-x64"
             $Env:PATH="$Env:DOTNET_ROOT;$Env:PATH"
@@ -719,9 +732,9 @@ jobs:
       - run:
           name: Setup Linux and Dependencies
           command: |
-            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
-            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            wget -qO- "https://packages.lunarg.com/lunarg-signing-key-pub.asc" | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
+            wget -qO- "https://packages.lunarg.com/vulkan/1.3.290/lunarg-vulkan-1.3.290-jammy.list" | sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.3.290-jammy.list
+            wget "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb"
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
               bison build-essential ccache cuda-compiler-11-8 flex g++-12 gperf libcublas-dev-11-8 libfontconfig1
@@ -736,7 +749,7 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            wget https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run
+            wget "https://gpt4all.io/ci/qt-unified-linux-x64-4.6.0-online.run"
             chmod +x qt-unified-linux-x64-4.6.0-online.run
             ./qt-unified-linux-x64-4.6.0-online.run --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
       - run:
@@ -780,21 +793,21 @@ jobs:
             - ccache-gpt4all-win-amd64-
       - run:
           name: Install dependencies
-          command: choco install -y ccache
+          command: choco install -y ccache wget
       - run:
           name: Installing Qt
           command: |
-            Invoke-WebRequest -Uri https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe -OutFile qt-unified-windows-x64-4.6.0-online.exe
+            wget.exe "https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe"
             & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
       - run:
           name: Install VulkanSDK
           command: |
-            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            wget.exe "https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe"
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
       - run:
           name: Install CUDA Toolkit
           command: |
-            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe -OutFile cuda_11.8.0_windows_network.exe
+            wget.exe "https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe"
             .\cuda_11.8.0_windows_network.exe -s cudart_11.8 nvcc_11.8 cublas_11.8 cublas_dev_11.8
       - run:
           name: Build
@@ -842,11 +855,11 @@ jobs:
           command: softwareupdate --install-rosetta --agree-to-license  # needed for QtIFW
       - run:
           name: Install dependencies
-          command: brew install ccache
+          command: brew install ccache wget
       - run:
           name: Installing Qt
           command: |
-            curl -o qt-unified-macOS-x64-4.6.0-online.dmg https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg
+            wget "https://gpt4all.io/ci/qt-unified-macOS-x64-4.6.0-online.dmg"
             hdiutil attach qt-unified-macOS-x64-4.6.0-online.dmg
             /Volumes/qt-unified-macOS-x64-4.6.0-online/qt-unified-macOS-x64-4.6.0-online.app/Contents/MacOS/qt-unified-macOS-x64-4.6.0-online --no-force-installations --no-default-installations --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations --email $QT_EMAIL --password $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
             hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
@@ -932,9 +945,9 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
-            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            wget -qO- "https://packages.lunarg.com/lunarg-signing-key-pub.asc" | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
+            wget -qO- "https://packages.lunarg.com/vulkan/1.3.290/lunarg-vulkan-1.3.290-jammy.list" | sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.3.290-jammy.list
+            wget "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb"
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
               build-essential ccache cmake cuda-compiler-11-8 g++-12 libcublas-dev-11-8 libnvidia-compute-550-server
@@ -1041,20 +1054,20 @@ jobs:
       - restore_cache:
           keys:
             - ccache-gpt4all-win-amd64-
+      - run:
+          name: Install dependencies
+          command:
+            choco install -y ccache cmake ninja wget --installargs 'ADD_CMAKE_TO_PATH=System'
       - run:
           name: Install VulkanSDK
           command: |
-            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            wget.exe "https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe"
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
       - run:
           name: Install CUDA Toolkit
           command: |
-            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe -OutFile cuda_11.8.0_windows_network.exe
+            wget.exe "https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe"
             .\cuda_11.8.0_windows_network.exe -s cudart_11.8 nvcc_11.8 cublas_11.8 cublas_dev_11.8
-      - run:
-          name: Install dependencies
-          command:
-            choco install -y ccache cmake ninja --installargs 'ADD_CMAKE_TO_PATH=System'
       - run:
           name: Install Python dependencies
           command: pip install setuptools wheel cmake
@@ -1096,7 +1109,7 @@ jobs:
           paths:
             - "*.whl"
 
-  store-and-upload-wheels:
+  publish-wheels:
     docker:
       - image: circleci/python:3.8
     steps:
@@ -1132,9 +1145,9 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
-            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+            wget -qO- "https://packages.lunarg.com/lunarg-signing-key-pub.asc" | sudo tee /etc/apt/trusted.gpg.d/lunarg.asc
+            wget -qO- "https://packages.lunarg.com/vulkan/1.3.290/lunarg-vulkan-1.3.290-jammy.list" | sudo tee /etc/apt/sources.list.d/lunarg-vulkan-1.3.290-jammy.list
+            wget "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb"
             sudo dpkg -i cuda-keyring_1.1-1_all.deb
             packages=(
               build-essential ccache cmake cuda-compiler-11-8 g++-12 libcublas-dev-11-8 libnvidia-compute-550-server
@@ -1238,20 +1251,20 @@ jobs:
       - restore_cache:
           keys:
             - ccache-gpt4all-win-amd64-
+      - run:
+          name: Install dependencies
+          command: |
+            choco install -y ccache cmake ninja wget --installargs 'ADD_CMAKE_TO_PATH=System'
       - run:
           name: Install VulkanSDK
           command: |
-            Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe -OutFile VulkanSDK-1.3.261.1-Installer.exe
+            wget.exe "https://sdk.lunarg.com/sdk/download/1.3.261.1/windows/VulkanSDK-1.3.261.1-Installer.exe"
             .\VulkanSDK-1.3.261.1-Installer.exe --accept-licenses --default-answer --confirm-command install
       - run:
           name: Install CUDA Toolkit
           command: |
-            Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe -OutFile cuda_11.8.0_windows_network.exe
+            wget.exe "https://developer.download.nvidia.com/compute/cuda/11.8.0/network_installers/cuda_11.8.0_windows_network.exe"
             .\cuda_11.8.0_windows_network.exe -s cudart_11.8 nvcc_11.8 cublas_11.8 cublas_dev_11.8
-      - run:
-          name: Install dependencies
-          command: |
-            choco install -y ccache cmake ninja --installargs 'ADD_CMAKE_TO_PATH=System'
       - run:
           name: Build Libraries
           no_output_timeout: 30m
@@ -1361,14 +1374,14 @@ jobs:
           at: /tmp/gpt4all-backend
       - run: choco install wget -y
       - run:
-         command: wget https://nodejs.org/dist/v18.16.0/node-v18.16.0-x86.msi -P C:\Users\circleci\Downloads\
-         shell: cmd.exe
-      - run: MsiExec.exe /i C:\Users\circleci\Downloads\node-v18.16.0-x86.msi /qn
+          command: |
+            wget.exe "https://nodejs.org/dist/v18.16.0/node-v18.16.0-x86.msi" -P C:\Users\circleci\Downloads\
+            MsiExec.exe /i C:\Users\circleci\Downloads\node-v18.16.0-x86.msi /qn
       - run:
           command: |
-              Start-Process powershell -verb runAs -Args "-start GeneralProfile"
-              nvm install 18.16.0
-              nvm use 18.16.0
+            Start-Process powershell -verb runAs -Args "-start GeneralProfile"
+            nvm install 18.16.0
+            nvm use 18.16.0
       - run: node --version
       - run: corepack enable
       - run:
@@ -1449,8 +1462,23 @@ jobs:
             npm set //registry.npmjs.org/:_authToken=$NPM_TOKEN
             npm publish
 
+main_only: &main_only
+  filters:
+    branches:
+      only: main
+
 workflows:
   version: 2
+  noop:
+    when:
+      not:
+        or:
+          - << pipeline.parameters.run-all-workflows >>
+          - << pipeline.parameters.run-python-workflow >>
+          - << pipeline.parameters.run-ts-workflow >>
+          - << pipeline.parameters.run-chat-workflow >>
+    jobs:
+      - noop
   build-chat-offline-installers:
     when:
       or:
@@ -1484,23 +1512,30 @@ workflows:
         - << pipeline.parameters.run-chat-workflow >>
     jobs:
       - hold:
+          <<: *main_only
           type: approval
       - build-online-chat-installer-macos:
+          <<: *main_only
           requires:
             - hold
       - sign-online-chat-installer-macos:
+          <<: *main_only
           requires:
             - build-online-chat-installer-macos
       - notarize-online-chat-installer-macos:
+          <<: *main_only
           requires:
             - sign-online-chat-installer-macos
       - build-online-chat-installer-windows:
+          <<: *main_only
           requires:
             - hold
       - sign-online-chat-installer-windows:
+          <<: *main_only
           requires:
             - build-online-chat-installer-windows
       - build-online-chat-installer-linux:
+          <<: *main_only
           requires:
             - hold
   build-and-test-gpt4all-chat:
@@ -1527,47 +1562,31 @@ workflows:
         - << pipeline.parameters.run-python-workflow >>
     jobs:
       - build-ts-docs:
-          filters:
-            branches:
-              only:
-                - main
+          <<: *main_only
       - build-py-docs:
-          filters:
-            branches:
-              only:
-                - main
-  build-py-deploy:
+          <<: *main_only
+  build-python:
     when:
       or:
         - << pipeline.parameters.run-all-workflows >>
         - << pipeline.parameters.run-python-workflow >>
     jobs:
       - pypi-hold:
+          <<: *main_only
           type: approval
       - hold:
           type: approval
       - build-py-linux:
-          filters:
-            branches:
-              only:
           requires:
             - hold
       - build-py-macos:
-          filters:
-            branches:
-              only:
           requires:
             - hold
       - build-py-windows:
-          filters:
-            branches:
-              only:
           requires:
             - hold
-      - store-and-upload-wheels:
-          filters:
-            branches:
-              only:
+      - publish-wheels:
+          <<: *main_only
           requires:
             - pypi-hold
             - build-py-windows
@@ -1582,59 +1601,38 @@ workflows:
     jobs:
       - hold:
           type: approval
-      - nuget-hold:
-          type: approval
       - nodejs-hold:
           type: approval
       - npm-hold:
+          <<: *main_only
           type: approval
       - build-bindings-backend-linux:
-          filters:
-            branches:
-              only:
           requires:
             - hold
       - build-bindings-backend-macos:
-          filters:
-            branches:
-              only:
           requires:
             - hold
       - build-bindings-backend-windows:
-          filters:
-            branches:
-              only:
           requires:
             - hold
 
       # NodeJs Jobs
       - prepare-npm-pkg:
-          filters:
-            branches:
-              only:
+          <<: *main_only
           requires:
             - npm-hold
             - build-nodejs-linux
             - build-nodejs-windows
             - build-nodejs-macos
       - build-nodejs-linux:
-          filters:
-            branches:
-              only:
           requires:
             - nodejs-hold
             - build-bindings-backend-linux
       - build-nodejs-windows:
-          filters:
-            branches:
-              only:
           requires:
             - nodejs-hold
             - build-bindings-backend-windows
       - build-nodejs-macos:
-          filters:
-            branches:
-              only:
           requires:
             - nodejs-hold
             - build-bindings-backend-macos

.codespellrc

@@ -1,3 +1,3 @@
 [codespell]
-ignore-words-list = blong, afterall, som, assistent, crasher
-skip = .git,*.pdf,*.svg,*.lock,*.ts
+ignore-words-list = blong, afterall, assistent, crasher, requestor
+skip = ./.git,./gpt4all-chat/translations,*.pdf,*.svg,*.lock

.gitignore

@@ -182,6 +182,7 @@ gpt4all-chat/models/*
 build_*
 build-*
 cmake-build-*
+/gpt4all-chat/tests/python/config.py
 
 # IntelliJ
 .idea/

.gitmodules

@@ -16,4 +16,10 @@
 	url = https://github.com/nomic-ai/DuckX.git
 [submodule "gpt4all-chat/deps/QXlsx"]
 	path = gpt4all-chat/deps/QXlsx
-	url = https://github.com/QtExcel/QXlsx.git
+	url = https://github.com/nomic-ai/QXlsx.git
+[submodule "gpt4all-chat/deps/Jinja2Cpp"]
+	path = gpt4all-chat/deps/Jinja2Cpp
+	url = https://github.com/nomic-ai/jinja2cpp.git
+[submodule "gpt4all-chat/deps/rapidjson"]
+	path = gpt4all-chat/deps/rapidjson
+	url = https://github.com/nomic-ai/rapidjson.git

README.md

@@ -15,7 +15,7 @@
   Read about what's new in <a href="https://www.nomic.ai/blog/tag/gpt4all">our blog</a>.
 </p>
 <p align="center">
-  <a href="https://nomic.ai/gpt4all">Subscribe to the newsletter</a>
+  <a href="https://nomic.ai/gpt4all/#newsletter-form">Subscribe to the newsletter</a>
 </p>
 
 https://github.com/nomic-ai/gpt4all/assets/70534565/513a0f15-4964-4109-89e4-4f9a9011f311

common/common.cmake

@@ -0,0 +1,42 @@
+function(gpt4all_add_warning_options target)
+    if (MSVC)
+        return()
+    endif()
+    target_compile_options("${target}" PRIVATE
+        # base options
+        -Wall
+        -Wextra
+        # extra options
+        -Wcast-align
+        -Wextra-semi
+        -Wformat=2
+        -Wmissing-include-dirs
+        -Wstrict-overflow=2
+        -Wsuggest-override
+        -Wvla
+        # errors
+        -Werror=format-security
+        -Werror=init-self
+        -Werror=pointer-arith
+        -Werror=undef
+        # disabled warnings
+        -Wno-sign-compare
+        -Wno-unused-parameter
+    )
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+        target_compile_options("${target}" PRIVATE
+            -Wduplicated-branches
+            -Wduplicated-cond
+            -Wlogical-op
+            -Wno-reorder
+            -Wno-null-dereference
+        )
+    elseif (CMAKE_CXX_COMPILER_ID MATCHES "^(Apple)?Clang$")
+        target_compile_options("${target}" PRIVATE
+            -Wunreachable-code-break
+            -Wunreachable-code-return
+            -Werror=pointer-integer-compare
+            -Wno-reorder-ctor
+        )
+    endif()
+endfunction()

gpt4all-backend/CMakeLists.txt

@@ -1,4 +1,7 @@
 cmake_minimum_required(VERSION 3.23)  # for FILE_SET
+
+include(../common/common.cmake)
+
 set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
@@ -94,8 +97,6 @@ if (LLMODEL_ROCM)
     list(APPEND BUILD_VARIANTS rocm rocm-avxonly)
 endif()
 
-set(CMAKE_VERBOSE_MAKEFILE ON)
-
 # Go through each build variant
 foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
     # Determine flags
@@ -151,6 +152,7 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
     # Add each individual implementations
     add_library(llamamodel-mainline-${BUILD_VARIANT} SHARED
         src/llamamodel.cpp src/llmodel_shared.cpp)
+    gpt4all_add_warning_options(llamamodel-mainline-${BUILD_VARIANT})
     target_compile_definitions(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
         LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
     target_include_directories(llamamodel-mainline-${BUILD_VARIANT} PRIVATE
@@ -169,6 +171,7 @@ add_library(llmodel
     src/llmodel_c.cpp
     src/llmodel_shared.cpp
 )
+gpt4all_add_warning_options(llmodel)
 target_sources(llmodel PUBLIC
     FILE_SET public_headers TYPE HEADERS BASE_DIRS include
     FILES include/gpt4all-backend/llmodel.h

gpt4all-backend/include/gpt4all-backend/llmodel.h

@@ -5,6 +5,7 @@
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
+#include <expected>
 #include <functional>
 #include <optional>
 #include <span>
@@ -24,6 +25,10 @@ using namespace std::string_literals;
 class LLModel {
 public:
     using Token = int32_t;
+    using PromptCallback      = std::function<bool(std::span<const Token> batch, bool cached)>;
+    using ResponseCallback    = std::function<bool(Token token, std::string_view piece)>;
+    using EmbedCancelCallback = bool(unsigned *batchSizes, unsigned nBatch, const char *backend);
+    using ProgressCallback    = std::function<bool(float progress)>;
 
     class BadArchError: public std::runtime_error {
     public:
@@ -101,6 +106,7 @@ public:
         static int32_t maxContextLength(const std::string &modelPath);
         static int32_t layerCount(const std::string &modelPath);
         static bool isEmbeddingModel(const std::string &modelPath);
+        static auto chatTemplate(const char *modelPath) -> std::expected<std::string, std::string>;
         static void setImplementationsSearchPath(const std::string &path);
         static const std::string &implementationsSearchPath();
         static bool hasSupportedCPU();
@@ -124,9 +130,6 @@ public:
     };
 
     struct PromptContext {
-        std::vector<int32_t> tokens;    // current tokens in the context window
-        int32_t n_past = 0;             // number of tokens in past conversation
-        int32_t n_ctx = 0;              // number of tokens possible in context window
         int32_t n_predict = 200;
         int32_t top_k = 40;
         float   top_p = 0.9f;
@@ -138,34 +141,28 @@ public:
         float   contextErase = 0.5f;    // percent of context to erase if we exceed the context window
     };
 
-    using ProgressCallback = std::function<bool(float progress)>;
-
     explicit LLModel() {}
     virtual ~LLModel() {}
 
     virtual bool supportsEmbedding() const = 0;
     virtual bool supportsCompletion() const = 0;
     virtual bool loadModel(const std::string &modelPath, int n_ctx, int ngl) = 0;
-    virtual bool isModelBlacklisted(const std::string &modelPath) const { (void)modelPath; return false; };
+    virtual bool isModelBlacklisted(const std::string &modelPath) const { (void)modelPath; return false; }
     virtual bool isEmbeddingModel(const std::string &modelPath) const { (void)modelPath; return false; }
     virtual bool isModelLoaded() const = 0;
     virtual size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) = 0;
     virtual size_t stateSize() const = 0;
-    virtual size_t saveState(std::span<uint8_t> dest) const = 0;
-    virtual size_t restoreState(std::span<const uint8_t> src) = 0;
+    virtual size_t saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const = 0;
+    virtual size_t restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens) = 0;
 
     // This method requires the model to return true from supportsCompletion otherwise it will throw
     // an error
-    virtual void prompt(const std::string &prompt,
-                        const std::string &promptTemplate,
-                        std::function<bool(int32_t)> promptCallback,
-                        std::function<bool(int32_t, const std::string&)> responseCallback,
-                        bool allowContextShift,
-                        PromptContext &ctx,
-                        bool special = false,
-                        std::optional<std::string_view> fakeReply = {});
+    virtual void prompt(std::string_view        prompt,
+                        const PromptCallback   &promptCallback,
+                        const ResponseCallback &responseCallback,
+                        const PromptContext    &ctx);
 
-    using EmbedCancelCallback = bool(unsigned *batchSizes, unsigned nBatch, const char *backend);
+    virtual int32_t countPromptTokens(std::string_view prompt) const;
 
     virtual size_t embeddingSize() const {
         throw std::logic_error(std::string(implementation().modelType()) + " does not support embeddings");
@@ -210,17 +207,24 @@ public:
 
     void setProgressCallback(ProgressCallback callback) { m_progressCallback = callback; }
 
+    virtual int32_t contextLength() const = 0;
+    virtual auto specialTokens() -> std::unordered_map<std::string, std::string> const = 0;
+
 protected:
     // These are pure virtual because subclasses need to implement as the default implementation of
     // 'prompt' above calls these functions
-    virtual std::vector<Token> tokenize(PromptContext &ctx, std::string_view str, bool special = false) = 0;
+    virtual std::vector<Token> tokenize(std::string_view str) const = 0;
     virtual bool isSpecialToken(Token id) const = 0;
     virtual std::string tokenToString(Token id) const = 0;
-    virtual void initSampler(PromptContext &ctx) = 0;
+    virtual void initSampler(const PromptContext &ctx) = 0;
     virtual Token sampleToken() const = 0;
-    virtual bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const = 0;
-    virtual void shiftContext(PromptContext &promptCtx) = 0;
-    virtual int32_t contextLength() const = 0;
+    virtual bool evalTokens(int32_t nPast, std::span<const Token> tokens) const = 0;
+    virtual void shiftContext(const PromptContext &promptCtx, int32_t *nPast) = 0;
+    virtual int32_t inputLength() const = 0;
+    virtual int32_t computeModelInputPosition(std::span<const Token> input) const = 0;
+    virtual void setModelInputPosition(int32_t pos) = 0;
+    virtual void appendInputToken(Token tok) = 0;
+    virtual std::span<const Token> inputTokens() const = 0;
     virtual const std::vector<Token> &endTokens() const = 0;
     virtual bool shouldAddBOS() const = 0;
 
@@ -236,6 +240,12 @@ protected:
         return -1;
     }
 
+    virtual auto chatTemplate(const char *modelPath) const -> std::expected<std::string, std::string>
+    {
+        (void)modelPath;
+        return std::unexpected("not implemented");
+    }
+
     const Implementation *m_implementation = nullptr;
 
     ProgressCallback m_progressCallback;
@@ -247,17 +257,15 @@ protected:
         return true;
     }
 
-    bool decodePrompt(std::function<bool(int32_t)> promptCallback,
-                      std::function<bool(int32_t, const std::string&)> responseCallback,
-                      bool allowContextShift,
-                      PromptContext &promptCtx,
-                      std::vector<Token> embd_inp,
-                      bool isResponse = false);
-    void generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
-                          bool allowContextShift,
-                          PromptContext &promptCtx);
-
-    Token m_tokenize_last_token = -1; // not serialized
+    // prefill context with prompt
+    auto decodePrompt(const PromptCallback &promptCallback,
+                      const PromptContext  &promptCtx,
+                      std::vector<Token>    embd_inp)
+        -> std::optional<int32_t>;
+    // generate a response
+    void generateResponse(const ResponseCallback &responseCallback,
+                          const PromptContext    &promptCtx,
+                          int32_t                 nPast);
 
     friend class LLMImplementation;
 };

gpt4all-backend/include/gpt4all-backend/llmodel_c.h

@@ -23,6 +23,11 @@ extern "C" {
  */
 typedef void *llmodel_model;
 
+/**
+ * A token.
+ */
+typedef int32_t token_t;
+
 /**
  * llmodel_prompt_context structure for holding the prompt context.
  * NOTE: The implementation takes care of all the memory handling of the raw logits pointer and the
@@ -30,19 +35,15 @@ typedef void *llmodel_model;
  * behavior.
  */
 struct llmodel_prompt_context {
-    int32_t *tokens;        // current tokens in the context window
-    size_t tokens_size;     // the size of the raw tokens vector
-    int32_t n_past;         // number of tokens in past conversation
-    int32_t n_ctx;          // number of tokens possible in context window
     int32_t n_predict;      // number of tokens to predict
     int32_t top_k;          // top k logits to sample from
-    float top_p;            // nucleus sampling probability threshold
-    float min_p;            // Min P sampling
-    float temp;             // temperature to adjust model's output distribution
+    float   top_p;          // nucleus sampling probability threshold
+    float   min_p;          // Min P sampling
+    float   temp;           // temperature to adjust model's output distribution
     int32_t n_batch;        // number of predictions to generate in parallel
-    float repeat_penalty;   // penalty factor for repeated tokens
+    float   repeat_penalty; // penalty factor for repeated tokens
     int32_t repeat_last_n;  // last n tokens to penalize
-    float context_erase;    // percent of context to erase if we exceed the context window
+    float   context_erase;  // percent of context to erase if we exceed the context window
 };
 
 struct llmodel_gpu_device {
@@ -61,10 +62,12 @@ typedef struct llmodel_gpu_device llmodel_gpu_device;
 
 /**
  * Callback type for prompt processing.
- * @param token_id The token id of the prompt.
+ * @param token_ids An array of token ids of the prompt.
+ * @param n_token_ids The number of tokens in the array.
+ * @param cached Whether the tokens were already in cache.
  * @return a bool indicating whether the model should keep processing.
  */
-typedef bool (*llmodel_prompt_callback)(int32_t token_id);
+typedef bool (*llmodel_prompt_callback)(const token_t *token_ids, size_t n_token_ids, bool cached);
 
 /**
  * Callback type for response.
@@ -72,7 +75,7 @@ typedef bool (*llmodel_prompt_callback)(int32_t token_id);
  * @param response The response string. NOTE: a token_id of -1 indicates the string is an error string.
  * @return a bool indicating whether the model should keep generating.
  */
-typedef bool (*llmodel_response_callback)(int32_t token_id, const char *response);
+typedef bool (*llmodel_response_callback)(token_t token_id, const char *response);
 
 /**
  * Embedding cancellation callback for use with llmodel_embed.
@@ -83,6 +86,8 @@ typedef bool (*llmodel_response_callback)(int32_t token_id, const char *response
  */
 typedef bool (*llmodel_emb_cancel_callback)(unsigned *batch_sizes, unsigned n_batch, const char *backend);
 
+typedef void (*llmodel_special_token_callback)(const char *name, const char *token);
+
 /**
  * Create a llmodel instance.
  * Recognises correct model type from file at model_path
@@ -141,48 +146,57 @@ bool llmodel_isModelLoaded(llmodel_model model);
  * @param model A pointer to the llmodel_model instance.
  * @return the size in bytes of the internal state of the model
  */
-uint64_t llmodel_get_state_size(llmodel_model model);
+uint64_t llmodel_state_get_size(llmodel_model model);
 
 /**
- * Saves the internal state of the model to the specified destination address.
+ * Saves the internal state of the model.
  * NOTE: This state data is specific to the type of model you have created.
  * @param model A pointer to the llmodel_model instance.
- * @param dest A pointer to the destination.
- * @param size The size of the destination buffer.
- * @return the number of bytes copied, or zero on error.
+ * @param state Where to store the state. This must be a buffer of at least llmodel_state_get_size() bytes.
+ * @param state_size The size of the destination for the state.
+ * @param input_tokens_out Where to store the address of the token cache state. This is dynamically allocated and must
+ * be freed with llmodel_state_free_input_tokens.
+ * @param n_input_tokens Where to store the size of the token cache state.
+ * @return The number of bytes copied. On error, zero is returned, the token cache is set to NULL, and the token cache
+ * size is set to zero.
  */
-uint64_t llmodel_save_state_data(llmodel_model model, uint8_t *dest, uint64_t size);
+uint64_t llmodel_state_get_data(llmodel_model model, uint8_t *state_out, uint64_t state_size,
+                                token_t **input_tokens_out, uint64_t *n_input_tokens);
+
+/**
+ * Frees the temporary token cache buffer created by a call to llmodel_state_get_data().
+ * @param input_tokens The token cache buffer.
+ */
+void llmodel_state_free_input_tokens(token_t *input_tokens);
 
 /**
  * Restores the internal state of the model using data from the specified address.
  * NOTE: This state data is specific to the type of model you have created.
  * @param model A pointer to the llmodel_model instance.
- * @param src A pointer to the state data.
- * @param size The size of the source data.
+ * @param state A pointer to the state data.
+ * @param state_size The size of the state data.
+ * @param input_tokens The token cache associated with the saved state.
+ * @param n_input_tokens The number of tokens in input_tokens.
  * @return The number of bytes read, or zero on error.
  */
-uint64_t llmodel_restore_state_data(llmodel_model model, const uint8_t *src, size_t size);
+uint64_t llmodel_state_set_data(llmodel_model model, const uint8_t *state, uint64_t state_size,
+                                const token_t *input_tokens, uint64_t n_input_tokens);
 
 /**
  * Generate a response using the model.
  * @param model A pointer to the llmodel_model instance.
  * @param prompt A string representing the input prompt.
- * @param prompt_template A string representing the input prompt template.
  * @param prompt_callback A callback function for handling the processing of prompt.
  * @param response_callback A callback function for handling the generated response.
- * @param allow_context_shift Whether to allow shifting of context to make room for more input.
- * @param special True if special tokens in the prompt should be processed, false otherwise.
- * @param fake_reply A string to insert into context as the model's reply, or NULL to generate one.
  * @param ctx A pointer to the llmodel_prompt_context structure.
+ * @param error A pointer to a string; will only be set on error.
  */
-void llmodel_prompt(llmodel_model model, const char *prompt,
-                    const char *prompt_template,
-                    llmodel_prompt_callback prompt_callback,
-                    llmodel_response_callback response_callback,
-                    bool allow_context_shift,
-                    llmodel_prompt_context *ctx,
-                    bool special,
-                    const char *fake_reply);
+bool llmodel_prompt(llmodel_model               model,
+                    const char                 *prompt,
+                    llmodel_prompt_callback     prompt_callback,
+                    llmodel_response_callback   response_callback,
+                    llmodel_prompt_context     *ctx,
+                    const char                **error);
 
 /**
  * Generate an embedding using the model.
@@ -294,6 +308,10 @@ const char *llmodel_model_backend_name(llmodel_model model);
  */
 const char *llmodel_model_gpu_device_name(llmodel_model model);
 
+int32_t llmodel_count_prompt_tokens(llmodel_model model, const char *prompt, const char **error);
+
+void llmodel_model_foreach_special_token(llmodel_model model, llmodel_special_token_callback callback);
+
 #ifdef __cplusplus
 }
 #endif

gpt4all-backend/src/llamamodel.cpp

@@ -202,7 +202,7 @@ static int32_t get_arch_key_u32(std::string const &modelPath, std::string const
         if (keyidx != -1) {
             value = gguf_get_val_u32(ctx, keyidx);
         } else {
-            std::cerr << __func__ << ": " << key << "not found in " << modelPath << "\n";
+            std::cerr << __func__ << ": " << key << " not found in " << modelPath << "\n";
         }
     }
 
@@ -218,6 +218,7 @@ struct LLamaPrivate {
     int64_t                      n_threads    = 0;
     std::vector<LLModel::Token>  end_tokens;
     const char                  *backend_name = nullptr;
+    std::vector<LLModel::Token>  inputTokens;
 
     llama_model          *model        = nullptr;
     llama_context        *ctx          = nullptr;
@@ -501,28 +502,29 @@ size_t LLamaModel::stateSize() const
     return llama_state_get_size(d_ptr->ctx);
 }
 
-size_t LLamaModel::saveState(std::span<uint8_t> dest) const
+size_t LLamaModel::saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const
 {
-    return llama_state_get_data(d_ptr->ctx, dest.data(), dest.size());
+    size_t bytesWritten = llama_state_get_data(d_ptr->ctx, stateOut.data(), stateOut.size());
+    if (bytesWritten)
+        inputTokensOut.assign(d_ptr->inputTokens.begin(), d_ptr->inputTokens.end());
+    return bytesWritten;
 }
 
-size_t LLamaModel::restoreState(std::span<const uint8_t> src)
+size_t LLamaModel::restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens)
 {
-    return llama_state_set_data(d_ptr->ctx, src.data(), src.size());
+    size_t bytesRead = llama_state_set_data(d_ptr->ctx, state.data(), state.size());
+    if (bytesRead)
+        d_ptr->inputTokens.assign(inputTokens.begin(), inputTokens.end());
+    return bytesRead;
 }
 
-std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, std::string_view str, bool special)
+std::vector<LLModel::Token> LLamaModel::tokenize(std::string_view str) const
 {
-    bool atStart = m_tokenize_last_token == -1;
-    bool insertSpace = atStart || isSpecialToken(m_tokenize_last_token);
     std::vector<LLModel::Token> fres(str.length() + 4);
-    int32_t fres_len = llama_tokenize_gpt4all(
-        d_ptr->model, str.data(), str.length(), fres.data(), fres.size(), /*add_special*/ atStart,
-        /*parse_special*/ special, /*insert_space*/ insertSpace
+    int32_t fres_len = llama_tokenize(
+        d_ptr->model, str.data(), str.length(), fres.data(), fres.size(), /*add_special*/ true, /*parse_special*/ true
     );
     fres.resize(fres_len);
-    if (fres_len)
-        m_tokenize_last_token = fres.back();
     return fres;
 }
 
@@ -548,7 +550,7 @@ std::string LLamaModel::tokenToString(Token id) const
     return std::string(result.data(), result.size());
 }
 
-void LLamaModel::initSampler(PromptContext &promptCtx)
+void LLamaModel::initSampler(const PromptContext &promptCtx)
 {
     auto *model = d_ptr->model;
     auto *chain = d_ptr->sampler_chain;
@@ -582,7 +584,8 @@ void LLamaModel::initSampler(PromptContext &promptCtx)
             llama_sampler_init_top_p(promptCtx.top_p, 1),
             llama_sampler_init_min_p(promptCtx.min_p, 1),
             llama_sampler_init_temp(promptCtx.temp),
-            llama_sampler_init_dist(LLAMA_DEFAULT_SEED)
+            llama_sampler_init_softmax(),
+            llama_sampler_init_dist(LLAMA_DEFAULT_SEED),
         };
         for (auto *smpl : samplers)
             llama_sampler_chain_add(chain, smpl);
@@ -594,9 +597,11 @@ LLModel::Token LLamaModel::sampleToken() const
     return llama_sampler_sample(d_ptr->sampler_chain, d_ptr->ctx, -1);
 }
 
-bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
+bool LLamaModel::evalTokens(int32_t nPast, std::span<const Token> tokens) const
 {
-    llama_kv_cache_seq_rm(d_ptr->ctx, 0, ctx.n_past, -1);
+    assert(!tokens.empty());
+
+    llama_kv_cache_seq_rm(d_ptr->ctx, 0, nPast, -1);
 
     llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
 
@@ -604,7 +609,7 @@ bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &toke
 
     for (int32_t i = 0; i < batch.n_tokens; i++) {
         batch.token   [i] = tokens[i];
-        batch.pos     [i] = ctx.n_past + i;
+        batch.pos     [i] = nPast + i;
         batch.n_seq_id[i] = 1;
         batch.seq_id  [i][0] = 0;
         batch.logits  [i] = false;
@@ -618,14 +623,14 @@ bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &toke
     return res == 0;
 }
 
-void LLamaModel::shiftContext(PromptContext &promptCtx)
+void LLamaModel::shiftContext(const PromptContext &promptCtx, int32_t *nPast)
 {
     // infinite text generation via context shifting
 
     // erase up to n_ctx*contextErase tokens
     int n_keep = shouldAddBOS();
-    int n_past = promptCtx.n_past;
-    int n_discard = std::min(n_past - n_keep, int(promptCtx.n_ctx * promptCtx.contextErase));
+    int n_past = *nPast;
+    int n_discard = std::min(n_past - n_keep, int(contextLength() * promptCtx.contextErase));
 
     assert(n_discard > 0);
     if (n_discard <= 0)
@@ -638,8 +643,9 @@ void LLamaModel::shiftContext(PromptContext &promptCtx)
     llama_kv_cache_seq_rm (d_ptr->ctx, 0, n_keep,             n_keep + n_discard);
     llama_kv_cache_seq_add(d_ptr->ctx, 0, n_keep + n_discard, n_past,             -n_discard);
 
-    promptCtx.tokens.erase(promptCtx.tokens.begin() + n_keep, promptCtx.tokens.begin() + n_keep + n_discard);
-    promptCtx.n_past = promptCtx.tokens.size();
+    auto &inp = d_ptr->inputTokens;
+    inp.erase(inp.begin() + n_keep, inp.begin() + n_keep + n_discard);
+    *nPast = inp.size();
 }
 
 int32_t LLamaModel::contextLength() const
@@ -647,6 +653,56 @@ int32_t LLamaModel::contextLength() const
     return llama_n_ctx(d_ptr->ctx);
 }
 
+auto LLamaModel::specialTokens() -> std::unordered_map<std::string, std::string> const
+{
+    if (!d_ptr->model)
+        throw std::logic_error("model not loaded");
+
+    std::unordered_map<std::string, std::string> tokens;
+    if (auto id = llama_token_bos(d_ptr->model); id != LLAMA_TOKEN_NULL)
+        tokens.emplace("bos_token", tokenToString(id));
+    if (auto id = llama_token_eos(d_ptr->model); id != LLAMA_TOKEN_NULL)
+        tokens.emplace("eos_token", tokenToString(id));
+    return tokens;
+}
+
+int32_t LLamaModel::inputLength() const
+{
+    return d_ptr->inputTokens.size();
+}
+
+int32_t LLamaModel::computeModelInputPosition(std::span<const Token> input) const
+{
+    // find common prefix
+    auto cacheIt = d_ptr->inputTokens.begin();
+    auto inputIt = input.begin();
+    while (cacheIt < d_ptr->inputTokens.end() && inputIt < input.end() && *cacheIt == *inputIt) {
+        ++cacheIt; ++inputIt;
+    }
+    // tell the caller to ignore the tokens between [begin, inputIt)
+    return inputIt - input.begin();
+}
+
+void LLamaModel::setModelInputPosition(int32_t pos)
+{
+    auto &inp = d_ptr->inputTokens;
+    assert(pos >= 0);
+    assert(pos <= inp.size());
+    // truncate token cache to end at the new n_past
+    if (pos < inp.size())
+        inp.resize(pos);
+}
+
+void LLamaModel::appendInputToken(Token tok)
+{
+    d_ptr->inputTokens.push_back(tok);
+}
+
+auto LLamaModel::inputTokens() const -> std::span<const Token>
+{
+    return d_ptr->inputTokens;
+}
+
 const std::vector<LLModel::Token> &LLamaModel::endTokens() const
 {
     return d_ptr->end_tokens;
@@ -667,6 +723,37 @@ int32_t LLamaModel::layerCount(std::string const &modelPath) const
     return get_arch_key_u32(modelPath, "block_count");
 }
 
+// TODO(jared): reduce redundant code and operations by combining all metadata getters for unloaded
+//              models into a class that keeps the model file open
+auto LLamaModel::chatTemplate(const char *modelPath) const -> std::expected<std::string, std::string>
+{
+    auto *ctx = load_gguf(modelPath);
+    if (!ctx)
+        return std::unexpected("failed to open model file");
+
+    std::expected<std::string, std::string> result;
+    enum gguf_type ktype;
+    const int kid = gguf_find_key(ctx, "tokenizer.chat_template");
+    if (kid == -1) {
+        result = std::unexpected("key not found");
+        goto cleanup;
+    }
+
+    ktype = gguf_get_kv_type(ctx, kid);
+    if (ktype != GGUF_TYPE_STRING) {
+        result = std::unexpected(
+            "expected key type STRING (" + std::to_string(GGUF_TYPE_STRING) + "), got " + std::to_string(ktype)
+        );
+        goto cleanup;
+    }
+
+    result = gguf_get_val_str(ctx, kid);
+
+cleanup:
+    gguf_free(ctx);
+    return result;
+}
+
 #ifdef GGML_USE_VULKAN
 static const char *getVulkanVendorName(uint32_t vendorID)
 {

gpt4all-backend/src/llamamodel_impl.h

@@ -11,6 +11,7 @@
 #include <string>
 #include <string_view>
 #include <vector>
+#include <unordered_map>
 
 struct LLamaPrivate;
 struct EmbModelSpec;
@@ -28,8 +29,8 @@ public:
     bool isModelLoaded() const override;
     size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) override;
     size_t stateSize() const override;
-    size_t saveState(std::span<uint8_t> dest) const override;
-    size_t restoreState(std::span<const uint8_t> src) override;
+    size_t saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const override;
+    size_t restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens) override;
     void setThreadCount(int32_t n_threads) override;
     int32_t threadCount() const override;
     std::vector<GPUDevice> availableGPUDevices(size_t memoryRequired = 0) const override;
@@ -48,28 +49,36 @@ public:
     void embed(const std::vector<std::string> &texts, float *embeddings, bool isRetrieval, int dimensionality = -1,
                size_t *tokenCount = nullptr, bool doMean = true, bool atlas = false) override;
 
-private:
-    std::unique_ptr<LLamaPrivate> d_ptr;
-    bool m_supportsEmbedding = false;
-    bool m_supportsCompletion = false;
+    int32_t contextLength() const override;
+    auto specialTokens() -> std::unordered_map<std::string, std::string> const override;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &ctx, std::string_view str, bool special) override;
+    std::vector<Token> tokenize(std::string_view str) const override;
     bool isSpecialToken(Token id) const override;
     std::string tokenToString(Token id) const override;
-    void initSampler(PromptContext &ctx) override;
+    void initSampler(const PromptContext &ctx) override;
     Token sampleToken() const override;
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;
-    void shiftContext(PromptContext &promptCtx) override;
-    int32_t contextLength() const override;
+    bool evalTokens(int32_t nPast, std::span<const Token> tokens) const override;
+    void shiftContext(const PromptContext &promptCtx, int32_t *nPast) override;
+    int32_t inputLength() const override;
+    int32_t computeModelInputPosition(std::span<const Token> input) const override;
+    void setModelInputPosition(int32_t pos) override;
+    void appendInputToken(Token tok) override;
+    std::span<const Token> inputTokens() const override;
     const std::vector<Token> &endTokens() const override;
     bool shouldAddBOS() const override;
     int32_t maxContextLength(std::string const &modelPath) const override;
     int32_t layerCount(std::string const &modelPath) const override;
+    auto chatTemplate(const char *modelPath) const -> std::expected<std::string, std::string> override;
 
     void embedInternal(const std::vector<std::string> &texts, float *embeddings, std::string prefix, int dimensionality,
                        size_t *tokenCount, bool doMean, bool atlas, EmbedCancelCallback *cancelCb,
                        const EmbModelSpec *spec);
+
+private:
+    std::unique_ptr<LLamaPrivate> d_ptr;
+    bool m_supportsEmbedding = false;
+    bool m_supportsCompletion = false;
 };
 
 #endif // LLAMAMODEL_H

gpt4all-backend/src/llmodel.cpp

@@ -326,6 +326,12 @@ bool LLModel::Implementation::isEmbeddingModel(const std::string &modelPath)
     return llama && llama->isEmbeddingModel(modelPath);
 }
 
+auto LLModel::Implementation::chatTemplate(const char *modelPath) -> std::expected<std::string, std::string>
+{
+    auto *llama = constructGlobalLlama();
+    return llama ? llama->chatTemplate(modelPath) : std::unexpected("backend not available");
+}
+
 void LLModel::Implementation::setImplementationsSearchPath(const std::string& path)
 {
     s_implementations_search_path = path;

gpt4all-backend/src/llmodel_c.cpp

@@ -7,17 +7,20 @@
 #include <cstdlib>
 #include <cstring>
 #include <exception>
-#include <functional>
 #include <iostream>
 #include <memory>
 #include <optional>
 #include <string>
 #include <string_view>
 #include <vector>
+#include <span>
+
+namespace ranges = std::ranges;
+
+static_assert(sizeof(token_t) == sizeof(LLModel::Token));
 
 struct LLModelWrapper {
     LLModel *llModel = nullptr;
-    LLModel::PromptContext promptContext;
     ~LLModelWrapper() { delete llModel; }
 };
 
@@ -85,74 +88,80 @@ bool llmodel_isModelLoaded(llmodel_model model)
     return wrapper->llModel->isModelLoaded();
 }
 
-uint64_t llmodel_get_state_size(llmodel_model model)
+uint64_t llmodel_state_get_size(llmodel_model model)
 {
     auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->stateSize();
 }
 
-uint64_t llmodel_save_state_data(llmodel_model model, uint8_t *dest, uint64_t size)
+uint64_t llmodel_state_get_data(llmodel_model model, uint8_t *state_out, uint64_t state_size,
+                                token_t **input_tokens_out, uint64_t *n_input_tokens)
 {
     auto *wrapper = static_cast<LLModelWrapper *>(model);
-    return wrapper->llModel->saveState({dest, size_t(size)});
+    std::vector<LLModel::Token> inputTokens;
+    auto bytesWritten = wrapper->llModel->saveState({state_out, size_t(state_size)}, inputTokens);
+    if (bytesWritten) {
+        auto *buf = new LLModel::Token[inputTokens.size()];
+        ranges::copy(inputTokens, buf);
+        *input_tokens_out = buf;
+        *n_input_tokens = uint64_t(inputTokens.size());
+    } else {
+        *input_tokens_out = nullptr;
+        *n_input_tokens = 0;
+    }
+    return bytesWritten;
 }
 
-uint64_t llmodel_restore_state_data(llmodel_model model, const uint8_t *src, uint64_t size)
+void llmodel_state_free_input_tokens(LLModel::Token *input_tokens)
 {
-    auto *wrapper = static_cast<LLModelWrapper *>(model);
-    return wrapper->llModel->restoreState({src, size_t(size)});
+    delete[] input_tokens;
 }
 
-void llmodel_prompt(llmodel_model model, const char *prompt,
-                    const char *prompt_template,
-                    llmodel_prompt_callback prompt_callback,
-                    llmodel_response_callback response_callback,
-                    bool allow_context_shift,
-                    llmodel_prompt_context *ctx,
-                    bool special,
-                    const char *fake_reply)
+uint64_t llmodel_state_set_data(llmodel_model model, const uint8_t *state, uint64_t state_size,
+                                const token_t *input_tokens, uint64_t n_input_tokens)
 {
     auto *wrapper = static_cast<LLModelWrapper *>(model);
+    return wrapper->llModel->restoreState({state, size_t(state_size)}, {input_tokens, size_t(n_input_tokens)});
+}
 
-    auto response_func = [response_callback](int32_t token_id, const std::string &response) {
-        return response_callback(token_id, response.c_str());
-    };
+bool llmodel_prompt(llmodel_model               model,
+                    const char                 *prompt,
+                    llmodel_prompt_callback     prompt_callback,
+                    llmodel_response_callback   response_callback,
+                    llmodel_prompt_context     *ctx,
+                    const char                **error)
+{
+    auto *wrapper = static_cast<LLModelWrapper *>(model);
 
     // Copy the C prompt context
-    wrapper->promptContext.n_past = ctx->n_past;
-    wrapper->promptContext.n_ctx = ctx->n_ctx;
-    wrapper->promptContext.n_predict = ctx->n_predict;
-    wrapper->promptContext.top_k = ctx->top_k;
-    wrapper->promptContext.top_p = ctx->top_p;
-    wrapper->promptContext.min_p = ctx->min_p;
-    wrapper->promptContext.temp = ctx->temp;
-    wrapper->promptContext.n_batch = ctx->n_batch;
-    wrapper->promptContext.repeat_penalty = ctx->repeat_penalty;
-    wrapper->promptContext.repeat_last_n = ctx->repeat_last_n;
-    wrapper->promptContext.contextErase = ctx->context_erase;
+    LLModel::PromptContext promptContext {
+        .n_predict      = ctx->n_predict,
+        .top_k          = ctx->top_k,
+        .top_p          = ctx->top_p,
+        .min_p          = ctx->min_p,
+        .temp           = ctx->temp,
+        .n_batch        = ctx->n_batch,
+        .repeat_penalty = ctx->repeat_penalty,
+        .repeat_last_n  = ctx->repeat_last_n,
+        .contextErase   = ctx->context_erase,
+    };
+
+    auto prompt_func = [prompt_callback](std::span<const LLModel::Token> token_ids, bool cached) {
+        return prompt_callback(token_ids.data(), token_ids.size(), cached);
+    };
+    auto response_func = [response_callback](LLModel::Token token_id, std::string_view piece) {
+        return response_callback(token_id, piece.data());
+    };
 
     // Call the C++ prompt method
-    wrapper->llModel->prompt(prompt, prompt_template, prompt_callback, response_func, allow_context_shift,
-                             wrapper->promptContext, special,
-                             fake_reply ? std::make_optional<std::string_view>(fake_reply) : std::nullopt);
+    try {
+        wrapper->llModel->prompt(prompt, prompt_func, response_func, promptContext);
+    } catch (std::exception const &e) {
+        llmodel_set_error(error, e.what());
+        return false;
+    }
 
-    // Update the C context by giving access to the wrappers raw pointers to std::vector data
-    // which involves no copies
-    ctx->tokens = wrapper->promptContext.tokens.data();
-    ctx->tokens_size = wrapper->promptContext.tokens.size();
-
-    // Update the rest of the C prompt context
-    ctx->n_past = wrapper->promptContext.n_past;
-    ctx->n_ctx = wrapper->promptContext.n_ctx;
-    ctx->n_predict = wrapper->promptContext.n_predict;
-    ctx->top_k = wrapper->promptContext.top_k;
-    ctx->top_p = wrapper->promptContext.top_p;
-    ctx->min_p = wrapper->promptContext.min_p;
-    ctx->temp = wrapper->promptContext.temp;
-    ctx->n_batch = wrapper->promptContext.n_batch;
-    ctx->repeat_penalty = wrapper->promptContext.repeat_penalty;
-    ctx->repeat_last_n = wrapper->promptContext.repeat_last_n;
-    ctx->context_erase = wrapper->promptContext.contextErase;
+    return true;
 }
 
 float *llmodel_embed(
@@ -291,3 +300,21 @@ const char *llmodel_model_gpu_device_name(llmodel_model model)
     const auto *wrapper = static_cast<LLModelWrapper *>(model);
     return wrapper->llModel->gpuDeviceName();
 }
+
+int32_t llmodel_count_prompt_tokens(llmodel_model model, const char *prompt, const char **error)
+{
+    auto *wrapper = static_cast<const LLModelWrapper *>(model);
+    try {
+        return wrapper->llModel->countPromptTokens(prompt);
+    } catch (const std::exception& e) {
+        llmodel_set_error(error, e.what());
+        return -1;
+    }
+}
+
+void llmodel_model_foreach_special_token(llmodel_model model, llmodel_special_token_callback callback)
+{
+    auto *wrapper = static_cast<const LLModelWrapper *>(model);
+    for (auto &[name, token] : wrapper->llModel->specialTokens())
+        callback(name.c_str(), token.c_str());
+}

gpt4all-backend/src/llmodel_shared.cpp

@@ -4,211 +4,120 @@
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
-#include <functional>
 #include <iostream>
+#include <iterator>
 #include <optional>
-#include <regex>
-#include <sstream>
+#include <ranges>
 #include <stdexcept>
 #include <string>
 #include <string_view>
 #include <vector>
 
 namespace ranges = std::ranges;
+namespace views  = std::ranges::views;
 
-static bool parsePromptTemplate(const std::string &tmpl, std::vector<std::smatch> &placeholders, std::string &err)
-{
-    static const std::regex placeholderRegex(R"(%[1-2](?![0-9]))");
+void LLModel::prompt(
+    std::string_view        prompt,
+    const PromptCallback   &promptCallback,
+    const ResponseCallback &responseCallback,
+    const PromptContext    &promptCtx
+) {
+    if (!isModelLoaded())
+        throw std::invalid_argument("Attempted to prompt an unloaded model.");
+    if (!supportsCompletion())
+        throw std::invalid_argument("Not a text completion model.");
+    if (!promptCtx.n_batch)
+        throw std::invalid_argument("Batch size cannot be zero.");
+    if (!promptCtx.n_predict)
+        return; // nothing requested
 
-    auto it = std::sregex_iterator(tmpl.begin(), tmpl.end(), placeholderRegex);
-    placeholders.clear();
-    placeholders.insert(placeholders.end(), it, std::sregex_iterator());
+    auto embd_inp = tokenize(prompt);
+    if (embd_inp.empty())
+        throw std::invalid_argument("Prompt tokenized to zero tokens.");
 
-    if (placeholders.size() > 2) {
-        err = "ERROR: expected at most two placeholders, got " + std::to_string(placeholders.size());
-        return false;
-    }
-    if (placeholders.size() >= 1 && placeholders[0].str() != "%1") {
-        err = "ERROR: first placeholder must be %1, got " + placeholders[0].str();
-        return false;
-    }
-    if (placeholders.size() >= 2 && placeholders[1].str() != "%2") {
-        err = "ERROR: second placeholder must be %2, got " + placeholders[1].str();
-        return false;
-    }
-    return true;
+    if (auto res = decodePrompt(promptCallback, promptCtx, std::move(embd_inp)))
+        generateResponse(responseCallback, promptCtx, /*n_past*/ *res);
 }
 
-void LLModel::prompt(const std::string &prompt,
-                     const std::string &promptTemplate,
-                     std::function<bool(int32_t)> promptCallback,
-                     std::function<bool(int32_t, const std::string&)> responseCallback,
-                     bool allowContextShift,
-                     PromptContext &promptCtx,
-                     bool special,
-                     std::optional<std::string_view> fakeReply)
+int32_t LLModel::countPromptTokens(std::string_view prompt) const
 {
-    if (!isModelLoaded()) {
-        std::cerr << implementation().modelType() << " ERROR: prompt won't work with an unloaded model!\n";
-        return;
-    }
-
-    if (!supportsCompletion()) {
-        std::string errorMessage = "ERROR: this model does not support text completion or chat!";
-        responseCallback(-1, errorMessage);
-        std::cerr << implementation().modelType() << " " << errorMessage << "\n";
-        return;
-    }
-
-    // sanity checks
-    if (promptCtx.n_past > contextLength()) {
-        std::ostringstream ss;
-        ss << "n_past=" << promptCtx.n_past << " is past end of context length=" << contextLength();
-        throw std::out_of_range(ss.str());
-    }
-    if (promptCtx.n_past > promptCtx.tokens.size()) {
-        std::ostringstream ss;
-        ss << "n_past=" << promptCtx.n_past << " is past end of token cache length=" << promptCtx.tokens.size();
-        throw std::out_of_range(ss.str());
-    }
-
-    promptCtx.n_ctx = contextLength();
-    promptCtx.n_batch = std::min(promptCtx.n_batch, LLMODEL_MAX_PROMPT_BATCH);
-
-    if (promptCtx.n_past < promptCtx.tokens.size())
-        promptCtx.tokens.resize(promptCtx.n_past);
-    m_tokenize_last_token = promptCtx.tokens.empty() ? -1 : promptCtx.tokens.back(); // not serialized
-
-    // parse the prompt template
-    std::vector<std::smatch> placeholders;
-    {
-        std::string err;
-        if (!parsePromptTemplate(promptTemplate, placeholders, err)) {
-            responseCallback(-1, err);
-            std::cerr << err << "\n";
-            return;
-        }
-    }
-
-    auto old_n_past = promptCtx.n_past; // prepare to fake n_past for tokenize
-
-    // tokenize the user prompt
-    std::vector<Token> embd_inp;
-    if (placeholders.empty()) {
-        // this is unusual, but well-defined
-        std::cerr << __func__ << ": prompt template has no placeholder\n";
-        embd_inp = tokenize(promptCtx, promptTemplate, true);
-    } else {
-        // template: beginning of user prompt
-        const auto &phUser = placeholders[0];
-        std::string userPrefix(phUser.prefix());
-        if (!userPrefix.empty()) {
-            embd_inp = tokenize(promptCtx, userPrefix, true);
-            promptCtx.n_past += embd_inp.size();
-        }
-
-        // user input (shouldn't have special token processing)
-        auto tokens = tokenize(promptCtx, prompt, special);
-        embd_inp.insert(embd_inp.end(), tokens.begin(), tokens.end());
-        promptCtx.n_past += tokens.size();
-
-        // template: end of user prompt + start of assistant prompt
-        size_t start = phUser.position() + phUser.length();
-        size_t end = placeholders.size() >= 2 ? placeholders[1].position() : promptTemplate.length();
-        auto userToAsst = promptTemplate.substr(start, end - start);
-        if (!userToAsst.empty()) {
-            tokens = tokenize(promptCtx, userToAsst, true);
-            embd_inp.insert(embd_inp.end(), tokens.begin(), tokens.end());
-            promptCtx.n_past += tokens.size();
-        }
-    }
-
-    promptCtx.n_past = old_n_past; // restore n_past so decodePrompt can increment it
-
-    // decode the user prompt
-    if (!decodePrompt(promptCallback, responseCallback, allowContextShift, promptCtx, embd_inp))
-        return; // error
-
-    // decode the assistant's reply, either generated or spoofed
-    if (!fakeReply) {
-        generateResponse(responseCallback, allowContextShift, promptCtx);
-    } else {
-        embd_inp = tokenize(promptCtx, *fakeReply, false);
-        if (!decodePrompt(promptCallback, responseCallback, allowContextShift, promptCtx, embd_inp, true))
-            return; // error
-    }
-
-    // decode the rest of the prompt template
-    // template: end of assistant prompt
-    std::string asstSuffix;
-    if (placeholders.size() >= 2) {
-        size_t start = placeholders[1].position() + placeholders[1].length();
-        asstSuffix = promptTemplate.substr(start);
-    } else {
-        asstSuffix = "\n\n"; // default to a blank link, good for e.g. Alpaca
-    }
-    if (!asstSuffix.empty()) {
-        embd_inp = tokenize(promptCtx, asstSuffix, true);
-        decodePrompt(promptCallback, responseCallback, allowContextShift, promptCtx, embd_inp);
-    }
+    if (!isModelLoaded())
+        throw std::invalid_argument("Attempted to tokenize with an unloaded model.");
+    return int32_t(tokenize(prompt).size());
 }
 
-// returns false on error
-bool LLModel::decodePrompt(std::function<bool(int32_t)> promptCallback,
-                           std::function<bool(int32_t, const std::string&)> responseCallback,
-                           bool allowContextShift,
-                           PromptContext &promptCtx,
-                           std::vector<Token> embd_inp,
-                           bool isResponse) {
-    if ((int) embd_inp.size() > promptCtx.n_ctx - 4) {
-        // FIXME: (Adam) We should find a way to bubble these strings to the UI level to allow for
-        // translation
-        responseCallback(-1, "Your message was too long and could not be processed. Please try again with something shorter.");
-        std::cerr << implementation().modelType() << " ERROR: The prompt is " << embd_inp.size() <<
-            " tokens and the context window is " << promptCtx.n_ctx << "!\n";
-        return false;
+auto LLModel::decodePrompt(
+    const PromptCallback &promptCallback,
+    const PromptContext  &promptCtx,
+    std::vector<Token>    embd_inp
+) -> std::optional<int32_t>
+{
+    assert(!embd_inp.empty());
+
+    int32_t nCtx = contextLength();
+    int32_t n_batch = std::min(promptCtx.n_batch, LLMODEL_MAX_PROMPT_BATCH);
+
+    // Find the greatest n_past where the beginning of embd_inp matches the end of the token cache, starting at the
+    // requested n_past.
+    // This is used to skip unnecessary work when the prompt shares a common prefix with the previous result.
+    int32_t nPast = computeModelInputPosition(embd_inp);
+
+    // always decode up to a full batch before generating, even if cached
+    nPast -= std::min(n_batch, nPast);
+
+    // TODO(jared): generalize this to find the smallest new_embd_inp.size() - nPast given the cache
+    if (!nPast && int32_t(embd_inp.size()) > nCtx) {
+        // no cache hit -> shift the input before even processing
+
+        int32_t nKeep     = shouldAddBOS();
+        auto    newLength = int32_t(nCtx * (1.f - promptCtx.contextErase));
+        int32_t nDiscard  = int32_t(embd_inp.size()) - std::max(1, std::min(nCtx, newLength));
+
+        // execute the callback even for skipped tokens. this misrepresents the position of BOS but we don't care
+        auto discardedTokens = embd_inp | views::drop(nKeep) | views::take(nDiscard);
+        if (!promptCallback(discardedTokens, true))
+            return std::nullopt;
+
+        // erase nDiscard tokens
+        embd_inp.erase(discardedTokens.begin(), discardedTokens.end());
+        assert(int32_t(embd_inp.size()) <= nCtx);
+
+        // check the cache again, just in case
+        nPast = computeModelInputPosition(embd_inp);
+        nPast -= std::min(n_batch, nPast);
     }
 
-    // FIXME(jared): There are mitigations for this situation, such as making room before
-    // copying the prompt context, or restoring the KV cache when we restore the prompt
-    // context.
-    if (!allowContextShift && promptCtx.n_past + embd_inp.size() > promptCtx.n_ctx) {
-        std::cerr << "LLModel Warning: Not enough space, n_past=" << promptCtx.n_past << ", n_eval=" << embd_inp.size()
-                  << ", n_ctx=" << promptCtx.n_ctx << "\n";
-        return false;
-    }
+    setModelInputPosition(nPast);
+
+    // execute the callback even for skipped tokens
+    if (!promptCallback(embd_inp | views::take(nPast), true))
+        return std::nullopt;
 
     // process the prompt in batches
-    size_t i = 0;
-    while (i < embd_inp.size()) {
-        size_t batch_end = std::min(i + promptCtx.n_batch, embd_inp.size());
-        std::vector<Token> batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
+    for (int32_t i = nPast; i < embd_inp.size();) {
+        auto batch_end = std::min(i + n_batch, int32_t(embd_inp.size()));
+        std::span batch(embd_inp.begin() + i, embd_inp.begin() + batch_end);
 
         // Check if the context has run out...
-        if (promptCtx.n_past + int32_t(batch.size()) > promptCtx.n_ctx) {
-            assert(allowContextShift);
-            shiftContext(promptCtx);
-            assert(promptCtx.n_past + int32_t(batch.size()) <= promptCtx.n_ctx);
+        if (nPast + int32_t(batch.size()) > nCtx) {
+            shiftContext(promptCtx, &nPast);
+            assert(nPast + int32_t(batch.size()) <= nCtx);
         }
 
-        if (!evalTokens(promptCtx, batch)) {
-            std::cerr << implementation().modelType() << " ERROR: Failed to process prompt\n";
-            return false;
-        }
+        // FIXME(Adam): We should find a way to bubble these strings to the UI level to allow for translation
+        if (!evalTokens(nPast, batch))
+            throw std::runtime_error("An internal error was encountered during prompt processing.");
 
-        size_t tokens = batch_end - i;
-        for (size_t t = 0; t < tokens; ++t) {
-            promptCtx.tokens.push_back(batch.at(t));
-            promptCtx.n_past += 1;
-            Token tok = batch.at(t);
-            bool res = isResponse ? responseCallback(tok, tokenToString(tok)) : promptCallback(tok);
-            if (!res)
-                return false;
+        for (auto &tok : batch) {
+            appendInputToken(tok);
+            nPast++;
+            if (!promptCallback({ &tok, 1 }, false))
+                return std::nullopt;
         }
         i = batch_end;
     }
 
-    return true;
+    return nPast;
 }
 
 /*
@@ -230,22 +139,16 @@ static std::string::size_type stringsOverlap(const std::string &s, const std::st
     return std::string::npos;
 }
 
-void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)> responseCallback,
-                               bool allowContextShift,
-                               PromptContext &promptCtx) {
+void LLModel::generateResponse(
+    const ResponseCallback &responseCallback,
+    const PromptContext    &promptCtx,
+    int32_t                 nPast
+) {
     static const char *stopSequences[] {
-        "### Instruction", "### Prompt", "### Response", "### Human", "### Assistant", "### Context",
+        "### System", "### Instruction", "### Human", "### User", "### Response", "### Assistant", "### Context",
+        "<|im_start|>", "<|im_end|>", "<|endoftext|>",
     };
 
-    // Don't even start if there is no room
-    if (!promptCtx.n_predict)
-        return;
-    if (!allowContextShift && promptCtx.n_past >= promptCtx.n_ctx) {
-        std::cerr << "LLModel Warning: Not enough space, n_past=" << promptCtx.n_past << ", n_ctx=" << promptCtx.n_ctx
-                  << "\n";
-        return;
-    }
-
     initSampler(promptCtx);
 
     std::string cachedResponse;
@@ -260,26 +163,20 @@ void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)>
         cachedTokens.push_back(new_tok.value());
         cachedResponse += new_piece;
 
-        auto accept = [this, &promptCtx, &cachedTokens, &new_tok, allowContextShift]() -> bool {
+        auto accept = [this, &promptCtx, &new_tok, &nPast] {
             // Shift context if out of space
-            if (promptCtx.n_past >= promptCtx.n_ctx) {
-                (void)allowContextShift;
-                assert(allowContextShift);
-                shiftContext(promptCtx);
-                assert(promptCtx.n_past < promptCtx.n_ctx);
+            if (nPast >= contextLength()) {
+                shiftContext(promptCtx, &nPast);
+                assert(nPast < contextLength());
             }
 
             // Accept the token
             Token tok = std::exchange(new_tok, std::nullopt).value();
-            if (!evalTokens(promptCtx, { tok })) {
-                // TODO(jared): raise an exception
-                std::cerr << implementation().modelType() << " ERROR: Failed to predict next token\n";
-                return false;
-            }
+            if (!evalTokens(nPast, { &tok, 1 }))
+                throw std::runtime_error("An internal error was encountered during response generation.");
 
-            promptCtx.tokens.push_back(tok);
-            promptCtx.n_past += 1;
-            return true;
+            appendInputToken(tok);
+            nPast++;
         };
 
         // Check for EOS
@@ -316,13 +213,6 @@ void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)>
             lengthLimit = cachedResponse.size() - new_piece.size();
         }
 
-        // Optionally stop if the context will run out
-        if (!allowContextShift && promptCtx.n_past + cachedTokens.size() >= promptCtx.n_ctx) {
-            std::cerr << "LLModel Warning: Not enough space, n_past=" << promptCtx.n_past << ", n_ctx="
-                      << promptCtx.n_ctx << "\n";
-            stop = true;
-        }
-
         // Empty the cache, up to the length limit
         std::string::size_type responseLength = 0;
         while (!cachedTokens.empty()) {
@@ -339,8 +229,8 @@ void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)>
             cachedResponse.erase(cachedResponse.begin(), cachedResponse.begin() + piece.size());
 
             // Accept the token, if needed (not cached)
-            if (cachedTokens.empty() && new_tok && !accept())
-                return;
+            if (cachedTokens.empty() && new_tok)
+                accept();
 
             // Send the token
             if (!responseCallback(tok, piece) || ++n_predicted >= promptCtx.n_predict) {
@@ -359,24 +249,23 @@ void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)>
             assert(!cachedTokens.empty() && cachedTokens.back() == new_tok);
             if (stop) {
                 cachedTokens.pop_back();
-            } else if (!accept()) {
-                return;
+            } else {
+                accept();
             }
         }
     }
 
-    auto &tokens = promptCtx.tokens;
-    if (tokens.size() < cachedTokens.size()) {
+    if (inputLength() < cachedTokens.size()) {
         /* This is theoretically possible if the longest stop sequence is greater than
          * n_ctx * contextErase tokens. */
         throw std::runtime_error("shifted too much context, can't go back");
     }
 
-    auto discard_start = tokens.end() - cachedTokens.size();
-    assert(std::equal(discard_start, tokens.end(), cachedTokens.begin()));
-    tokens.erase(discard_start, tokens.end());
-
-    promptCtx.n_past -= cachedTokens.size();
+#ifndef NDEBUG
+    auto inp = inputTokens();
+    auto discard_start = inp.end() - cachedTokens.size();
+    assert(std::equal(discard_start, inp.end(), cachedTokens.begin()));
+#endif
 }
 
 void LLModel::embed(

gpt4all-bindings/cli/app.py

@@ -113,10 +113,7 @@ def _old_loop(gpt4all_instance):
         full_response = gpt4all_instance.chat_completion(
             MESSAGES,
             # preferential kwargs for chat ux
-            logits_size=0,
-            tokens_size=0,
             n_past=0,
-            n_ctx=0,
             n_predict=200,
             top_k=40,
             top_p=0.9,

gpt4all-bindings/python/CHANGELOG.md

@@ -8,10 +8,12 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ### Added
 - Warn on Windows if the Microsoft Visual C++ runtime libraries are not found ([#2920](https://github.com/nomic-ai/gpt4all/pull/2920))
+- Basic cache for faster prefill when the input shares a prefix with previous context ([#3073](https://github.com/nomic-ai/gpt4all/pull/3073))
 
 ### Changed
 - Rebase llama.cpp on latest upstream as of September 26th ([#2998](https://github.com/nomic-ai/gpt4all/pull/2998))
 - Change the error message when a message is too long ([#3004](https://github.com/nomic-ai/gpt4all/pull/3004))
+- Fix CalledProcessError on Intel Macs since v2.8.0 ([#3045](https://github.com/nomic-ai/gpt4all/pull/3045))
 
 ## [2.8.2] - 2024-08-14
 

gpt4all-bindings/python/docs/gpt4all_desktop/chat_templates.md

@@ -0,0 +1,206 @@
+## What are chat templates?
+Natively, large language models only know how to complete plain text and do not know the difference between their input and their output. In order to support a chat with a person, LLMs are designed to use a template to convert the conversation to plain text using a specific format.
+
+For a given model, it is important to use an appropriate chat template, as each model is designed to work best with a specific format. The chat templates included with the built-in models should be sufficient for most purposes.
+
+There are two reasons you would want to alter the chat template:
+
+- You are sideloading a model and there is no chat template available,
+- You would like to have greater control over the input to the LLM than a system message provides.
+
+
+## What is a system message?
+A system message is a message that controls the responses from the LLM in a way that affects the entire conversation. System messages can be short, such as "Speak like a pirate.", or they can be long and contain a lot of context for the LLM to keep in mind.
+
+Not all models are designed to use a system message, so they work with some models better than others.
+
+
+## How do I customize the chat template or system message?
+To customize the chat template or system message, go to Settings > Model. Make sure to select the correct model at the top. If you clone a model, you can use a different chat template or system message from the base model, enabling you to use different settings for each conversation.
+
+These settings take effect immediately. After changing them, you can click "Redo last response" in the chat view, and the response will take the new settings into account.
+
+
+## Do I need to write a chat template?
+You typically do not need to write your own chat template. The exception is models that are not in the official model list and do not come with a chat template built-in. These will show a "Clear" option above the chat template field in the Model Settings page instead of a "Reset" option. See the section on [finding] or [creating] a chat template.
+
+[finding]: #how-do-i-find-a-chat-template
+[creating]: #advanced-how-do-chat-templates-work
+
+
+## What changed in GPT4All v3.5?
+GPT4All v3.5 overhauled the chat template system. There are three crucial differences:
+
+- The chat template now formats an entire conversation instead of a single pair of messages,
+- The chat template now uses Jinja syntax instead of `%1` and `%2` placeholders,
+- And the system message should no longer contain control tokens or trailing whitespace.
+
+If you are using any chat templates or system messages that had been added or altered from the default before upgrading to GPT4All v3.5 or newer, these will no longer work. See below for how to solve common errors you may see after upgrading.
+
+
+## Error/Warning: System message is not plain text.
+This is easy to fix. Go to the model's settings and look at the system prompt. There are three things to look for:
+
+- Control tokens such as `<|im_start|>`, `<|start_header_id|>`, or `<|system|>`
+- A prefix such as `### System` or `SYSTEM:`
+- Trailing whitespace, such as a space character or blank line.
+
+If you see any of these things, remove them. For example, this legacy system prompt:
+```
+<|start_header_id|>system<|end_header_id|>
+You are a helpful assistant.<|eot_id|>
+```
+
+Should become this:
+```
+You are a helpful assistant.
+```
+
+If you do not see anything that needs to be changed, you can dismiss the error by making a minor modification to the message and then changing it back.
+
+If you see a warning, your system message does not appear to be plain text. If you believe this warning is incorrect, it can be safely ignored. If in doubt, ask on the [Discord].
+
+[Discord]: https://discord.gg/mGZE39AS3e
+
+
+## Error: Legacy system prompt needs to be updated in Settings.
+This is the same as [above][above-1], but appears on the chat page.
+
+[above-1]: #errorwarning-system-message-is-not-plain-text
+
+
+## Error/Warning: Chat template is not in Jinja format.
+This is the result of attempting to use an old-style template (possibly from a previous version) in GPT4All 3.5+.
+
+Go to the Model Settings page and select the affected model. If you see a "Reset" button, and you have not intentionally modified the prompt template, you can click "Reset". Otherwise, this is what you can do:
+
+1. Back up your chat template by copying it safely to a text file and saving it. In the next step, it will be removed from GPT4All.
+2. Click "Reset" or "Clear".
+3. If you clicked "Clear", the chat template is now gone. Follow the steps to [find][finding] or [create][creating] a basic chat template for your model.
+4. Customize the chat template to suit your needs. For help, read the section about [creating] a chat template.
+
+
+## Error: Legacy prompt template needs to be updated in Settings.
+This is the same as [above][above-2], but appears on the chat page.
+
+[above-2]: #errorwarning-chat-template-is-not-in-jinja-format
+
+
+## The chat template has a syntax error.
+If there is a syntax error while editing the chat template, the details will be displayed in an error message above the input box. This could be because the chat template is not actually in Jinja format (see [above][above-2]).
+
+Otherwise, you have either typed something correctly, or the model comes with a template that is incompatible with GPT4All. See [the below section][creating] on creating chat templates and make sure that everything is correct. When in doubt, ask on the [Discord].
+
+
+## Error: No chat template configured.
+This may appear for models that are not from the official model list and do not include a chat template. Older versions of GPT4All picked a poor default in this case. You will get much better results if you follow the steps to [find][finding] or [create][creating] a chat template for your model.
+
+
+## Error: The chat template cannot be blank.
+If the button above the chat template on the Model Settings page says "Clear", see [above][above-3]. If you see "Reset", click that button to restore a reasonable default. Also see the section on [syntax errors][chat-syntax-error].
+
+[above-3]: #error-no-chat-template-configured
+[chat-syntax-error]: #the-chat-template-has-a-syntax-error
+
+
+## How do I find a chat template?
+When in doubt, you can always ask the [Discord] community for help. Below are the instructions to find one on your own.
+
+The authoritative source for a model's chat template is the HuggingFace repo that the original (non-GGUF) model came from. First, you should find this page. If you just have a model file, you can try a google search for the model's name. If you know the page you downloaded the GGUF model from, its README usually links to the original non-GGUF model.
+
+Once you have located the original model, there are two methods you can use to extract its chat template. Pick whichever one you are most comfortable with.
+
+### Using the CLI (all models)
+1. Install `jq` using your preferred package manager - e.g. Chocolatey (Windows), Homebrew (macOS), or apt (Ubuntu).
+2. Download `tokenizer_config.json` from the model's "Files and versions" tab.
+3. Open a command prompt in the directory which you have downloaded the model file.
+4. Run `jq -r ".chat_template" tokenizer_config.json`. This shows the chat template in a human-readable form. You can copy this and paste it into the settings page.
+5. (Optional) You can save the output to a text file like this: `jq -r ".chat_template" tokenizer_config.json >chat_template.txt`
+
+If the output is "null", the model does not provide a chat template. See the [below instructions][creating] on creating a chat template.
+
+### Python (open models)
+1. Install `transformers` using your preferred python package manager, e.g. `pip install transformers`. Make sure it is at least version v4.43.0.
+2. Copy the ID of the HuggingFace model, using the clipboard icon next to the name. For example, if the URL is `https://huggingface.co/NousResearch/Hermes-2-Pro-Llama-3-8B`, the ID is `NousResearch/Hermes-2-Pro-Llama-3-8B`.
+3. Open a python interpreter (`python`) and run the following commands. Change the model ID in the example to the one you copied.
+```
+>>> from transformers import AutoTokenizer
+>>> tokenizer = AutoTokenizer.from_pretrained('NousResearch/Hermes-2-Pro-Llama-3-8B')
+>>> print(tokenizer.get_chat_template())
+```
+You can copy the output and paste it into the settings page.
+4. (Optional) You can save the output to a text file like this:
+```
+>>> open('chat_template.txt', 'w').write(tokenizer.get_chat_template())
+```
+
+If you get a ValueError exception, this model does not provide a chat template. See the [below instructions][creating] on creating a chat template.
+
+
+### Python (gated models)
+Some models, such as Llama and Mistral, do not allow public access to their chat template. You must either use the CLI method above, or follow the following instructions to use Python:
+
+1. For these steps, you must have git and git-lfs installed.
+2. You must have a HuggingFace account and be logged in.
+3. You must already have access to the gated model. Otherwise, request access.
+4. You must have an SSH key configured for git access to HuggingFace.
+5. `git clone` the model's HuggingFace repo using the SSH clone URL. There is no need to download the entire model, which is very large. A good way to do this on Linux is:
+```console
+$ GIT_LFS_SKIP_SMUDGE=1 git clone hf.co:meta-llama/Llama-3.1-8B-Instruct.git
+$ cd Llama-3.1-8B-Instruct
+$ git lfs pull -I "tokenizer.*"
+```
+6. Follow the above instructions for open models, but replace the model ID with the path to the directory containing `tokenizer\_config.json`:
+```
+>>> tokenizer = AutoTokenizer.from_pretrained('.')
+```
+
+
+## Advanced: How do chat templates work?
+The chat template is applied to the entire conversation you see in the chat window. The template loops over the list of messages, each containing `role` and `content` fields. `role` is either `user`, `assistant`, or `system`.
+
+GPT4All also supports the special variables `bos_token`, `eos_token`, and `add_generation_prompt`. See the [HuggingFace docs] for what those do.
+
+[HuggingFace docs]: https://huggingface.co/docs/transformers/v4.46.3/en/chat_templating#special-variables
+
+
+## Advanced: How do I make a chat template?
+The best way to create a chat template is to start by using an existing one as a reference. Then, modify it to use the format documented for the given model. Its README page may explicitly give an example of its template. Or, it may mention the name of a well-known standard template, such as ChatML, Alpaca, Vicuna. GPT4All does not yet include presets for these templates, so they will have to be found in other models or taken from the community.
+
+For more information, see the very helpful [HuggingFace guide]. Some of this is not applicable, such as the information about tool calling and RAG - GPT4All implements those features differently.
+
+Some models use a prompt template that does not intuitively map to a multi-turn chat, because it is more intended for single instructions. The [FastChat] implementation of these templates is a useful reference for the correct way to extend them to multiple messages.
+
+[HuggingFace guide]: https://huggingface.co/docs/transformers/v4.46.3/en/chat_templating#advanced-template-writing-tips
+[FastChat]: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+
+
+# Advanced: What are GPT4All v1 templates?
+GPT4All supports its own template syntax, which is nonstandard but provides complete control over the way LocalDocs sources and file attachments are inserted into the conversation. These templates begin with `{# gpt4all v1 #}` and look similar to the example below.
+
+For standard templates, GPT4All combines the user message, sources, and attachments into the `content` field. For GPT4All v1 templates, this is not done, so they must be used directly in the template for those features to work correctly.
+
+```jinja
+{# gpt4all v1 #}
+{%- for message in messages %}
+    {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n' }}
+    {%- if message['role'] == 'user' %}
+        {%- for source in message['sources'] %}
+            {%- if loop.first %}
+                {{- '### Context:\n' }}
+            {%- endif %}
+            {{- 'Collection: ' + source['collection'] + '\n'   +
+                'Path: '       + source['path']       + '\n'   +
+                'Excerpt: '    + source['text']       + '\n\n' }}
+        {%- endfor %}
+    {%- endif %}
+    {%- for attachment in message['prompt_attachments'] %}
+        {{- attachment['processed_content'] + '\n\n' }}
+    {%- endfor %}
+    {{- message['content'] | trim }}
+    {{- '<|eot_id|>' }}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
+```

gpt4all-bindings/python/docs/gpt4all_desktop/cookbook/use-local-ai-models-to-privately-chat-with-microsoft-excel.md

@@ -0,0 +1,85 @@
+# Using GPT4All to Privately Chat with your Microsoft Excel Spreadsheets
+Local and Private AI Chat with your Microsoft Excel Spreadsheets
+
+Microsoft Excel allows you to create, manage, and analyze data in spreadsheet format. By attaching your spreadsheets directly to GPT4All, you can privately chat with the AI to query and explore the data, enabling you to summarize, generate reports, and glean insights from your files—all within your conversation.
+
+<div style="position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;">
+  <iframe src="../../assets/gpt4all_xlsx_attachment.mp4" style="position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;" allowfullscreen title="YouTube Video"></iframe>
+</div>
+
+
+## Attach Microsoft Excel to your GPT4All Conversation
+
+!!! note "Attach Microsoft Excel to your GPT4All Conversation"
+
+    1. **Install GPT4All and Open **:
+
+        - Go to [nomic.ai/gpt4all](https://nomic.ai/gpt4all) to install GPT4All for your operating system.
+
+        - Navigate to the Chats view within GPT4All.
+
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Chat view -->
+                  <img width="1348" alt="Chat view" src="../../assets/chat_window.png">
+               </td>
+            </tr>
+         </table>
+
+    2. **Example Spreadsheet **:
+
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Spreadsheet view -->
+                  <img width="1348" alt="Spreadsheet view" src="../../assets/disney_spreadsheet.png">
+               </td>
+            </tr>
+         </table>
+
+    3. **Attach to GPT4All conversration**
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Attach view -->
+                  <img width="1348" alt="Attach view" src="../../assets/attach_spreadsheet.png">
+               </td>
+            </tr>
+         </table>
+
+    4. **Have GPT4All Summarize and Generate a Report**
+        <table>
+            <tr>
+               <td>
+                  <!-- Screenshot of Attach view -->
+                  <img width="1348" alt="Attach view" src="../../assets/spreadsheet_chat.png">
+               </td>
+            </tr>
+         </table>
+
+
+## How It Works
+
+GPT4All parses your attached excel spreadsheet into Markdown, a format understandable to LLMs, and adds the markdown text to the context for your LLM chat. You can view the code that converts `.xslx` to Markdown [here](https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-chat/src/xlsxtomd.cpp) in the GPT4All github repo.
+
+For example, the above spreadsheet titled `disney_income_stmt.xlsx` would be formatted the following way:
+
+```markdown
+## disney_income_stmt
+
+|Walt Disney Co.|||||||
+|---|---|---|---|---|---|---|
+|Consolidated Income Statement|||||||
+|||||||||
+|US$ in millions|||||||
+|12 months ended:|2023-09-30 00:00:00|2022-10-01 00:00:00|2021-10-02 00:00:00|2020-10-03 00:00:00|2019-09-28 00:00:00|2018-09-29 00:00:00|
+|Services|79562|74200|61768|59265|60542|50869|
+...
+...
+...
+```
+
+## Limitations
+
+It is important to double-check the claims LLMs make about the spreadsheets you provide. LLMs can make mistakes about the data they are presented, particularly for the LLMs with smaller parameter counts (~8B) that fit within the memory of consumer hardware.

gpt4all-bindings/python/gpt4all/_pyllmodel.py

@@ -3,14 +3,13 @@ from __future__ import annotations
 import ctypes
 import os
 import platform
-import re
 import subprocess
 import sys
 import textwrap
 import threading
 from enum import Enum
 from queue import Queue
-from typing import TYPE_CHECKING, Any, Callable, Generic, Iterable, Literal, NoReturn, TypeVar, overload
+from typing import TYPE_CHECKING, Any, Callable, Generic, Iterable, Iterator, Literal, NoReturn, TypeVar, overload
 
 if sys.version_info >= (3, 9):
     import importlib.resources as importlib_resources
@@ -24,20 +23,31 @@ else:
     from typing import TypedDict
 
 if TYPE_CHECKING:
-    from typing_extensions import TypeAlias
+    from typing_extensions import ParamSpec, TypeAlias
+    T = TypeVar("T")
+    P = ParamSpec("P")
 
 EmbeddingsType = TypeVar('EmbeddingsType', bound='list[Any]')
 
+cuda_found: bool = False
+
+
+# TODO(jared): use operator.call after we drop python 3.10 support
+def _operator_call(obj: Callable[P, T], /, *args: P.args, **kwargs: P.kwargs) -> T:
+    return obj(*args, **kwargs)
+
 
 # Detect Rosetta 2
-if platform.system() == "Darwin" and platform.processor() == "i386":
-    if subprocess.run(
-        "sysctl -n sysctl.proc_translated".split(), check=True, capture_output=True, text=True,
-    ).stdout.strip() == "1":
-        raise RuntimeError(textwrap.dedent("""\
-            Running GPT4All under Rosetta is not supported due to CPU feature requirements.
-            Please install GPT4All in an environment that uses a native ARM64 Python interpreter.
-        """).strip())
+@_operator_call
+def check_rosetta() -> None:
+    if platform.system() == "Darwin" and platform.processor() == "i386":
+        p = subprocess.run("sysctl -n sysctl.proc_translated".split(), capture_output=True, text=True)
+        if p.returncode == 0 and p.stdout.strip() == "1":
+            raise RuntimeError(textwrap.dedent("""\
+                Running GPT4All under Rosetta is not supported due to CPU feature requirements.
+                Please install GPT4All in an environment that uses a native ARM64 Python interpreter.
+            """).strip())
+
 
 # Check for C++ runtime libraries
 if platform.system() == "Windows":
@@ -53,33 +63,35 @@ if platform.system() == "Windows":
         """), file=sys.stderr)
 
 
-def _load_cuda(rtver: str, blasver: str) -> None:
-    if platform.system() == "Linux":
-        cudalib   = f"lib/libcudart.so.{rtver}"
-        cublaslib = f"lib/libcublas.so.{blasver}"
-    else:  # Windows
-        cudalib   = fr"bin\cudart64_{rtver.replace('.', '')}.dll"
-        cublaslib = fr"bin\cublas64_{blasver}.dll"
+@_operator_call
+def find_cuda() -> None:
+    global cuda_found
 
-    # preload the CUDA libs so the backend can find them
-    ctypes.CDLL(os.path.join(cuda_runtime.__path__[0], cudalib), mode=ctypes.RTLD_GLOBAL)
-    ctypes.CDLL(os.path.join(cublas.__path__[0], cublaslib), mode=ctypes.RTLD_GLOBAL)
+    def _load_cuda(rtver: str, blasver: str) -> None:
+        if platform.system() == "Linux":
+            cudalib   = f"lib/libcudart.so.{rtver}"
+            cublaslib = f"lib/libcublas.so.{blasver}"
+        else:  # Windows
+            cudalib   = fr"bin\cudart64_{rtver.replace('.', '')}.dll"
+            cublaslib = fr"bin\cublas64_{blasver}.dll"
 
+        # preload the CUDA libs so the backend can find them
+        ctypes.CDLL(os.path.join(cuda_runtime.__path__[0], cudalib), mode=ctypes.RTLD_GLOBAL)
+        ctypes.CDLL(os.path.join(cublas.__path__[0], cublaslib), mode=ctypes.RTLD_GLOBAL)
 
-# Find CUDA libraries from the official packages
-cuda_found = False
-if platform.system() in ("Linux", "Windows"):
-    try:
-        from nvidia import cuda_runtime, cublas
-    except ImportError:
-        pass  # CUDA is optional
-    else:
-        for rtver, blasver in [("12", "12"), ("11.0", "11")]:
-            try:
-                _load_cuda(rtver, blasver)
-                cuda_found = True
-            except OSError:  # dlopen() does not give specific error codes
-                pass  # try the next one
+    # Find CUDA libraries from the official packages
+    if platform.system() in ("Linux", "Windows"):
+        try:
+            from nvidia import cuda_runtime, cublas
+        except ImportError:
+            pass  # CUDA is optional
+        else:
+            for rtver, blasver in [("12", "12"), ("11.0", "11")]:
+                try:
+                    _load_cuda(rtver, blasver)
+                    cuda_found = True
+                except OSError:  # dlopen() does not give specific error codes
+                    pass  # try the next one
 
 
 # TODO: provide a config file to make this more robust
@@ -106,21 +118,18 @@ llmodel = load_llmodel_library()
 
 class LLModelPromptContext(ctypes.Structure):
     _fields_ = [
-        ("tokens", ctypes.POINTER(ctypes.c_int32)),
-        ("tokens_size", ctypes.c_size_t),
-        ("n_past", ctypes.c_int32),
-        ("n_ctx", ctypes.c_int32),
-        ("n_predict", ctypes.c_int32),
-        ("top_k", ctypes.c_int32),
-        ("top_p", ctypes.c_float),
-        ("min_p", ctypes.c_float),
-        ("temp", ctypes.c_float),
-        ("n_batch", ctypes.c_int32),
+        ("n_predict",      ctypes.c_int32),
+        ("top_k",          ctypes.c_int32),
+        ("top_p",          ctypes.c_float),
+        ("min_p",          ctypes.c_float),
+        ("temp",           ctypes.c_float),
+        ("n_batch",        ctypes.c_int32),
         ("repeat_penalty", ctypes.c_float),
-        ("repeat_last_n", ctypes.c_int32),
-        ("context_erase", ctypes.c_float),
+        ("repeat_last_n",  ctypes.c_int32),
+        ("context_erase",  ctypes.c_float),
     ]
 
+
 class LLModelGPUDevice(ctypes.Structure):
     _fields_ = [
         ("backend", ctypes.c_char_p),
@@ -131,6 +140,7 @@ class LLModelGPUDevice(ctypes.Structure):
         ("vendor", ctypes.c_char_p),
     ]
 
+
 # Define C function signatures using ctypes
 llmodel.llmodel_model_create.argtypes = [ctypes.c_char_p]
 llmodel.llmodel_model_create.restype = ctypes.c_void_p
@@ -148,23 +158,21 @@ llmodel.llmodel_required_mem.restype = ctypes.c_size_t
 llmodel.llmodel_isModelLoaded.argtypes = [ctypes.c_void_p]
 llmodel.llmodel_isModelLoaded.restype = ctypes.c_bool
 
-PromptCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_int32)
-ResponseCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_int32, ctypes.c_char_p)
-EmbCancelCallback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.POINTER(ctypes.c_uint), ctypes.c_uint, ctypes.c_char_p)
+PromptCallback       = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.POINTER(ctypes.c_int32), ctypes.c_size_t, ctypes.c_bool)
+ResponseCallback     = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_int32, ctypes.c_char_p)
+EmbCancelCallback    = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.POINTER(ctypes.c_uint), ctypes.c_uint, ctypes.c_char_p)
+SpecialTokenCallback = ctypes.CFUNCTYPE(None, ctypes.c_char_p, ctypes.c_char_p)
 
 llmodel.llmodel_prompt.argtypes = [
     ctypes.c_void_p,
     ctypes.c_char_p,
-    ctypes.c_char_p,
     PromptCallback,
     ResponseCallback,
-    ctypes.c_bool,
     ctypes.POINTER(LLModelPromptContext),
-    ctypes.c_bool,
-    ctypes.c_char_p,
+    ctypes.POINTER(ctypes.c_char_p),
 ]
 
-llmodel.llmodel_prompt.restype = None
+llmodel.llmodel_prompt.restype = ctypes.c_bool
 
 llmodel.llmodel_embed.argtypes = [
     ctypes.c_void_p,
@@ -213,6 +221,12 @@ llmodel.llmodel_model_backend_name.restype = ctypes.c_char_p
 llmodel.llmodel_model_gpu_device_name.argtypes = [ctypes.c_void_p]
 llmodel.llmodel_model_gpu_device_name.restype = ctypes.c_char_p
 
+llmodel.llmodel_count_prompt_tokens.argtypes = [ctypes.c_void_p, ctypes.POINTER(ctypes.c_char_p)]
+llmodel.llmodel_count_prompt_tokens.restype = ctypes.c_int32
+
+llmodel.llmodel_model_foreach_special_token.argtypes = [ctypes.c_void_p, SpecialTokenCallback]
+llmodel.llmodel_model_foreach_special_token.restype = None
+
 ResponseCallbackType = Callable[[int, str], bool]
 RawResponseCallbackType = Callable[[int, bytes], bool]
 EmbCancelCallbackType: TypeAlias = 'Callable[[list[int], str], bool]'
@@ -257,7 +271,6 @@ class LLModel:
         self.model_path = model_path.encode()
         self.n_ctx = n_ctx
         self.ngl = ngl
-        self.context: LLModelPromptContext | None = None
         self.buffer = bytearray()
         self.buff_expecting_cont_bytes: int = 0
 
@@ -277,6 +290,10 @@ class LLModel:
 
             raise RuntimeError(f"Unable to instantiate model: {errmsg}")
         self.model: ctypes.c_void_p | None = model
+        self.special_tokens_map: dict[str, str] = {}
+        llmodel.llmodel_model_foreach_special_token(
+            self.model, lambda n, t: self.special_tokens_map.__setitem__(n.decode(), t.decode()),
+        )
 
     def __del__(self, llmodel=llmodel):
         if hasattr(self, 'model'):
@@ -303,6 +320,19 @@ class LLModel:
         dev = llmodel.llmodel_model_gpu_device_name(self.model)
         return None if dev is None else dev.decode()
 
+    def count_prompt_tokens(self, prompt: str) -> int:
+        if self.model is None:
+            self._raise_closed()
+        err = ctypes.c_char_p()
+        n_tok = llmodel.llmodel_count_prompt_tokens(self.model, prompt, ctypes.byref(err))
+        if n_tok < 0:
+            s = err.value
+            errmsg = 'null' if s is None else s.decode()
+            raise RuntimeError(f'Unable to count prompt tokens: {errmsg}')
+        return n_tok
+
+    llmodel.llmodel_count_prompt_tokens.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
+
     @staticmethod
     def list_gpus(mem_required: int = 0) -> list[str]:
         """
@@ -366,50 +396,6 @@ class LLModel:
             raise Exception("Model not loaded")
         return llmodel.llmodel_threadCount(self.model)
 
-    def _set_context(
-        self,
-        n_predict: int = 4096,
-        top_k: int = 40,
-        top_p: float = 0.9,
-        min_p: float = 0.0,
-        temp: float = 0.1,
-        n_batch: int = 8,
-        repeat_penalty: float = 1.2,
-        repeat_last_n: int = 10,
-        context_erase: float = 0.75,
-        reset_context: bool = False,
-    ):
-        if self.context is None:
-            context = LLModelPromptContext(
-                tokens_size=0,
-                n_past=0,
-                n_ctx=0,
-                n_predict=n_predict,
-                top_k=top_k,
-                top_p=top_p,
-                min_p=min_p,
-                temp=temp,
-                n_batch=n_batch,
-                repeat_penalty=repeat_penalty,
-                repeat_last_n=repeat_last_n,
-                context_erase=context_erase,
-            )
-            self.context = context
-        else:
-            context = self.context
-            if reset_context:
-                self.context.n_past = 0
-
-        self.context.n_predict = n_predict
-        self.context.top_k = top_k
-        self.context.top_p = top_p
-        self.context.min_p = min_p
-        self.context.temp = temp
-        self.context.n_batch = n_batch
-        self.context.repeat_penalty = repeat_penalty
-        self.context.repeat_last_n = repeat_last_n
-        self.context.context_erase = context_erase
-
     @overload
     def generate_embeddings(
         self, text: str, prefix: str | None, dimensionality: int, do_mean: bool, atlas: bool,
@@ -479,20 +465,18 @@ class LLModel:
 
     def prompt_model(
         self,
-        prompt: str,
-        prompt_template: str,
-        callback: ResponseCallbackType,
-        n_predict: int = 4096,
-        top_k: int = 40,
-        top_p: float = 0.9,
-        min_p: float = 0.0,
-        temp: float = 0.1,
-        n_batch: int = 8,
-        repeat_penalty: float = 1.2,
-        repeat_last_n: int = 10,
-        context_erase: float = 0.75,
-        reset_context: bool = False,
-        special: bool = False,
+        prompt          : str,
+        callback        : ResponseCallbackType,
+        n_predict       : int                  = 4096,
+        top_k           : int                  = 40,
+        top_p           : float                = 0.9,
+        min_p           : float                = 0.0,
+        temp            : float                = 0.1,
+        n_batch         : int                  = 8,
+        repeat_penalty  : float                = 1.2,
+        repeat_last_n   : int                  = 10,
+        context_erase   : float                = 0.75,
+        reset_context   : bool                 = False,
     ):
         """
         Generate response from model from a prompt.
@@ -515,35 +499,38 @@ class LLModel:
         self.buffer.clear()
         self.buff_expecting_cont_bytes = 0
 
-        self._set_context(
-            n_predict=n_predict,
-            top_k=top_k,
-            top_p=top_p,
-            min_p=min_p,
-            temp=temp,
-            n_batch=n_batch,
-            repeat_penalty=repeat_penalty,
-            repeat_last_n=repeat_last_n,
-            context_erase=context_erase,
-            reset_context=reset_context,
+        context = LLModelPromptContext(
+            n_predict      = n_predict,
+            top_k          = top_k,
+            top_p          = top_p,
+            min_p          = min_p,
+            temp           = temp,
+            n_batch        = n_batch,
+            repeat_penalty = repeat_penalty,
+            repeat_last_n  = repeat_last_n,
+            context_erase  = context_erase,
         )
 
-        llmodel.llmodel_prompt(
+        error_msg: bytes | None = None
+        def error_callback(msg: bytes) -> None:
+            nonlocal error_msg
+            error_msg = msg
+
+        err = ctypes.c_char_p()
+        if not llmodel.llmodel_prompt(
             self.model,
             ctypes.c_char_p(prompt.encode()),
-            ctypes.c_char_p(prompt_template.encode()),
             PromptCallback(self._prompt_callback),
             ResponseCallback(self._callback_decoder(callback)),
-            True,
-            self.context,
-            special,
-            ctypes.c_char_p(),
-        )
-
+            context,
+            ctypes.byref(err),
+        ):
+            s = err.value
+            raise RuntimeError(f"prompt error: {'null' if s is None else s.decode()}")
 
     def prompt_model_streaming(
-        self, prompt: str, prompt_template: str, callback: ResponseCallbackType = empty_response_callback, **kwargs
-    ) -> Iterable[str]:
+        self, prompt: str, callback: ResponseCallbackType = empty_response_callback, **kwargs: Any,
+    ) -> Iterator[str]:
         if self.model is None:
             self._raise_closed()
 
@@ -562,15 +549,15 @@ class LLModel:
 
             return _generator_callback
 
-        def run_llmodel_prompt(prompt: str, prompt_template: str, callback: ResponseCallbackType, **kwargs):
-            self.prompt_model(prompt, prompt_template, callback, **kwargs)
+        def run_llmodel_prompt(prompt: str, callback: ResponseCallbackType, **kwargs):
+            self.prompt_model(prompt, callback, **kwargs)
             output_queue.put(Sentinel.TERMINATING_SYMBOL)
 
         # Kick off llmodel_prompt in separate thread so we can return generator
         # immediately
         thread = threading.Thread(
             target=run_llmodel_prompt,
-            args=(prompt, prompt_template, _generator_callback_wrapper(callback)),
+            args=(prompt, _generator_callback_wrapper(callback)),
             kwargs=kwargs,
         )
         thread.start()
@@ -589,16 +576,16 @@ class LLModel:
             decoded = []
 
             for byte in response:
-                
+
                 bits = "{:08b}".format(byte)
                 (high_ones, _, _) = bits.partition('0')
 
-                if len(high_ones) == 1: 
+                if len(high_ones) == 1:
                     # continuation byte
                     self.buffer.append(byte)
                     self.buff_expecting_cont_bytes -= 1
 
-                else: 
+                else:
                     # beginning of a byte sequence
                     if len(self.buffer) > 0:
                         decoded.append(self.buffer.decode(errors='replace'))
@@ -608,22 +595,22 @@ class LLModel:
                     self.buffer.append(byte)
                     self.buff_expecting_cont_bytes = max(0, len(high_ones) - 1)
 
-                if self.buff_expecting_cont_bytes <= 0: 
+                if self.buff_expecting_cont_bytes <= 0:
                     # received the whole sequence or an out of place continuation byte
                     decoded.append(self.buffer.decode(errors='replace'))
 
                     self.buffer.clear()
                     self.buff_expecting_cont_bytes = 0
-                    
+
             if len(decoded) == 0 and self.buff_expecting_cont_bytes > 0:
                 # wait for more continuation bytes
                 return True
-            
-            return callback(token_id, ''.join(decoded))     
+
+            return callback(token_id, ''.join(decoded))
 
         return _raw_callback
 
     # Empty prompt callback
     @staticmethod
-    def _prompt_callback(token_id: int) -> bool:
+    def _prompt_callback(token_ids: ctypes._Pointer[ctypes.c_int32], n_token_ids: int, cached: bool) -> bool:
         return True

gpt4all-bindings/python/gpt4all/gpt4all.py

@@ -4,38 +4,66 @@ Python only API for running all GPT4All models.
 from __future__ import annotations
 
 import hashlib
+import json
 import os
 import platform
 import re
 import sys
-import time
 import warnings
 from contextlib import contextmanager
+from datetime import datetime
 from pathlib import Path
 from types import TracebackType
-from typing import TYPE_CHECKING, Any, Iterable, Literal, Protocol, overload
+from typing import TYPE_CHECKING, Any, Iterable, Iterator, Literal, NamedTuple, NoReturn, Protocol, TypedDict, overload
 
+import jinja2
 import requests
+from jinja2.sandbox import ImmutableSandboxedEnvironment
 from requests.exceptions import ChunkedEncodingError
 from tqdm import tqdm
 from urllib3.exceptions import IncompleteRead, ProtocolError
 
 from ._pyllmodel import (CancellationError as CancellationError, EmbCancelCallbackType, EmbedResult as EmbedResult,
-                         LLModel, ResponseCallbackType, empty_response_callback)
+                         LLModel, ResponseCallbackType, _operator_call, empty_response_callback)
 
 if TYPE_CHECKING:
     from typing_extensions import Self, TypeAlias
 
-if sys.platform == 'darwin':
+if sys.platform == "darwin":
     import fcntl
 
 # TODO: move to config
 DEFAULT_MODEL_DIRECTORY = Path.home() / ".cache" / "gpt4all"
 
-DEFAULT_PROMPT_TEMPLATE = "### Human:\n{0}\n\n### Assistant:\n"
+ConfigType: TypeAlias = "dict[str, Any]"
 
-ConfigType: TypeAlias = 'dict[str, Any]'
-MessageType: TypeAlias = 'dict[str, str]'
+# Environment setup adapted from HF transformers
+@_operator_call
+def _jinja_env() -> ImmutableSandboxedEnvironment:
+    def raise_exception(message: str) -> NoReturn:
+        raise jinja2.exceptions.TemplateError(message)
+
+    def tojson(obj: Any, indent: int | None = None) -> str:
+        return json.dumps(obj, ensure_ascii=False, indent=indent)
+
+    def strftime_now(fmt: str) -> str:
+        return datetime.now().strftime(fmt)
+
+    env = ImmutableSandboxedEnvironment(trim_blocks=True, lstrip_blocks=True)
+    env.filters["tojson"         ] = tojson
+    env.globals["raise_exception"] = raise_exception
+    env.globals["strftime_now"   ] = strftime_now
+    return env
+
+
+class MessageType(TypedDict):
+    role: str
+    content: str
+
+
+class ChatSession(NamedTuple):
+    template: jinja2.Template
+    history: list[MessageType]
 
 
 class Embed4All:
@@ -55,7 +83,7 @@ class Embed4All:
             kwargs: Remaining keyword arguments are passed to the `GPT4All` constructor.
         """
         if model_name is None:
-            model_name = 'all-MiniLM-L6-v2.gguf2.f16.gguf'
+            model_name = "all-MiniLM-L6-v2.gguf2.f16.gguf"
         self.gpt4all = GPT4All(model_name, n_threads=n_threads, device=device, **kwargs)
 
     def __enter__(self) -> Self:
@@ -146,18 +174,18 @@ class Embed4All:
             dimensionality = -1
         else:
             if dimensionality <= 0:
-                raise ValueError(f'Dimensionality must be None or a positive integer, got {dimensionality}')
+                raise ValueError(f"Dimensionality must be None or a positive integer, got {dimensionality}")
             if dimensionality < self.MIN_DIMENSIONALITY:
                 warnings.warn(
-                    f'Dimensionality {dimensionality} is less than the suggested minimum of {self.MIN_DIMENSIONALITY}.'
-                    ' Performance may be degraded.'
+                    f"Dimensionality {dimensionality} is less than the suggested minimum of {self.MIN_DIMENSIONALITY}."
+                    " Performance may be degraded."
                 )
         try:
             do_mean = {"mean": True, "truncate": False}[long_text_mode]
         except KeyError:
             raise ValueError(f"Long text mode must be one of 'mean' or 'truncate', got {long_text_mode!r}")
         result = self.gpt4all.model.generate_embeddings(text, prefix, dimensionality, do_mean, atlas, cancel_cb)
-        return result if return_dict else result['embeddings']
+        return result if return_dict else result["embeddings"]
 
 
 class GPT4All:
@@ -205,8 +233,7 @@ class GPT4All:
         """
 
         self.model_type = model_type
-        self._history: list[MessageType] | None = None
-        self._current_prompt_template: str = "{0}"
+        self._chat_session: ChatSession | None = None
 
         device_init = None
         if sys.platform == "darwin":
@@ -265,7 +292,13 @@ class GPT4All:
 
     @property
     def current_chat_session(self) -> list[MessageType] | None:
-        return None if self._history is None else list(self._history)
+        return None if self._chat_session is None else self._chat_session.history
+
+    @current_chat_session.setter
+    def current_chat_session(self, history: list[MessageType]) -> None:
+        if self._chat_session is None:
+            raise ValueError("current_chat_session may only be set when there is an active chat session")
+        self._chat_session.history[:] = history
 
     @staticmethod
     def list_models() -> list[ConfigType]:
@@ -277,7 +310,7 @@ class GPT4All:
         """
         resp = requests.get("https://gpt4all.io/models/models3.json")
         if resp.status_code != 200:
-            raise ValueError(f'Request failed: HTTP {resp.status_code} {resp.reason}')
+            raise ValueError(f"Request failed: HTTP {resp.status_code} {resp.reason}")
         return resp.json()
 
     @classmethod
@@ -307,15 +340,9 @@ class GPT4All:
         # get the config for the model
         config: ConfigType = {}
         if allow_download:
-            available_models = cls.list_models()
-
-            for m in available_models:
-                if model_filename == m["filename"]:
-                    tmpl = m.get("promptTemplate", DEFAULT_PROMPT_TEMPLATE)
-                    # change to Python-style formatting
-                    m["promptTemplate"] = tmpl.replace("%1", "{0}", 1).replace("%2", "{1}", 1)
-                    config.update(m)
-                    break
+            models = cls.list_models()
+            if (model := next((m for m in models if m["filename"] == model_filename), None)) is not None:
+                config.update(model)
 
         # Validate download directory
         if model_path is None:
@@ -379,13 +406,13 @@ class GPT4All:
             headers = {}
             if offset:
                 print(f"\nDownload interrupted, resuming from byte position {offset}", file=sys.stderr)
-                headers['Range'] = f'bytes={offset}-'  # resume incomplete response
+                headers["Range"] = f"bytes={offset}-"  # resume incomplete response
                 headers["Accept-Encoding"] = "identity"  # Content-Encoding changes meaning of ranges
             response = requests.get(url, stream=True, headers=headers)
             if response.status_code not in (200, 206):
-                raise ValueError(f'Request failed: HTTP {response.status_code} {response.reason}')
-            if offset and (response.status_code != 206 or str(offset) not in response.headers.get('Content-Range', '')):
-                raise ValueError('Connection was interrupted and server does not support range requests')
+                raise ValueError(f"Request failed: HTTP {response.status_code} {response.reason}")
+            if offset and (response.status_code != 206 or str(offset) not in response.headers.get("Content-Range", "")):
+                raise ValueError("Connection was interrupted and server does not support range requests")
             if (enc := response.headers.get("Content-Encoding")) is not None:
                 raise ValueError(f"Expected identity Content-Encoding, got {enc}")
             return response
@@ -484,19 +511,19 @@ class GPT4All:
 
     def generate(
         self,
-        prompt: str,
+        prompt         : str,
         *,
-        max_tokens: int = 200,
-        temp: float = 0.7,
-        top_k: int = 40,
-        top_p: float = 0.4,
-        min_p: float = 0.0,
-        repeat_penalty: float = 1.18,
-        repeat_last_n: int = 64,
-        n_batch: int = 8,
-        n_predict: int | None = None,
-        streaming: bool = False,
-        callback: ResponseCallbackType = empty_response_callback,
+        max_tokens     : int                  = 200,
+        temp           : float                = 0.7,
+        top_k          : int                  = 40,
+        top_p          : float                = 0.4,
+        min_p          : float                = 0.0,
+        repeat_penalty : float                = 1.18,
+        repeat_last_n  : int                  = 64,
+        n_batch        : int                  = 8,
+        n_predict      : int | None           = None,
+        streaming      : bool                 = False,
+        callback       : ResponseCallbackType = empty_response_callback,
     ) -> Any:
         """
         Generate outputs from any GPT4All model.
@@ -521,122 +548,94 @@ class GPT4All:
 
         # Preparing the model request
         generate_kwargs: dict[str, Any] = dict(
-            temp=temp,
-            top_k=top_k,
-            top_p=top_p,
-            min_p=min_p,
-            repeat_penalty=repeat_penalty,
-            repeat_last_n=repeat_last_n,
-            n_batch=n_batch,
-            n_predict=n_predict if n_predict is not None else max_tokens,
+            temp           = temp,
+            top_k          = top_k,
+            top_p          = top_p,
+            min_p          = min_p,
+            repeat_penalty = repeat_penalty,
+            repeat_last_n  = repeat_last_n,
+            n_batch        = n_batch,
+            n_predict      = n_predict if n_predict is not None else max_tokens,
         )
 
-        if self._history is not None:
-            # check if there is only one message, i.e. system prompt:
-            reset = len(self._history) == 1
-            self._history.append({"role": "user", "content": prompt})
-
-            fct_func = self._format_chat_prompt_template.__func__  # type: ignore[attr-defined]
-            if fct_func is GPT4All._format_chat_prompt_template:
-                if reset:
-                    # ingest system prompt
-                    # use "%1%2" and not "%1" to avoid implicit whitespace
-                    self.model.prompt_model(self._history[0]["content"], "%1%2",
-                                            empty_response_callback,
-                                            n_batch=n_batch, n_predict=0, reset_context=True, special=True)
-                prompt_template = self._current_prompt_template.format("%1", "%2")
-            else:
-                warnings.warn(
-                    "_format_chat_prompt_template is deprecated. Please use a chat session with a prompt template.",
-                    DeprecationWarning,
-                )
-                # special tokens won't be processed
-                prompt = self._format_chat_prompt_template(
-                    self._history[-1:],
-                    self._history[0]["content"] if reset else "",
-                )
-                prompt_template = "%1"
-                generate_kwargs["reset_context"] = reset
-        else:
-            prompt_template = "%1"
-            generate_kwargs["reset_context"] = True
-
         # Prepare the callback, process the model response
-        output_collector: list[MessageType]
-        output_collector = [
-            {"content": ""}
-        ]  # placeholder for the self._history if chat session is not activated
+        full_response = ""
 
-        if self._history is not None:
-            self._history.append({"role": "assistant", "content": ""})
-            output_collector = self._history
+        def _callback_wrapper(token_id: int, response: str) -> bool:
+            nonlocal full_response
+            full_response += response
+            return callback(token_id, response)
 
-        def _callback_wrapper(
-            callback: ResponseCallbackType,
-            output_collector: list[MessageType],
-        ) -> ResponseCallbackType:
-            def _callback(token_id: int, response: str) -> bool:
-                nonlocal callback, output_collector
+        last_msg_rendered = prompt
+        if self._chat_session is not None:
+            session = self._chat_session
+            def render(messages: list[MessageType]) -> str:
+                return session.template.render(
+                    messages=messages,
+                    add_generation_prompt=True,
+                    **self.model.special_tokens_map,
+                )
+            session.history.append(MessageType(role="user", content=prompt))
+            prompt = render(session.history)
+            if len(session.history) > 1:
+                last_msg_rendered = render(session.history[-1:])
 
-                output_collector[-1]["content"] += response
-
-                return callback(token_id, response)
-
-            return _callback
+        # Check request length
+        last_msg_len = self.model.count_prompt_tokens(last_msg_rendered)
+        if last_msg_len > (limit := self.model.n_ctx - 4):
+            raise ValueError(f"Your message was too long and could not be processed ({last_msg_len} > {limit}).")
 
         # Send the request to the model
         if streaming:
-            return self.model.prompt_model_streaming(
-                prompt,
-                prompt_template,
-                _callback_wrapper(callback, output_collector),
-                **generate_kwargs,
-            )
+            def stream() -> Iterator[str]:
+                yield from self.model.prompt_model_streaming(prompt, _callback_wrapper, **generate_kwargs)
+                if self._chat_session is not None:
+                    self._chat_session.history.append(MessageType(role="assistant", content=full_response))
+            return stream()
 
-        self.model.prompt_model(
-            prompt,
-            prompt_template,
-            _callback_wrapper(callback, output_collector),
-            **generate_kwargs,
-        )
-
-        return output_collector[-1]["content"]
+        self.model.prompt_model(prompt, _callback_wrapper, **generate_kwargs)
+        if self._chat_session is not None:
+            self._chat_session.history.append(MessageType(role="assistant", content=full_response))
+        return full_response
 
     @contextmanager
     def chat_session(
         self,
-        system_prompt: str | None = None,
-        prompt_template: str | None = None,
+        system_message: str | Literal[False] | None = None,
+        chat_template: str | None = None,
     ):
         """
         Context manager to hold an inference optimized chat session with a GPT4All model.
 
         Args:
-            system_prompt: An initial instruction for the model.
-            prompt_template: Template for the prompts with {0} being replaced by the user message.
+            system_message: An initial instruction for the model, None to use the model default, or False to disable. Defaults to None.
+            chat_template: Jinja template for the conversation, or None to use the model default. Defaults to None.
         """
 
-        if system_prompt is None:
-            system_prompt = self.config.get("systemPrompt", "")
+        if system_message is None:
+            system_message = self.config.get("systemMessage", False)
 
-        if prompt_template is None:
-            if (tmpl := self.config.get("promptTemplate")) is None:
-                warnings.warn("Use of a sideloaded model or allow_download=False without specifying a prompt template "
-                              "is deprecated. Defaulting to Alpaca.", DeprecationWarning)
-                tmpl = DEFAULT_PROMPT_TEMPLATE
-            prompt_template = tmpl
+        if chat_template is None:
+            if "name" not in self.config:
+                raise ValueError("For sideloaded models or with allow_download=False, you must specify a chat template.")
+            if "chatTemplate" not in self.config:
+                raise NotImplementedError("This model appears to have a built-in chat template, but loading it is not "
+                                          "currently implemented. Please pass a template to chat_session() directly.")
+            if (tmpl := self.config["chatTemplate"]) is None:
+                raise ValueError(f"The model {self.config['name']!r} does not support chat.")
+            chat_template = tmpl
 
-        if re.search(r"%1(?![0-9])", prompt_template):
-            raise ValueError("Prompt template containing a literal '%1' is not supported. For a prompt "
-                             "placeholder, please use '{0}' instead.")
-
-        self._history = [{"role": "system", "content": system_prompt}]
-        self._current_prompt_template = prompt_template
+        history = []
+        if system_message is not False:
+            history.append(MessageType(role="system", content=system_message))
+        self._chat_session = ChatSession(
+            template=_jinja_env.from_string(chat_template),
+            history=history,
+        )
         try:
             yield self
         finally:
-            self._history = None
-            self._current_prompt_template = "{0}"
+            self._chat_session = None
 
     @staticmethod
     def list_gpus() -> list[str]:
@@ -648,43 +647,6 @@ class GPT4All:
         """
         return LLModel.list_gpus()
 
-    def _format_chat_prompt_template(
-        self,
-        messages: list[MessageType],
-        default_prompt_header: str = "",
-        default_prompt_footer: str = "",
-    ) -> str:
-        """
-        Helper method for building a prompt from list of messages using the self._current_prompt_template as a template for each message.
-
-        Warning:
-            This function was deprecated in version 2.3.0, and will be removed in a future release.
-
-        Args:
-            messages:  List of dictionaries. Each dictionary should have a "role" key
-                with value of "system", "assistant", or "user" and a "content" key with a
-                string value. Messages are organized such that "system" messages are at top of prompt,
-                and "user" and "assistant" messages are displayed in order. Assistant messages get formatted as
-                "Response: {content}".
-
-        Returns:
-            Formatted prompt.
-        """
-
-        full_prompt = default_prompt_header + "\n\n" if default_prompt_header != "" else ""
-
-        for message in messages:
-            if message["role"] == "user":
-                user_message = self._current_prompt_template.format(message["content"])
-                full_prompt += user_message
-            if message["role"] == "assistant":
-                assistant_message = message["content"] + "\n"
-                full_prompt += assistant_message
-
-        full_prompt += "\n\n" + default_prompt_footer if default_prompt_footer != "" else ""
-
-        return full_prompt
-
 
 def append_extension_if_missing(model_name):
     if not model_name.endswith((".bin", ".gguf")):
@@ -697,7 +659,7 @@ class _HasFileno(Protocol):
 
 
 def _fsync(fd: int | _HasFileno) -> None:
-    if sys.platform == 'darwin':
+    if sys.platform == "darwin":
         # Apple's fsync does not flush the drive write cache
         try:
             fcntl.fcntl(fd, fcntl.F_FULLFSYNC)

gpt4all-bindings/python/mkdocs.yml

@@ -14,7 +14,9 @@ nav:
     - 'Models' : 'gpt4all_desktop/models.md'
     - 'LocalDocs' : 'gpt4all_desktop/localdocs.md'
     - 'Settings' : 'gpt4all_desktop/settings.md'
+    - 'Chat Templates' : 'gpt4all_desktop/chat_templates.md'
     - 'Cookbook':
+      - 'Local AI Chat with Microsoft Excel': 'gpt4all_desktop/cookbook/use-local-ai-models-to-privately-chat-with-microsoft-excel.md'
       - 'Local AI Chat with your Google Drive': 'gpt4all_desktop/cookbook/use-local-ai-models-to-privately-chat-with-google-drive.md'
       - 'Local AI Chat with your Obsidian Vault': 'gpt4all_desktop/cookbook/use-local-ai-models-to-privately-chat-with-Obsidian.md'
       - 'Local AI Chat with your OneDrive': 'gpt4all_desktop/cookbook/use-local-ai-models-to-privately-chat-with-One-Drive.md'

gpt4all-bindings/python/setup.py

@@ -88,9 +88,10 @@ setup(
     python_requires='>=3.8',
     packages=find_packages(),
     install_requires=[
+        'importlib_resources; python_version < "3.9"',
+        'jinja2~=3.1',
         'requests',
         'tqdm',
-        'importlib_resources; python_version < "3.9"',
         'typing-extensions>=4.3.0; python_version >= "3.9" and python_version < "3.11"',
     ],
     extras_require={

gpt4all-chat/.flake8

@@ -0,0 +1,5 @@
+# vim: set syntax=dosini:
+[flake8]
+exclude = .*,__pycache__
+max-line-length = 120
+extend-ignore = B001,C408,D,DAR,E221,E303,E722,E741,E800,N801,N806,P101,S101,S324,S404,S406,S410,S603,WPS100,WPS110,WPS111,WPS113,WPS114,WPS115,WPS120,WPS2,WPS300,WPS301,WPS304,WPS305,WPS306,WPS309,WPS316,WPS317,WPS318,WPS319,WPS322,WPS323,WPS326,WPS329,WPS330,WPS332,WPS336,WPS337,WPS347,WPS360,WPS361,WPS407,WPS414,WPS420,WPS421,WPS429,WPS430,WPS431,WPS432,WPS433,WPS437,WPS440,WPS440,WPS441,WPS442,WPS457,WPS458,WPS460,WPS462,WPS463,WPS473,WPS501,WPS504,WPS505,WPS508,WPS509,WPS510,WPS515,WPS516,WPS519,WPS520,WPS529,WPS531,WPS602,WPS604,WPS605,WPS608,WPS609,WPS613,WPS615

gpt4all-chat/CHANGELOG.md

@@ -4,24 +4,59 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
-## [Unreleased]
+## [3.5.0-rc1] - 2024-12-04
+
+### Added
+- Add ability to attach text, markdown, and rst files to chat ([#3135](https://github.com/nomic-ai/gpt4all/pull/3135))
+- Add feature to minimize to system tray (by [@bgallois](https://github.com/bgallois) in ([#3109](https://github.com/nomic-ai/gpt4all/pull/3109))
+- Basic cache for faster prefill when the input shares a prefix with previous context ([#3073](https://github.com/nomic-ai/gpt4all/pull/3073))
+
+### Changed
+- Implement Qt 6.8 compatibility ([#3121](https://github.com/nomic-ai/gpt4all/pull/3121))
+
+### Fixed
+- Fix bug in GUI when localdocs encounters binary data ([#3137](https://github.com/nomic-ai/gpt4all/pull/3137))
+- Fix LocalDocs bugs that prevented some docx files from fully chunking ([#3140](https://github.com/nomic-ai/gpt4all/pull/3140))
+- Fix missing softmax that was causing crashes and effectively infinite temperature since 3.4.0 ([#3202](https://github.com/nomic-ai/gpt4all/pull/3202))
+
+## [3.4.2] - 2024-10-16
+
+### Fixed
+- Limit bm25 retrieval to only specified collections ([#3083](https://github.com/nomic-ai/gpt4all/pull/3083))
+- Fix bug removing documents because of a wrong case sensitive file suffix check ([#3083](https://github.com/nomic-ai/gpt4all/pull/3083))
+- Fix bug with hybrid localdocs search where database would get out of sync ([#3083](https://github.com/nomic-ai/gpt4all/pull/3083))
+- Fix GUI bug where the localdocs embedding device appears blank ([#3083](https://github.com/nomic-ai/gpt4all/pull/3083))
+- Prevent LocalDocs from not making progress in certain cases ([#3094](https://github.com/nomic-ai/gpt4all/pull/3094))
+
+## [3.4.1] - 2024-10-11
+
+### Fixed
+- Improve the Italian translation ([#3048](https://github.com/nomic-ai/gpt4all/pull/3048))
+- Fix models.json cache location ([#3052](https://github.com/nomic-ai/gpt4all/pull/3052))
+- Fix LocalDocs regressions caused by docx change ([#3079](https://github.com/nomic-ai/gpt4all/pull/3079))
+- Fix Go code being highlighted as Java ([#3080](https://github.com/nomic-ai/gpt4all/pull/3080))
+
+## [3.4.0] - 2024-10-08
 
 ### Added
 - Add bm25 hybrid search to localdocs ([#2969](https://github.com/nomic-ai/gpt4all/pull/2969))
 - LocalDocs support for .docx files ([#2986](https://github.com/nomic-ai/gpt4all/pull/2986))
-- Add support for attaching Excel spreadsheet to chat ([#3007](https://github.com/nomic-ai/gpt4all/pull/3007))
+- Add support for attaching Excel spreadsheet to chat ([#3007](https://github.com/nomic-ai/gpt4all/pull/3007), [#3028](https://github.com/nomic-ai/gpt4all/pull/3028))
 
 ### Changed
 - Rebase llama.cpp on latest upstream as of September 26th ([#2998](https://github.com/nomic-ai/gpt4all/pull/2998))
 - Change the error message when a message is too long ([#3004](https://github.com/nomic-ai/gpt4all/pull/3004))
 - Simplify chatmodel to get rid of unnecessary field and bump chat version ([#3016](https://github.com/nomic-ai/gpt4all/pull/3016))
 - Allow ChatLLM to have direct access to ChatModel for restoring state from text ([#3018](https://github.com/nomic-ai/gpt4all/pull/3018))
+- Improvements to XLSX conversion and UI fix ([#3022](https://github.com/nomic-ai/gpt4all/pull/3022))
 
 ### Fixed
 - Fix a crash when attempting to continue a chat loaded from disk ([#2995](https://github.com/nomic-ai/gpt4all/pull/2995))
 - Fix the local server rejecting min\_p/top\_p less than 1 ([#2996](https://github.com/nomic-ai/gpt4all/pull/2996))
 - Fix "regenerate" always forgetting the most recent message ([#3011](https://github.com/nomic-ai/gpt4all/pull/3011))
 - Fix loaded chats forgetting context when there is a system prompt ([#3015](https://github.com/nomic-ai/gpt4all/pull/3015))
+- Make it possible to downgrade and keep some chats, and avoid crash for some model types ([#3030](https://github.com/nomic-ai/gpt4all/pull/3030))
+- Fix scroll positition being reset in model view, and attempt a better fix for the clone issue ([#3042](https://github.com/nomic-ai/gpt4all/pull/3042))
 
 ## [3.3.1] - 2024-09-27 ([v3.3.y](https://github.com/nomic-ai/gpt4all/tree/v3.3.y))
 
@@ -144,7 +179,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - Fix several Vulkan resource management issues ([#2694](https://github.com/nomic-ai/gpt4all/pull/2694))
 - Fix crash/hang when some models stop generating, by showing special tokens ([#2701](https://github.com/nomic-ai/gpt4all/pull/2701))
 
-[Unreleased]: https://github.com/nomic-ai/gpt4all/compare/v3.3.0...HEAD
+[3.5.0-rc1]: https://github.com/nomic-ai/gpt4all/compare/v3.4.2...v3.5.0-rc1
+[3.4.2]: https://github.com/nomic-ai/gpt4all/compare/v3.4.1...v3.4.2
+[3.4.1]: https://github.com/nomic-ai/gpt4all/compare/v3.4.0...v3.4.1
+[3.4.0]: https://github.com/nomic-ai/gpt4all/compare/v3.3.0...v3.4.0
 [3.3.1]: https://github.com/nomic-ai/gpt4all/compare/v3.3.0...v3.3.1
 [3.3.0]: https://github.com/nomic-ai/gpt4all/compare/v3.2.1...v3.3.0
 [3.2.1]: https://github.com/nomic-ai/gpt4all/compare/v3.2.0...v3.2.1

gpt4all-chat/CMakeLists.txt

@@ -1,10 +1,12 @@
 cmake_minimum_required(VERSION 3.25)  # for try_compile SOURCE_FROM_VAR
 
+include(../common/common.cmake)
+
 set(APP_VERSION_MAJOR 3)
-set(APP_VERSION_MINOR 4)
+set(APP_VERSION_MINOR 5)
 set(APP_VERSION_PATCH 0)
 set(APP_VERSION_BASE "${APP_VERSION_MAJOR}.${APP_VERSION_MINOR}.${APP_VERSION_PATCH}")
-set(APP_VERSION "${APP_VERSION_BASE}-dev0")
+set(APP_VERSION "${APP_VERSION_BASE}-rc1")
 
 project(gpt4all VERSION ${APP_VERSION_BASE} LANGUAGES CXX C)
 
@@ -20,6 +22,9 @@ if(APPLE)
   endif()
 endif()
 
+find_package(Python3 3.12 QUIET COMPONENTS Interpreter)
+
+option(GPT4ALL_TEST "Build the tests" ${Python3_FOUND})
 option(GPT4ALL_LOCALHOST "Build installer for localhost repo" OFF)
 option(GPT4ALL_OFFLINE_INSTALLER "Build an offline installer" OFF)
 option(GPT4ALL_SIGN_INSTALL "Sign installed binaries and installers (requires signing identities)" OFF)
@@ -28,6 +33,10 @@ option(GPT4ALL_SIGN_INSTALL "Sign installed binaries and installers (requires si
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 set(CMAKE_CXX_STANDARD 23)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
+if (MSVC)
+    # Enable accurate __cplusplus macro to fix errors in Jinja2Cpp
+    add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/Zc:__cplusplus>)
+endif()
 
 
 # conftests
@@ -70,7 +79,12 @@ configure_file(
   "${CMAKE_CURRENT_BINARY_DIR}/config.h"
 )
 
-find_package(Qt6 6.4 COMPONENTS Core HttpServer LinguistTools Pdf Quick QuickDialogs2 Sql Svg REQUIRED)
+set(CMAKE_FIND_PACKAGE_TARGETS_GLOBAL ON)
+find_package(Qt6 6.5 COMPONENTS Core HttpServer LinguistTools Pdf Quick QuickDialogs2 Sql Svg REQUIRED)
+
+if (QT_KNOWN_POLICY_QTP0004)
+    qt_policy(SET QTP0004 NEW)  # generate extra qmldir files on Qt 6.8+
+endif()
 
 # Get the Qt6Core target properties
 get_target_property(Qt6Core_INCLUDE_DIRS Qt6::Core INTERFACE_INCLUDE_DIRECTORIES)
@@ -88,9 +102,41 @@ message(STATUS "Qt 6 root directory: ${Qt6_ROOT_DIR}")
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
+if (CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
+  set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install CACHE PATH "..." FORCE)
+endif()
+
 add_subdirectory(deps)
 add_subdirectory(../gpt4all-backend llmodel)
 
+if (GPT4ALL_TEST)
+    enable_testing()
+
+    # Llama-3.2-1B model
+    set(TEST_MODEL "Llama-3.2-1B-Instruct-Q4_0.gguf")
+    set(TEST_MODEL_MD5 "48ff0243978606fdba19d899b77802fc")
+    set(TEST_MODEL_PATH "${CMAKE_BINARY_DIR}/resources/${TEST_MODEL}")
+    set(TEST_MODEL_URL "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/${TEST_MODEL}")
+
+    # Create a custom command to download the file if it does not exist or if the checksum does not match
+    add_custom_command(
+        OUTPUT "${TEST_MODEL_PATH}"
+        COMMAND ${CMAKE_COMMAND} -E echo "Downloading test model from ${TEST_MODEL_URL} ..."
+        COMMAND ${CMAKE_COMMAND} -DURL="${TEST_MODEL_URL}" -DOUTPUT_PATH="${TEST_MODEL_PATH}" -DEXPECTED_MD5="${TEST_MODEL_MD5}" -P "${CMAKE_SOURCE_DIR}/cmake/download_model.cmake"
+        DEPENDS "${CMAKE_SOURCE_DIR}/cmake/download_model.cmake"
+    )
+
+    # Define a custom target that depends on the downloaded model
+    add_custom_target(download_test_model
+        DEPENDS "${TEST_MODEL_PATH}"
+    )
+
+    add_subdirectory(tests)
+
+    # The 'check' target makes sure the tests and their dependencies are up-to-date before running them
+    add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure DEPENDS download_test_model chat gpt4all_tests)
+endif()
+
 set(CHAT_EXE_RESOURCES)
 
 # Metal shader library
@@ -135,6 +181,12 @@ if (APPLE)
     set_source_files_properties(${CHAT_EXE_RESOURCES} PROPERTIES MACOSX_PACKAGE_LOCATION Resources)
 endif()
 
+set(MACOS_SOURCES)
+if (APPLE)
+    find_library(COCOA_LIBRARY Cocoa)
+    list(APPEND MACOS_SOURCES src/macosdock.mm src/macosdock.h)
+endif()
+
 qt_add_executable(chat
     src/main.cpp
     src/chat.cpp                  src/chat.h
@@ -146,6 +198,7 @@ qt_add_executable(chat
     src/database.cpp              src/database.h
     src/download.cpp              src/download.h
     src/embllm.cpp                src/embllm.h
+    src/jinja_helpers.cpp         src/jinja_helpers.h
     src/llm.cpp                   src/llm.h
     src/localdocs.cpp             src/localdocs.h
     src/localdocsmodel.cpp        src/localdocsmodel.h
@@ -156,7 +209,9 @@ qt_add_executable(chat
     src/server.cpp                src/server.h
     src/xlsxtomd.cpp              src/xlsxtomd.h
     ${CHAT_EXE_RESOURCES}
+    ${MACOS_SOURCES}
 )
+gpt4all_add_warning_options(chat)
 
 qt_add_qml_module(chat
     URI gpt4all
@@ -168,6 +223,8 @@ qt_add_qml_module(chat
       qml/AddModelView.qml
       qml/ApplicationSettings.qml
       qml/ChatDrawer.qml
+      qml/ChatItemView.qml
+      qml/ChatMessageButton.qml
       qml/ChatView.qml
       qml/CollectionsDrawer.qml
       qml/HomeView.qml
@@ -180,7 +237,7 @@ qt_add_qml_module(chat
       qml/PopupDialog.qml
       qml/SettingsView.qml
       qml/StartupDialog.qml
-      qml/SwitchModelDialog.qml
+      qml/ConfirmationDialog.qml
       qml/Theme.qml
       qml/ThumbsDownDialog.qml
       qml/Toast.qml
@@ -192,6 +249,7 @@ qt_add_qml_module(chat
       qml/MyDialog.qml
       qml/MyDirectoryField.qml
       qml/MyFileDialog.qml
+      qml/MyFileIcon.qml
       qml/MyFolderDialog.qml
       qml/MyFancyLink.qml
       qml/MyMenu.qml
@@ -226,6 +284,7 @@ qt_add_qml_module(chat
       icons/eject.svg
       icons/email.svg
       icons/file-doc.svg
+      icons/file-docx.svg
       icons/file-md.svg
       icons/file-pdf.svg
       icons/file-txt.svg
@@ -297,15 +356,17 @@ if (APPLE)
         OUTPUT_NAME gpt4all
     )
     add_dependencies(chat ggml-metal)
+endif()
 
-    if(NOT MAC_SIGNING_IDENTITY)
-        if(NOT DEFINED ENV{MAC_SIGNING_CERT_NAME} AND GPT4ALL_SIGN_INSTALL)
+if (APPLE AND GPT4ALL_SIGN_INSTALL)
+    if (NOT MAC_SIGNING_IDENTITY)
+        if (NOT DEFINED ENV{MAC_SIGNING_CERT_NAME})
             REPORT_MISSING_SIGNING_CONTEXT()
         endif()
         set(MAC_SIGNING_IDENTITY $ENV{MAC_SIGNING_CERT_NAME})
     endif()
-    if(NOT MAC_SIGNING_TID)
-        if(NOT DEFINED ENV{MAC_NOTARIZATION_TID} AND GPT4ALL_SIGN_INSTALL)
+    if (NOT MAC_SIGNING_TID)
+        if (NOT DEFINED ENV{MAC_NOTARIZATION_TID})
             REPORT_MISSING_SIGNING_CONTEXT()
         endif()
         set(MAC_SIGNING_TID $ENV{MAC_NOTARIZATION_TID})
@@ -335,17 +396,16 @@ target_include_directories(chat PRIVATE deps/usearch/include
 target_link_libraries(chat
     PRIVATE Qt6::Core Qt6::HttpServer Qt6::Pdf Qt6::Quick Qt6::Sql Qt6::Svg)
 target_link_libraries(chat
-    PRIVATE llmodel SingleApplication fmt::fmt duckx::duckx QXlsx)
+    PRIVATE llmodel SingleApplication fmt::fmt duckx::duckx QXlsx jinja2cpp)
 
+if (APPLE)
+    target_link_libraries(chat PRIVATE ${COCOA_LIBRARY})
+endif()
 
 # -- install --
 
 set(COMPONENT_NAME_MAIN ${PROJECT_NAME})
 
-if(CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
-  set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install CACHE PATH "..." FORCE)
-endif()
-
 install(TARGETS chat DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN})
 
 install(
@@ -458,6 +518,7 @@ elseif(${CMAKE_SYSTEM_NAME} MATCHES Darwin)
     set(CPACK_BUNDLE_ICON "${CMAKE_CURRENT_SOURCE_DIR}/resources/gpt4all.icns")
 endif()
 
+set(CPACK_COMPONENTS_ALL gpt4all)  # exclude development components
 set(CPACK_PACKAGE_INSTALL_DIRECTORY ${COMPONENT_NAME_MAIN})
 set(CPACK_PACKAGE_VERSION_MAJOR ${PROJECT_VERSION_MAJOR})
 set(CPACK_PACKAGE_VERSION_MINOR ${PROJECT_VERSION_MINOR})

gpt4all-chat/cmake/deploy-qt-mac.cmake.in

@@ -2,7 +2,10 @@ set(MACDEPLOYQT "@MACDEPLOYQT@")
 set(COMPONENT_NAME_MAIN "@COMPONENT_NAME_MAIN@")
 set(CMAKE_CURRENT_SOURCE_DIR "@CMAKE_CURRENT_SOURCE_DIR@")
 set(GPT4ALL_SIGNING_ID "@MAC_SIGNING_IDENTITY@")
-execute_process(COMMAND ${MACDEPLOYQT} ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app -qmldir=${CMAKE_CURRENT_SOURCE_DIR} -verbose=2 -sign-for-notarization=${GPT4ALL_SIGNING_ID})
+if (GPT4ALL_SIGNING_ID)
+    set(MAC_NOTARIZE -sign-for-notarization=${GPT4ALL_SIGNING_ID})
+endif()
+execute_process(COMMAND ${MACDEPLOYQT} ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app -qmldir=${CMAKE_CURRENT_SOURCE_DIR} -verbose=2 ${MAC_NOTARIZE})
 file(GLOB MYLLAMALIBS ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/lib/libllama*)
 file(GLOB MYLLMODELLIBS ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/lib/libllmodel.*)
 file(COPY ${MYLLAMALIBS}

gpt4all-chat/cmake/download_model.cmake

@@ -0,0 +1,12 @@
+if(NOT DEFINED URL OR NOT DEFINED OUTPUT_PATH OR NOT DEFINED EXPECTED_MD5)
+    message(FATAL_ERROR "Usage: cmake -DURL=<url> -DOUTPUT_PATH=<path> -DEXPECTED_MD5=<md5> -P download_model.cmake")
+endif()
+
+message(STATUS "Downloading model from ${URL} to ${OUTPUT_PATH} ...")
+
+file(DOWNLOAD "${URL}" "${OUTPUT_PATH}" EXPECTED_MD5 "${EXPECTED_MD5}" STATUS status)
+
+list(GET status 0 status_code)
+if(NOT status_code EQUAL 0)
+    message(FATAL_ERROR "Failed to download model: ${status}")
+endif()

gpt4all-chat/deps/CMakeLists.txt

@@ -3,7 +3,7 @@ set(BUILD_SHARED_LIBS OFF)
 set(FMT_INSTALL OFF)
 add_subdirectory(fmt)
 
-set(QAPPLICATION_CLASS QGuiApplication)
+set(QAPPLICATION_CLASS QApplication)
 add_subdirectory(SingleApplication)
 
 set(DUCKX_INSTALL OFF)
@@ -11,3 +11,12 @@ add_subdirectory(DuckX)
 
 set(QT_VERSION_MAJOR 6)
 add_subdirectory(QXlsx/QXlsx)
+
+# forked dependency of Jinja2Cpp
+set(RAPIDJSON_BUILD_DOC                  OFF)
+set(RAPIDJSON_BUILD_EXAMPLES             OFF)
+set(RAPIDJSON_BUILD_TESTS                OFF)
+set(RAPIDJSON_ENABLE_INSTRUMENTATION_OPT OFF)
+add_subdirectory(rapidjson)
+
+add_subdirectory(Jinja2Cpp)

gpt4all-chat/dev-requirements.txt

@@ -0,0 +1,11 @@
+-r test-requirements.txt
+
+# dev tools
+flake8~=7.1
+mypy~=1.12
+pytype>=2024.10.11
+wemake-python-styleguide~=0.19.2
+
+# type stubs and other optional modules
+types-requests~=2.32
+urllib3[socks]

gpt4all-chat/icons/edit.svg

@@ -1,3 +1 @@
-<svg width="32" height="32" viewBox="0 0 32 32" fill="none" xmlns="http://www.w3.org/2000/svg">
-<path d="M28.4138 9.17125L22.8288 3.585C22.643 3.39924 22.4225 3.25188 22.1799 3.15134C21.9372 3.0508 21.6771 2.99905 21.4144 2.99905C21.1517 2.99905 20.8916 3.0508 20.6489 3.15134C20.4062 3.25188 20.1857 3.39924 20 3.585L4.58626 19C4.39973 19.185 4.25185 19.4053 4.15121 19.648C4.05057 19.8907 3.99917 20.151 4.00001 20.4138V26C4.00001 26.5304 4.21072 27.0391 4.5858 27.4142C4.96087 27.7893 5.46958 28 6.00001 28H11.5863C11.849 28.0008 12.1093 27.9494 12.352 27.8488C12.5947 27.7482 12.815 27.6003 13 27.4138L28.4138 12C28.5995 11.8143 28.7469 11.5938 28.8474 11.3511C28.948 11.1084 28.9997 10.8483 28.9997 10.5856C28.9997 10.3229 28.948 10.0628 28.8474 9.82015C28.7469 9.57747 28.5995 9.35698 28.4138 9.17125ZM6.41376 20L17 9.41375L19.0863 11.5L8.50001 22.085L6.41376 20ZM6.00001 22.4138L9.58626 26H6.00001V22.4138ZM12 25.5863L9.91376 23.5L20.5 12.9138L22.5863 15L12 25.5863ZM24 13.5863L18.4138 8L21.4138 5L27 10.585L24 13.5863Z" fill="black"/>
-</svg>
+<svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" fill="#000000" viewBox="0 0 256 256"><path d="M227.31,73.37,182.63,28.68a16,16,0,0,0-22.63,0L36.69,152A15.86,15.86,0,0,0,32,163.31V208a16,16,0,0,0,16,16H92.69A15.86,15.86,0,0,0,104,219.31L227.31,96a16,16,0,0,0,0-22.63ZM92.69,208H48V163.31l88-88L180.69,120ZM192,108.68,147.31,64l24-24L216,84.68Z"></path></svg>

gpt4all-chat/icons/file-docx.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 256 256"><rect width="256" height="256" fill="none"/><line x1="152" y1="96" x2="208" y2="96" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="16"/><line x1="152" y1="160" x2="208" y2="160" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="16"/><path d="M64,72V40a8,8,0,0,1,8-8H200a8,8,0,0,1,8,8V216a8,8,0,0,1-8,8H72a8,8,0,0,1-8-8V184" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="16"/><polyline points="64 104 76 152 92 120 108 152 120 104" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="16"/><rect x="32" y="72" width="120" height="112" rx="8" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="16"/></svg>

gpt4all-chat/main.qml

@@ -12,6 +12,7 @@ import network
 import gpt4all
 import localdocs
 import mysettings
+import Qt.labs.platform
 
 Window {
     id: window
@@ -22,6 +23,43 @@ Window {
     visible: true
     title: qsTr("GPT4All v%1").arg(Qt.application.version)
 
+    SystemTrayIcon {
+        id: systemTrayIcon
+        property bool shouldClose: false
+        visible: MySettings.systemTray && !shouldClose
+        icon.source: "qrc:/gpt4all/icons/gpt4all.svg"
+
+        function restore() {
+            LLM.showDockIcon();
+            window.show();
+            window.raise();
+            window.requestActivate();
+        }
+        onActivated: function(reason) {
+            if (reason === SystemTrayIcon.Context && Qt.platform.os !== "osx")
+                menu.open();
+            else if (reason === SystemTrayIcon.Trigger)
+                restore();
+        }
+
+        menu: Menu {
+            MenuItem {
+                text: qsTr("Restore")
+                onTriggered: systemTrayIcon.restore()
+            }
+            MenuItem {
+                text: qsTr("Quit")
+                onTriggered: {
+                    systemTrayIcon.restore();
+                    systemTrayIcon.shouldClose = true;
+                    window.shouldClose = true;
+                    savingPopup.open();
+                    ChatListModel.saveChats();
+                }
+            }
+        }
+    }
+
     Settings {
         property alias x: window.x
         property alias y: window.y
@@ -156,7 +194,7 @@ Window {
         font.pixelSize: theme.fontSizeLarge
     }
 
-    property bool hasSaved: false
+    property bool shouldClose: false
 
     PopupDialog {
         id: savingPopup
@@ -180,9 +218,18 @@ Window {
     }
 
     onClosing: function(close) {
-        if (window.hasSaved)
+        if (systemTrayIcon.visible) {
+            LLM.hideDockIcon();
+            window.visible = false;
+            ChatListModel.saveChats();
+            close.accepted = false;
+            return;
+        }
+
+        if (window.shouldClose)
             return;
 
+        window.shouldClose = true;
         savingPopup.open();
         ChatListModel.saveChats();
         close.accepted = false
@@ -191,9 +238,9 @@ Window {
     Connections {
         target: ChatListModel
         function onSaveChatsFinished() {
-            window.hasSaved = true;
             savingPopup.close();
-            window.close()
+            if (window.shouldClose)
+                window.close()
         }
     }
 

gpt4all-chat/metadata/latestnews.md

@@ -1,11 +1,20 @@
 ## Latest News
 
-GPT4All v3.3.1 from September 26th fixes a crash after loading a chat and an incorrect local server parameter check in v3.3.0.
+GPT4All v3.4.1 was released on October 11th, and fixes several issues with LocalDocs from the previous release.
 
-GPT4All v3.3.0 was released on September 23rd. Changes include:
+GPT4All v3.4.2 was released on October 16th, and fixes more issues with LocalDocs.
 
-* **UI Improvements**: The minimum window size now adapts to the font size. A few labels and links have been fixed. The Embeddings Device selection of "Auto"/"Application default" works again. The window icon is now set on Linux. The antenna icon now displays when the API server is listening.
-* **Single Instance**: Only one instance of GPT4All can be opened at a time. This is now enforced.
-* **Greedy Sampling**: Set temperature to zero to enable greedy sampling.
-* **API Server Changes**: The built-in API server now responds correctly to both legacy completions, and chats with message history. Also, it now uses the system prompt configured in the UI.
-* **Translation Improvements**: The Italian, Romanian, and Traditional Chinese translations have been updated.
+**IMPORTANT NOTE:** If you are coming from v3.4.0, be sure to "Rebuild" your collections at least once after updating!
+
+---
+
+GPT4All v3.4.0 was released on October 8th. Changes include:
+
+* **Attached Files:** You can now attach a small Microsoft Excel spreadsheet (.xlsx) to a chat message and ask the model about it.
+* **LocalDocs Accuracy:** The LocalDocs algorithm has been enhanced to find more accurate references for some queries.
+* **Word Document Support:** LocalDocs now supports Microsoft Word (.docx) documents natively.
+  * **IMPORTANT NOTE:** If .docx files are not found, make sure Settings > LocalDocs > Allowed File Extensions includes "docx".
+* **Forgetful Model Fixes:** Issues with the "Redo last chat response" button, and with continuing chats from previous sessions, have been fixed.
+* **Chat Saving Improvements:** On exit, GPT4All will no longer save chats that are not new or modified. As a bonus, downgrading without losing access to all chats will be possible in the future, should the need arise.
+* **UI Fixes:** The model list no longer scrolls to the top when you start downloading a model.
+* **New Models:** LLama 3.2 Instruct 3B and 1B models now available in model list.

gpt4all-chat/metadata/models3.json

@@ -17,6 +17,40 @@
   },
   {
     "order": "b",
+    "md5sum": "27b44e8ae1817525164ddf4f8dae8af4",
+    "name": "Llama 3.2 3B Instruct",
+    "filename": "Llama-3.2-3B-Instruct-Q4_0.gguf",
+    "filesize": "1921909280",
+    "requires": "3.4.0",
+    "ramrequired": "4",
+    "parameters": "3 billion",
+    "quant": "q4_0",
+    "type": "LLaMA3",
+    "description": "<ul><li>Fast responses</li><li>Instruct model</li><li>Multilingual dialogue use</li><li>Agentic system capable</li><li>Trained by Meta</li><li>License: <a href=\"https://llama.meta.com/llama3_2/license/\">Meta Llama 3.2 Community License</a></li></ul>",
+    "url": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-Q4_0.gguf",
+    "promptTemplate": "<|start_header_id|>user<|end_header_id|>\n\n%1<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n%2",
+    "systemPrompt": "<|start_header_id|>system<|end_header_id|>\nCutting Knowledge Date: December 2023\n\nYou are a helpful assistant.<|eot_id|>",
+    "chatTemplate": "{{- bos_token }}\n{%- if not date_string is defined %}\n    {%- if strftime_now is defined %}\n        {%- set date_string = strftime_now(\"%d %b %Y\") %}\n    {%- else %}\n        {%- set date_string = \"26 Jul 2024\" %}\n    {%- endif %}\n{%- endif %}\n\n{#- This block extracts the system message, so we can slot it into the right place. #}\n{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content']|trim %}\n    {%- set messages = messages[1] %}\n{%- else %}\n    {%- set system_message = \"\" %}\n{%- endif %}\n\n{#- System message #}\n{{- \"<|start_header_id|>system<|end_header_id|>\\n\\n\" }}\n{{- \"Cutting Knowledge Date: December 2023\\n\" }}\n{{- \"Today Date: \" + date_string + \"\\n\\n\" }}\n{{- system_message }}\n{{- \"<|eot_id|>\" }}\n\n{%- for message in messages %}\n    {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}\n"
+  },
+  {
+    "order": "c",
+    "md5sum": "48ff0243978606fdba19d899b77802fc",
+    "name": "Llama 3.2 1B Instruct",
+    "filename": "Llama-3.2-1B-Instruct-Q4_0.gguf",
+    "filesize": "773025920",
+    "requires": "3.4.0",
+    "ramrequired": "2",
+    "parameters": "1 billion",
+    "quant": "q4_0",
+    "type": "LLaMA3",
+    "description": "<ul><li>Fast responses</li><li>Instruct model</li><li>Multilingual dialogue use</li><li>Agentic system capable</li><li>Trained by Meta</li><li>License: <a href=\"https://llama.meta.com/llama3_2/license/\">Meta Llama 3.2 Community License</a></li></ul>",
+    "url": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf",
+    "promptTemplate": "<|start_header_id|>user<|end_header_id|>\n\n%1<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n%2",
+    "systemPrompt": "<|start_header_id|>system<|end_header_id|>\nCutting Knowledge Date: December 2023\n\nYou are a helpful assistant.<|eot_id|>",
+    "chatTemplate": "{{- bos_token }}\n{%- if not date_string is defined %}\n    {%- if strftime_now is defined %}\n        {%- set date_string = strftime_now(\"%d %b %Y\") %}\n    {%- else %}\n        {%- set date_string = \"26 Jul 2024\" %}\n    {%- endif %}\n{%- endif %}\n\n{#- This block extracts the system message, so we can slot it into the right place. #}\n{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content']|trim %}\n    {%- set messages = messages[1] %}\n{%- else %}\n    {%- set system_message = \"\" %}\n{%- endif %}\n\n{#- System message #}\n{{- \"<|start_header_id|>system<|end_header_id|>\\n\\n\" }}\n{{- \"Cutting Knowledge Date: December 2023\\n\" }}\n{{- \"Today Date: \" + date_string + \"\\n\\n\" }}\n{{- system_message }}\n{{- \"<|eot_id|>\" }}\n\n{%- for message in messages %}\n    {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}\n"
+  },
+  {
+    "order": "d",
     "md5sum": "a5f6b4eabd3992da4d7fb7f020f921eb",
     "name": "Nous Hermes 2 Mistral DPO",
     "filename": "Nous-Hermes-2-Mistral-7B-DPO.Q4_0.gguf",
@@ -32,7 +66,7 @@
     "systemPrompt": ""
   },
   {
-    "order": "c",
+    "order": "e",
     "md5sum": "97463be739b50525df56d33b26b00852",
     "name": "Mistral Instruct",
     "filename": "mistral-7b-instruct-v0.1.Q4_0.gguf",
@@ -45,10 +79,11 @@
     "systemPrompt": "",
     "description": "<strong>Strong overall fast instruction following model</strong><br><ul><li>Fast responses</li><li>Trained by Mistral AI<li>Uncensored</li><li>Licensed for commercial use</li></ul>",
     "url": "https://gpt4all.io/models/gguf/mistral-7b-instruct-v0.1.Q4_0.gguf",
-    "promptTemplate": "[INST] %1 [/INST]"
+    "promptTemplate": "[INST] %1 [/INST]",
+    "chatTemplate": "{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content'] %}\n    {%- set loop_start = 1 %}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if (message['role'] == 'user') != ((loop.index0 - loop_start) % 2 == 0) %}\n            {{- raise_exception('After the optional system message, conversation roles must alternate user/assistant/user/assistant/...') }}\n        {%- endif %}\n        {%- if message['role'] == 'user' %}\n            {%- if loop.first and system_message is defined %}\n                {{- ' [INST] ' + system_message + '\\n\\n' + message['content'] + ' [/INST]' }}\n            {%- else %}\n                {{- ' [INST] ' + message['content'] + ' [/INST]' }}\n            {%- endif %}\n        {%- elif message['role'] == 'assistant' %}\n            {{- ' ' + message['content'] + eos_token }}\n        {%- else %}\n            {{- raise_exception('Only user and assistant roles are supported, with the exception of an initial optional system message!') }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}"
   },
   {
-    "order": "d",
+    "order": "f",
     "md5sum": "8a9c75bcd8a66b7693f158ec96924eeb",
     "name": "Llama 3.1 8B Instruct 128k",
     "filename": "Meta-Llama-3.1-8B-Instruct-128k-Q4_0.gguf",
@@ -64,7 +99,7 @@
     "systemPrompt": "<|start_header_id|>system<|end_header_id|>\nCutting Knowledge Date: December 2023\n\nYou are a helpful assistant.<|eot_id|>"
   },
   {
-    "order": "e",
+    "order": "g",
     "md5sum": "f692417a22405d80573ac10cb0cd6c6a",
     "name": "Mistral OpenOrca",
     "filename": "mistral-7b-openorca.gguf2.Q4_0.gguf",
@@ -80,7 +115,7 @@
     "systemPrompt": "<|im_start|>system\nYou are MistralOrca, a large language model trained by Alignment Lab AI.\n<|im_end|>\n"
   },
   {
-    "order": "f",
+    "order": "h",
     "md5sum": "c4c78adf744d6a20f05c8751e3961b84",
     "name": "GPT4All Falcon",
     "filename": "gpt4all-falcon-newbpe-q4_0.gguf",
@@ -93,10 +128,11 @@
     "systemPrompt": "",
     "description": "<strong>Very fast model with good quality</strong><br><ul><li>Fastest responses</li><li>Instruction based</li><li>Trained by TII<li>Finetuned by Nomic AI<li>Licensed for commercial use</ul>",
     "url": "https://gpt4all.io/models/gguf/gpt4all-falcon-newbpe-q4_0.gguf",
-    "promptTemplate": "### Instruction:\n%1\n\n### Response:\n"
+    "promptTemplate": "### Instruction:\n%1\n\n### Response:\n",
+    "chatTemplate": "{%- if messages[0]['role'] == 'system' %}\n    {%- set loop_start = 1 %}\n    {{- messages[0]['content'] + '\\n\\n' }}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if message['role'] == 'user' %}\n            {{- '### User: ' + message['content'] + '\\n\\n' }}\n        {%- elif message['role'] == 'assistant' %}\n            {{- '### Assistant: ' + message['content'] + '\\n\\n' }}\n        {%- else %}\n            {{- raise_exception('Only user and assistant roles are supported, with the exception of an initial optional system message!') }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '### Assistant:' }}\n{%- endif %}"
   },
   {
-    "order": "g",
+    "order": "i",
     "md5sum": "00c8593ba57f5240f59662367b3ed4a5",
     "name": "Orca 2 (Medium)",
     "filename": "orca-2-7b.Q4_0.gguf",
@@ -108,10 +144,11 @@
     "type": "LLaMA2",
     "systemPrompt": "",
     "description": "<ul><li>Instruction based<li>Trained by Microsoft<li>Cannot be used commercially</ul>",
-    "url": "https://gpt4all.io/models/gguf/orca-2-7b.Q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/orca-2-7b.Q4_0.gguf",
+    "chatTemplate": "{%- for message in messages %}\n    {{- '<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>\\n' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}"
   },
   {
-    "order": "h",
+    "order": "j",
     "md5sum": "3c0d63c4689b9af7baa82469a6f51a19",
     "name": "Orca 2 (Full)",
     "filename": "orca-2-13b.Q4_0.gguf",
@@ -123,10 +160,11 @@
     "type": "LLaMA2",
     "systemPrompt": "",
     "description": "<ul><li>Instruction based<li>Trained by Microsoft<li>Cannot be used commercially</ul>",
-    "url": "https://gpt4all.io/models/gguf/orca-2-13b.Q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/orca-2-13b.Q4_0.gguf",
+    "chatTemplate": "{%- for message in messages %}\n    {{- '<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>\\n' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}"
   },
   {
-    "order": "i",
+    "order": "k",
     "md5sum": "5aff90007499bce5c64b1c0760c0b186",
     "name": "Wizard v1.2",
     "filename": "wizardlm-13b-v1.2.Q4_0.gguf",
@@ -138,10 +176,12 @@
     "type": "LLaMA2",
     "systemPrompt": "",
     "description": "<strong>Strong overall larger model</strong><br><ul><li>Instruction based<li>Gives very long responses<li>Finetuned with only 1k of high-quality data<li>Trained by Microsoft and Peking University<li>Cannot be used commercially</ul>",
-    "url": "https://gpt4all.io/models/gguf/wizardlm-13b-v1.2.Q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/wizardlm-13b-v1.2.Q4_0.gguf",
+    "chatTemplate": "{%- if messages[0]['role'] == 'system' %}\n    {%- set loop_start = 1 %}\n    {{- messages[0]['content'] + ' ' }}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in loop_messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if message['role'] == 'user' %}\n            {{- 'USER: ' + message['content'] }}\n        {%- elif message['role'] == 'assistant' %}\n            {{- 'ASSISTANT: ' + message['content'] }}\n        {%- else %}\n            {{- raise_exception('Only user and assistant roles are supported, with the exception of an initial optional system message!') }}\n        {%- endif %}\n        {%- if (loop.index0 - loop_start) % 2 == 0 %}\n            {{- ' ' }}\n        {%- else %}\n            {{- eos_token }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- 'ASSISTANT:' }}\n{%- endif %}",
+    "systemMessage": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions."
   },
   {
-    "order": "j",
+    "order": "l",
     "md5sum": "31b47b4e8c1816b62684ac3ca373f9e1",
     "name": "Ghost 7B v0.9.1",
     "filename": "ghost-7b-v0.9.1-Q4_0.gguf",
@@ -154,10 +194,11 @@
     "description": "<strong>Ghost 7B v0.9.1</strong> fast, powerful and smooth for Vietnamese and English languages.",
     "url": "https://huggingface.co/lamhieu/ghost-7b-v0.9.1-gguf/resolve/main/ghost-7b-v0.9.1-Q4_0.gguf",
     "promptTemplate": "<|user|>\n%1</s>\n<|assistant|>\n%2</s>\n",
-    "systemPrompt": "<|system|>\nYou are Ghost created by Lam Hieu. You are a helpful and knowledgeable assistant. You like to help and always give honest information, in its original language. In communication, you are always respectful, equal and promote positive behavior.\n</s>"
+    "systemPrompt": "<|system|>\nYou are Ghost created by Lam Hieu. You are a helpful and knowledgeable assistant. You like to help and always give honest information, in its original language. In communication, you are always respectful, equal and promote positive behavior.\n</s>",
+    "systemMessage": "You are Ghost created by Lam Hieu. You are a helpful and knowledgeable assistant. You like to help and always give honest information, in its original language. In communication, you are always respectful, equal and promote positive behavior."
   },
   {
-    "order": "k",
+    "order": "m",
     "md5sum": "3d12810391d04d1153b692626c0c6e16",
     "name": "Hermes",
     "filename": "nous-hermes-llama2-13b.Q4_0.gguf",
@@ -170,10 +211,11 @@
     "systemPrompt": "",
     "description": "<strong>Extremely good model</strong><br><ul><li>Instruction based<li>Gives long responses<li>Curated with 300,000 uncensored instructions<li>Trained by Nous Research<li>Cannot be used commercially</ul>",
     "url": "https://gpt4all.io/models/gguf/nous-hermes-llama2-13b.Q4_0.gguf",
-    "promptTemplate": "### Instruction:\n%1\n\n### Response:\n"
+    "promptTemplate": "### Instruction:\n%1\n\n### Response:\n",
+    "chatTemplate": "{%- if messages[0]['role'] == 'system' %}\n    {%- set loop_start = 1 %}\n    {{- messages[0]['content'] + '\\n\\n' }}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if message['role'] == 'user' %}\n            {{- '### Instruction:\\n' + message['content'] + '\\n\\n' }}\n        {%- elif message['role'] == 'assistant' %}\n            {{- '### Response:\\n' + message['content'] + '\\n\\n' }}\n        {%- else %}\n            {{- raise_exception('Only user and assistant roles are supported, with the exception of an initial optional system message!') }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '### Instruction:\\n' }}\n{%- endif %}"
   },
   {
-    "order": "l",
+    "order": "n",
     "md5sum": "40388eb2f8d16bb5d08c96fdfaac6b2c",
     "name": "Snoozy",
     "filename": "gpt4all-13b-snoozy-q4_0.gguf",
@@ -185,10 +227,12 @@
     "type": "LLaMA",
     "systemPrompt": "",
     "description": "<strong>Very good overall model</strong><br><ul><li>Instruction based<li>Based on the same dataset as Groovy<li>Slower than Groovy, with higher quality responses<li>Trained by Nomic AI<li>Cannot be used commercially</ul>",
-    "url": "https://gpt4all.io/models/gguf/gpt4all-13b-snoozy-q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/gpt4all-13b-snoozy-q4_0.gguf",
+    "chatTemplate": "{%- if messages[0]['role'] == 'system' %}\n    {%- set loop_start = 1 %}\n    {{- messages[0]['content'] + '\\n\\n' }}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if message['role'] == 'user' %}\n            {{- '### Instruction:\\n' + message['content'] + '\\n\\n' }}\n        {%- elif message['role'] == 'assistant' %}\n            {{- '### Response:\\n' + message['content'] + '\\n\\n' }}\n        {%- else %}\n            {{- raise_exception('Only user and assistant roles are supported, with the exception of an initial optional system message!') }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '### Response:\\n' }}\n{%- endif %}",
+    "systemMessage": "Below is an instruction that describes a task. Write a response that appropriately completes the request."
   },
   {
-    "order": "m",
+    "order": "o",
     "md5sum": "15dcb4d7ea6de322756449c11a0b7545",
     "name": "MPT Chat",
     "filename": "mpt-7b-chat-newbpe-q4_0.gguf",
@@ -202,10 +246,11 @@
     "description": "<strong>Good model with novel architecture</strong><br><ul><li>Fast responses<li>Chat based<li>Trained by Mosaic ML<li>Cannot be used commercially</ul>",
     "url": "https://gpt4all.io/models/gguf/mpt-7b-chat-newbpe-q4_0.gguf",
     "promptTemplate": "<|im_start|>user\n%1<|im_end|>\n<|im_start|>assistant\n%2<|im_end|>\n",
-    "systemPrompt": "<|im_start|>system\n- You are a helpful assistant chatbot trained by MosaicML.\n- You answer questions.\n- You are excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.\n- You are more than just an information source, you are also able to write poetry, short stories, and make jokes.<|im_end|>\n"
+    "systemPrompt": "<|im_start|>system\n- You are a helpful assistant chatbot trained by MosaicML.\n- You answer questions.\n- You are excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.\n- You are more than just an information source, you are also able to write poetry, short stories, and make jokes.<|im_end|>\n",
+    "chatTemplate": "{%- for message in messages %}\n    {{- '<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>\\n' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}"
   },
   {
-    "order": "n",
+    "order": "p",
     "md5sum": "ab5d8e8a2f79365ea803c1f1d0aa749d",
     "name": "MPT Chat",
     "filename": "mpt-7b-chat.gguf4.Q4_0.gguf",
@@ -218,10 +263,11 @@
     "description": "<strong>Good model with novel architecture</strong><br><ul><li>Fast responses<li>Chat based<li>Trained by Mosaic ML<li>Cannot be used commercially</ul>",
     "url": "https://gpt4all.io/models/gguf/mpt-7b-chat.gguf4.Q4_0.gguf",
     "promptTemplate": "<|im_start|>user\n%1<|im_end|>\n<|im_start|>assistant\n%2<|im_end|>\n",
-    "systemPrompt": "<|im_start|>system\n- You are a helpful assistant chatbot trained by MosaicML.\n- You answer questions.\n- You are excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.\n- You are more than just an information source, you are also able to write poetry, short stories, and make jokes.<|im_end|>\n"
+    "systemPrompt": "<|im_start|>system\n- You are a helpful assistant chatbot trained by MosaicML.\n- You answer questions.\n- You are excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.\n- You are more than just an information source, you are also able to write poetry, short stories, and make jokes.<|im_end|>\n",
+    "chatTemplate": "{%- for message in messages %}\n    {{- '<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>\\n' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}"
   },
   {
-    "order": "o",
+    "order": "q",
     "md5sum": "f8347badde9bfc2efbe89124d78ddaf5",
     "name": "Phi-3 Mini Instruct",
     "filename": "Phi-3-mini-4k-instruct.Q4_0.gguf",
@@ -237,7 +283,7 @@
     "systemPrompt": ""
   },
   {
-    "order": "p",
+    "order": "r",
     "md5sum": "0e769317b90ac30d6e09486d61fefa26",
     "name": "Mini Orca (Small)",
     "filename": "orca-mini-3b-gguf2-q4_0.gguf",
@@ -250,10 +296,11 @@
     "description": "<strong>Small version of new model with novel dataset</strong><br><ul><li>Very fast responses</li><li>Instruction based</li><li>Explain tuned datasets</li><li>Orca Research Paper dataset construction approaches</li><li>Cannot be used commercially</li></ul>",
     "url": "https://gpt4all.io/models/gguf/orca-mini-3b-gguf2-q4_0.gguf",
     "promptTemplate": "### User:\n%1\n\n### Response:\n",
-    "systemPrompt": "### System:\nYou are an AI assistant that follows instruction extremely well. Help as much as you can.\n\n"
+    "systemPrompt": "### System:\nYou are an AI assistant that follows instruction extremely well. Help as much as you can.\n\n",
+    "chatTemplate": "{%- if messages[0]['role'] == 'system' %}\n    {%- set loop_start = 1 %}\n    {{- '### System:\\n' + messages[0]['content'] + '\\n\\n' }}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if message['role'] == 'user' %}\n            {{- '### User:\\n' + message['content'] + '\\n\\n' }}\n        {%- elif message['role'] == 'assistant' %}\n            {{- '### Response:\\n' + message['content'] + '\\n\\n' }}\n        {%- else %}\n            {{- raise_exception('Only user and assistant roles are supported, with the exception of an initial optional system message!') }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '### Response:\\n' }}\n{%- endif %}"
   },
   {
-    "order": "q",
+    "order": "s",
     "md5sum": "c232f17e09bca4b7ee0b5b1f4107c01e",
     "disableGUI": "true",
     "name": "Replit",
@@ -267,10 +314,11 @@
     "systemPrompt": "",
     "promptTemplate": "%1",
     "description": "<strong>Trained on subset of the Stack</strong><br><ul><li>Code completion based<li>Licensed for commercial use<li>WARNING: Not available for chat GUI</ul>",
-    "url": "https://gpt4all.io/models/gguf/replit-code-v1_5-3b-newbpe-q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/replit-code-v1_5-3b-newbpe-q4_0.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "r",
+    "order": "t",
     "md5sum": "70841751ccd95526d3dcfa829e11cd4c",
     "disableGUI": "true",
     "name": "Starcoder",
@@ -284,10 +332,11 @@
     "systemPrompt": "",
     "promptTemplate": "%1",
     "description": "<strong>Trained on subset of the Stack</strong><br><ul><li>Code completion based<li>WARNING: Not available for chat GUI</ul>",
-    "url": "https://gpt4all.io/models/gguf/starcoder-newbpe-q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/starcoder-newbpe-q4_0.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "s",
+    "order": "u",
     "md5sum": "e973dd26f0ffa6e46783feaea8f08c83",
     "disableGUI": "true",
     "name": "Rift coder",
@@ -301,10 +350,11 @@
     "systemPrompt": "",
     "promptTemplate": "%1",
     "description": "<strong>Trained on collection of Python and TypeScript</strong><br><ul><li>Code completion based<li>WARNING: Not available for chat GUI</li>",
-    "url": "https://gpt4all.io/models/gguf/rift-coder-v0-7b-q4_0.gguf"
+    "url": "https://gpt4all.io/models/gguf/rift-coder-v0-7b-q4_0.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "t",
+    "order": "v",
     "md5sum": "e479e6f38b59afc51a470d1953a6bfc7",
     "disableGUI": "true",
     "name": "SBert",
@@ -319,10 +369,11 @@
     "embeddingModel": true,
     "systemPrompt": "",
     "description": "<strong>LocalDocs text embeddings model</strong><br><ul><li>For use with LocalDocs feature<li>Used for retrieval augmented generation (RAG)",
-    "url": "https://gpt4all.io/models/gguf/all-MiniLM-L6-v2-f16.gguf"
+    "url": "https://gpt4all.io/models/gguf/all-MiniLM-L6-v2-f16.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "u",
+    "order": "w",
     "md5sum": "dd90e2cb7f8e9316ac3796cece9883b5",
     "name": "SBert",
     "filename": "all-MiniLM-L6-v2.gguf2.f16.gguf",
@@ -335,10 +386,11 @@
     "type": "Bert",
     "embeddingModel": true,
     "description": "<strong>LocalDocs text embeddings model</strong><br><ul><li>For use with LocalDocs feature<li>Used for retrieval augmented generation (RAG)",
-    "url": "https://gpt4all.io/models/gguf/all-MiniLM-L6-v2.gguf2.f16.gguf"
+    "url": "https://gpt4all.io/models/gguf/all-MiniLM-L6-v2.gguf2.f16.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "v",
+    "order": "x",
     "md5sum": "919de4dd6f25351bcb0223790db1932d",
     "name": "EM German Mistral",
     "filename": "em_german_mistral_v01.Q4_0.gguf",
@@ -351,10 +403,12 @@
     "description": "<strong>Mistral-based model for German-language applications</strong><br><ul><li>Fast responses</li><li>Chat based model</li><li>Trained by ellamind<li>Finetuned on German instruction and chat data</a><li>Licensed for commercial use</ul>",
     "url": "https://huggingface.co/TheBloke/em_german_mistral_v01-GGUF/resolve/main/em_german_mistral_v01.Q4_0.gguf",
     "promptTemplate": "USER: %1 ASSISTANT: ",
-    "systemPrompt": "Du bist ein hilfreicher Assistent. "
+    "systemPrompt": "Du bist ein hilfreicher Assistent. ",
+    "chatTemplate": "{%- set system_message = false %}\n{%- if messages[0]['role'] == 'system' %}\n    {%- set loop_start = 1 %}\n    {%- set system_message = true %}\n    {{- messages[0]['content'] }}\n{%- else %}\n    {%- set loop_start = 0 %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if loop.index0 >= loop_start %}\n        {%- if (not loop.first) or (system_message is not none)  %}\n            {{- ' ' }}\n        {%- endif %}\n        {%- if message['role'] == 'user' %}\n            {{- 'USER: ' + message['content'] }}\n        {%- elif message['role'] == 'assistant' %}\n            {{- 'ASSISTANT: ' + message['content'] }}\n        {%- else %}\n            {{- raise_exception('After the optional system message, conversation roles must be either user or assistant.') }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {%- if messages %}\n        {{- ' ' }}\n    {%- endif %}\n    {{- 'ASSISTANT:' }}\n{%- endif %}",
+    "systemMessage": "Du bist ein hilfreicher Assistent."
   },
   {
-    "order": "w",
+    "order": "y",
     "md5sum": "60ea031126f82db8ddbbfecc668315d2",
     "disableGUI": "true",
     "name": "Nomic Embed Text v1",
@@ -368,10 +422,11 @@
     "embeddingModel": true,
     "systemPrompt": "",
     "description": "nomic-embed-text-v1",
-    "url": "https://gpt4all.io/models/gguf/nomic-embed-text-v1.f16.gguf"
+    "url": "https://gpt4all.io/models/gguf/nomic-embed-text-v1.f16.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "x",
+    "order": "z",
     "md5sum": "a5401e7f7e46ed9fcaed5b60a281d547",
     "disableGUI": "true",
     "name": "Nomic Embed Text v1.5",
@@ -385,10 +440,11 @@
     "embeddingModel": true,
     "systemPrompt": "",
     "description": "nomic-embed-text-v1.5",
-    "url": "https://gpt4all.io/models/gguf/nomic-embed-text-v1.5.f16.gguf"
+    "url": "https://gpt4all.io/models/gguf/nomic-embed-text-v1.5.f16.gguf",
+    "chatTemplate": null
   },
   {
-    "order": "z",
+    "order": "zzz",
     "md5sum": "a8c5a783105f87a481543d4ed7d7586d",
     "name": "Qwen2-1.5B-Instruct",
     "filename": "qwen2-1_5b-instruct-q4_0.gguf",

gpt4all-chat/metadata/release.json

@@ -213,5 +213,20 @@
     "version": "3.3.1",
     "notes": "* Fixed a crash when attempting to continue a chat loaded from disk\n* Fixed the local server rejecting min\\_p/top\\_p less than 1\n",
     "contributors": "* Jared Van Bortel (Nomic AI)"
+  },
+  {
+    "version": "3.4.0",
+    "notes": "* **Attached Files:** You can now attach a small Microsoft Excel spreadsheet (.xlsx) to a chat message and ask the model about it.\n* **LocalDocs Accuracy:** The LocalDocs algorithm has been enhanced to find more accurate references for some queries.\n* **Word Document Support:** LocalDocs now supports Microsoft Word (.docx) documents natively.\n  * **IMPORTANT NOTE:** If .docx files are not found, make sure Settings > LocalDocs > Allowed File Extensions includes \"docx\".\n* **Forgetful Model Fixes:** Issues with the \"Redo last chat response\" button, and with continuing chats from previous sessions, have been fixed.\n* **Chat Saving Improvements:** On exit, GPT4All will no longer save chats that are not new or modified. As a bonus, downgrading without losing access to all chats will be possible in the future, should the need arise.\n* **UI Fixes:** The model list no longer scrolls to the top when you start downloading a model.\n* **New Models:** LLama 3.2 Instruct 3B and 1B models now available in model list.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* Andriy Mulyar (Nomic AI)\n* Ikko Eltociear Ashimine (`@eltociear`)\n* Victor Emanuel (`@SINAPSA-IC`)\n* Shiranui (`@supersonictw`)"
+  },
+  {
+    "version": "3.4.1",
+    "notes": "* **LocalDocs Fixes:** Several issues with LocalDocs in v3.4.0 have been fixed, including missing words and very slow indexing.\n* **Syntax Highlighting:** Go code is now highlighted with the correct colors.\n* **Cache Fixes:** The model list cache is now stored with a version number, and in a more appropriate directory.\n* **Translation Updates:** The Italian translation has been improved.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* John Parent (Kitware)\n* Riccardo Giovanetti (`@Harvester62`)"
+  },
+  {
+    "version": "3.4.2",
+    "notes": "* **LocalDocs Fixes:** Several issues with LocalDocs, some of which were introduced in v3.4.0, have been fixed.\n  * Fixed the possible use of references from unselected collections.\n  * Fixed unnecessary reindexing of files with uppercase extensions.\n  * Fixed hybrid search failure due to inconsistent database state.\n  * Fully fixed the blank Embeddings Device selection in LocalDocs settings.\n  * Fixed LocalDocs indexing of large PDFs making very slow progress or even stalling.\n",
+    "contributors": "* Adam Treat (Nomic AI)\n* Jared Van Bortel (Nomic AI)"
   }
 ]

gpt4all-chat/pyproject.toml

@@ -0,0 +1,29 @@
+[tool.pytest.ini_options]
+addopts = ['--import-mode=importlib']
+
+[tool.mypy]
+files = 'tests/python'
+pretty = true
+strict = true
+warn_unused_ignores = false
+
+[tool.pytype]
+inputs = ['tests/python']
+jobs = 'auto'
+bind_decorated_methods = true
+none_is_not_bool = true
+overriding_renamed_parameter_count_checks = true
+strict_none_binding = true
+precise_return = true
+# protocols:
+# - https://github.com/google/pytype/issues/1423
+# - https://github.com/google/pytype/issues/1424
+strict_import = true
+strict_parameter_checks = true
+strict_primitive_comparisons = true
+# strict_undefined_checks: too many false positives
+
+[tool.isort]
+src_paths = ['tests/python']
+line_length = 120
+combine_as_imports = true

gpt4all-chat/qml/ApplicationSettings.qml

@@ -10,7 +10,7 @@ import network
 import llm
 
 MySettingsTab {
-    onRestoreDefaultsClicked: {
+    onRestoreDefaults: {
         MySettings.restoreApplicationDefaults();
     }
     title: qsTr("Application")
@@ -487,32 +487,32 @@ MySettingsTab {
             Accessible.description: ToolTip.text
         }
         MySettingsLabel {
-            id: saveChatsContextLabel
-            text: qsTr("Save Chat Context")
-            helpText: qsTr("Save the chat model's state to disk for faster loading. WARNING: Uses ~2GB per chat.")
-            Layout.row: 12
+            id: trayLabel
+            text: qsTr("Enable System Tray")
+            helpText: qsTr("The application will minimize to the system tray when the window is closed.")
+            Layout.row: 13
             Layout.column: 0
         }
         MyCheckBox {
-            id: saveChatsContextBox
-            Layout.row: 12
+            id: trayBox
+            Layout.row: 13
             Layout.column: 2
             Layout.alignment: Qt.AlignRight
-            checked: MySettings.saveChatsContext
+            checked: MySettings.systemTray
             onClicked: {
-                MySettings.saveChatsContext = !MySettings.saveChatsContext
+                MySettings.systemTray = !MySettings.systemTray
             }
         }
         MySettingsLabel {
             id: serverChatLabel
             text: qsTr("Enable Local API Server")
             helpText: qsTr("Expose an OpenAI-Compatible server to localhost. WARNING: Results in increased resource usage.")
-            Layout.row: 13
+            Layout.row: 14
             Layout.column: 0
         }
         MyCheckBox {
             id: serverChatBox
-            Layout.row: 13
+            Layout.row: 14
             Layout.column: 2
             Layout.alignment: Qt.AlignRight
             checked: MySettings.serverChat
@@ -524,7 +524,7 @@ MySettingsTab {
             id: serverPortLabel
             text: qsTr("API Server Port")
             helpText: qsTr("The port to use for the local server. Requires restart.")
-            Layout.row: 14
+            Layout.row: 15
             Layout.column: 0
         }
         MyTextField {
@@ -532,7 +532,7 @@ MySettingsTab {
             text: MySettings.networkPort
             color: theme.textColor
             font.pixelSize: theme.fontSizeLarge
-            Layout.row: 14
+            Layout.row: 15
             Layout.column: 2
             Layout.minimumWidth: 200
             Layout.maximumWidth: 200
@@ -577,12 +577,12 @@ MySettingsTab {
             id: updatesLabel
             text: qsTr("Check For Updates")
             helpText: qsTr("Manually check for an update to GPT4All.");
-            Layout.row: 15
+            Layout.row: 16
             Layout.column: 0
         }
 
         MySettingsButton {
-            Layout.row: 15
+            Layout.row: 16
             Layout.column: 2
             Layout.alignment: Qt.AlignRight
             text: qsTr("Updates");
@@ -593,7 +593,7 @@ MySettingsTab {
         }
 
         Rectangle {
-            Layout.row: 16
+            Layout.row: 17
             Layout.column: 0
             Layout.columnSpan: 3
             Layout.fillWidth: true

gpt4all-chat/qml/ChatItemView.qml

@@ -0,0 +1,884 @@
+import Qt5Compat.GraphicalEffects
+import QtCore
+import QtQuick
+import QtQuick.Controls
+import QtQuick.Controls.Basic
+import QtQuick.Layouts
+
+import gpt4all
+import mysettings
+
+ColumnLayout {
+
+property var inputBoxText: null
+signal setInputBoxText(text: string)
+
+Item {
+
+Layout.fillWidth: true
+Layout.maximumWidth: parent.width
+Layout.preferredHeight: gridLayout.height
+
+HoverHandler { id: hoverArea }
+
+GridLayout {
+    id: gridLayout
+    anchors.left: parent.left
+    anchors.right: parent.right
+    columns: 2
+
+    Item {
+        Layout.row: 0
+        Layout.column: 0
+        Layout.alignment: Qt.AlignVCenter | Qt.AlignRight
+        Layout.preferredWidth: 32
+        Layout.preferredHeight: 32
+        Image {
+            id: logo
+            sourceSize: Qt.size(32, 32)
+            fillMode: Image.PreserveAspectFit
+            mipmap: true
+            visible: false
+            source: name !== "Response: " ? "qrc:/gpt4all/icons/you.svg" : "qrc:/gpt4all/icons/gpt4all_transparent.svg"
+        }
+
+        ColorOverlay {
+            id: colorOver
+            anchors.fill: logo
+            source: logo
+            color: theme.conversationHeader
+            RotationAnimation {
+                id: rotationAnimation
+                target: colorOver
+                property: "rotation"
+                from: 0
+                to: 360
+                duration: 1000
+                loops: Animation.Infinite
+                running: isCurrentResponse && currentChat.responseInProgress
+            }
+        }
+    }
+
+    Item {
+        Layout.row: 0
+        Layout.column: 1
+        Layout.fillWidth: true
+        Layout.preferredHeight: 38
+        RowLayout {
+            spacing: 5
+            anchors.left: parent.left
+            anchors.top: parent.top
+            anchors.bottom: parent.bottom
+
+            TextArea {
+                text: name === "Response: " ? qsTr("GPT4All") : qsTr("You")
+                padding: 0
+                font.pixelSize: theme.fontSizeLarger
+                font.bold: true
+                color: theme.conversationHeader
+                enabled: false
+                focus: false
+                readOnly: true
+            }
+            Text {
+                visible: name === "Response: "
+                font.pixelSize: theme.fontSizeLarger
+                text: currentModelName()
+                color: theme.mutedTextColor
+            }
+            RowLayout {
+                visible: isCurrentResponse && (value === "" && currentChat.responseInProgress)
+                Text {
+                    color: theme.mutedTextColor
+                    font.pixelSize: theme.fontSizeLarger
+                    text: {
+                        switch (currentChat.responseState) {
+                        case Chat.ResponseStopped: return qsTr("response stopped ...");
+                        case Chat.LocalDocsRetrieval: return qsTr("retrieving localdocs: %1 ...").arg(currentChat.collectionList.join(", "));
+                        case Chat.LocalDocsProcessing: return qsTr("searching localdocs: %1 ...").arg(currentChat.collectionList.join(", "));
+                        case Chat.PromptProcessing: return qsTr("processing ...")
+                        case Chat.ResponseGeneration: return qsTr("generating response ...");
+                        case Chat.GeneratingQuestions: return qsTr("generating questions ...");
+                        default: return ""; // handle unexpected values
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    ColumnLayout {
+        Layout.row: 1
+        Layout.column: 1
+        Layout.fillWidth: true
+        spacing: 10
+        Flow {
+            id: attachedUrlsFlow
+            Layout.fillWidth: true
+            Layout.bottomMargin: 10
+            spacing: 10
+            visible: promptAttachments.length !== 0
+            Repeater {
+                model: promptAttachments
+
+                delegate: Rectangle {
+                    width: 350
+                    height: 50
+                    radius: 5
+                    color: theme.attachmentBackground
+                    border.color: theme.controlBorder
+
+                    Row {
+                        spacing: 5
+                        anchors.fill: parent
+                        anchors.margins: 5
+
+                        MyFileIcon {
+                            iconSize: 40
+                            fileName: modelData.file
+                        }
+
+                        Text {
+                            width: 295
+                            height: 40
+                            text: modelData.file
+                            color: theme.textColor
+                            horizontalAlignment: Text.AlignHLeft
+                            verticalAlignment: Text.AlignVCenter
+                            font.pixelSize: theme.fontSizeMedium
+                            font.bold: true
+                            wrapMode: Text.WrapAnywhere
+                            elide: Qt.ElideRight
+                        }
+                    }
+                }
+            }
+        }
+
+        TextArea {
+            id: myTextArea
+            Layout.fillWidth: true
+            padding: 0
+            color: {
+                if (!currentChat.isServer)
+                    return theme.textColor
+                return theme.white
+            }
+            wrapMode: Text.WordWrap
+            textFormat: TextEdit.PlainText
+            focus: false
+            readOnly: true
+            font.pixelSize: theme.fontSizeLarge
+            cursorVisible: isCurrentResponse ? currentChat.responseInProgress : false
+            cursorPosition: text.length
+            TapHandler {
+                id: tapHandler
+                onTapped: function(eventPoint, button) {
+                    var clickedPos = myTextArea.positionAt(eventPoint.position.x, eventPoint.position.y);
+                    var success = textProcessor.tryCopyAtPosition(clickedPos);
+                    if (success)
+                        copyCodeMessage.open();
+                }
+            }
+
+            MouseArea {
+                id: conversationMouseArea
+                anchors.fill: parent
+                acceptedButtons: Qt.RightButton
+
+                onClicked: (mouse) => {
+                               if (mouse.button === Qt.RightButton) {
+                                   conversationContextMenu.x = conversationMouseArea.mouseX
+                                   conversationContextMenu.y = conversationMouseArea.mouseY
+                                   conversationContextMenu.open()
+                               }
+                           }
+            }
+
+            onLinkActivated: function(link) {
+                if (!isCurrentResponse || !currentChat.responseInProgress)
+                    Qt.openUrlExternally(link)
+            }
+
+            onLinkHovered: function (link) {
+                if (!isCurrentResponse || !currentChat.responseInProgress)
+                    statusBar.externalHoveredLink = link
+            }
+
+            MyMenu {
+                id: conversationContextMenu
+                MyMenuItem {
+                    text: qsTr("Copy")
+                    enabled: myTextArea.selectedText !== ""
+                    height: enabled ? implicitHeight : 0
+                    onTriggered: myTextArea.copy()
+                }
+                MyMenuItem {
+                    text: qsTr("Copy Message")
+                    enabled: myTextArea.selectedText === ""
+                    height: enabled ? implicitHeight : 0
+                    onTriggered: {
+                        myTextArea.selectAll()
+                        myTextArea.copy()
+                        myTextArea.deselect()
+                    }
+                }
+                MyMenuItem {
+                    text: textProcessor.shouldProcessText ? qsTr("Disable markdown") : qsTr("Enable markdown")
+                    height: enabled ? implicitHeight : 0
+                    onTriggered: {
+                        textProcessor.shouldProcessText = !textProcessor.shouldProcessText;
+                        textProcessor.setValue(value);
+                    }
+                }
+            }
+
+            ChatViewTextProcessor {
+                id: textProcessor
+            }
+
+            function resetChatViewTextProcessor() {
+                textProcessor.fontPixelSize                = myTextArea.font.pixelSize
+                textProcessor.codeColors.defaultColor      = theme.codeDefaultColor
+                textProcessor.codeColors.keywordColor      = theme.codeKeywordColor
+                textProcessor.codeColors.functionColor     = theme.codeFunctionColor
+                textProcessor.codeColors.functionCallColor = theme.codeFunctionCallColor
+                textProcessor.codeColors.commentColor      = theme.codeCommentColor
+                textProcessor.codeColors.stringColor       = theme.codeStringColor
+                textProcessor.codeColors.numberColor       = theme.codeNumberColor
+                textProcessor.codeColors.headerColor       = theme.codeHeaderColor
+                textProcessor.codeColors.backgroundColor   = theme.codeBackgroundColor
+                textProcessor.textDocument                 = textDocument
+                textProcessor.setValue(value);
+            }
+
+            property bool textProcessorReady: false
+
+            Component.onCompleted: {
+                resetChatViewTextProcessor();
+                textProcessorReady = true;
+            }
+
+            Connections {
+                target: chatModel
+                function onValueChanged(i, value) {
+                    if (myTextArea.textProcessorReady && index === i)
+                        textProcessor.setValue(value);
+                }
+            }
+
+            Connections {
+                target: MySettings
+                function onFontSizeChanged() {
+                    myTextArea.resetChatViewTextProcessor();
+                }
+                function onChatThemeChanged() {
+                    myTextArea.resetChatViewTextProcessor();
+                }
+            }
+
+            Accessible.role: Accessible.Paragraph
+            Accessible.name: text
+            Accessible.description: name === "Response: " ? "The response by the model" : "The prompt by the user"
+        }
+
+        ThumbsDownDialog {
+            id: thumbsDownDialog
+            x: Math.round((parent.width - width) / 2)
+            y: Math.round((parent.height - height) / 2)
+            width: 640
+            height: 300
+            property string text: value
+            response: newResponse === undefined || newResponse === "" ? text : newResponse
+            onAccepted: {
+                var responseHasChanged = response !== text && response !== newResponse
+                if (thumbsDownState && !thumbsUpState && !responseHasChanged)
+                    return
+
+                chatModel.updateNewResponse(index, response)
+                chatModel.updateThumbsUpState(index, false)
+                chatModel.updateThumbsDownState(index, true)
+                Network.sendConversation(currentChat.id, getConversationJson());
+            }
+        }
+    }
+
+    Item {
+        Layout.row: 2
+        Layout.column: 1
+        Layout.topMargin: 5
+        Layout.alignment: Qt.AlignVCenter
+        Layout.preferredWidth: childrenRect.width
+        Layout.preferredHeight: childrenRect.height
+        visible: {
+            if (name !== "Response: ")
+                return false
+            if (consolidatedSources.length === 0)
+                return false
+            if (!MySettings.localDocsShowReferences)
+                return false
+            if (isCurrentResponse && currentChat.responseInProgress
+                    && currentChat.responseState !== Chat.GeneratingQuestions )
+                return false
+            return true
+        }
+
+        MyButton {
+            backgroundColor: theme.sourcesBackground
+            backgroundColorHovered: theme.sourcesBackgroundHovered
+            contentItem: RowLayout {
+                anchors.centerIn: parent
+
+                Item {
+                    Layout.preferredWidth: 24
+                    Layout.preferredHeight: 24
+
+                    Image {
+                        id: sourcesIcon
+                        visible: false
+                        anchors.fill: parent
+                        sourceSize.width: 24
+                        sourceSize.height: 24
+                        mipmap: true
+                        source: "qrc:/gpt4all/icons/db.svg"
+                    }
+
+                    ColorOverlay {
+                        anchors.fill: sourcesIcon
+                        source: sourcesIcon
+                        color: theme.textColor
+                    }
+                }
+
+                Text {
+                    text: qsTr("%n Source(s)", "", consolidatedSources.length)
+                    padding: 0
+                    font.pixelSize: theme.fontSizeLarge
+                    font.bold: true
+                    color: theme.styledTextColor
+                }
+
+                Item {
+                    Layout.preferredWidth: caret.width
+                    Layout.preferredHeight: caret.height
+                    Image {
+                        id: caret
+                        anchors.centerIn: parent
+                        visible: false
+                        sourceSize.width: theme.fontSizeLarge
+                        sourceSize.height: theme.fontSizeLarge
+                        mipmap: true
+                        source: {
+                            if (sourcesLayout.state === "collapsed")
+                                return "qrc:/gpt4all/icons/caret_right.svg";
+                            else
+                                return "qrc:/gpt4all/icons/caret_down.svg";
+                        }
+                    }
+
+                    ColorOverlay {
+                        anchors.fill: caret
+                        source: caret
+                        color: theme.textColor
+                    }
+                }
+            }
+
+            onClicked: {
+                if (sourcesLayout.state === "collapsed")
+                    sourcesLayout.state = "expanded";
+                else
+                    sourcesLayout.state = "collapsed";
+            }
+        }
+    }
+
+    ColumnLayout {
+        id: sourcesLayout
+        Layout.row: 3
+        Layout.column: 1
+        Layout.topMargin: 5
+        visible: {
+            if (consolidatedSources.length === 0)
+                return false
+            if (!MySettings.localDocsShowReferences)
+                return false
+            if (isCurrentResponse && currentChat.responseInProgress
+                    && currentChat.responseState !== Chat.GeneratingQuestions )
+                return false
+            return true
+        }
+        clip: true
+        Layout.fillWidth: true
+        Layout.preferredHeight: 0
+        state: "collapsed"
+        states: [
+            State {
+                name: "expanded"
+                PropertyChanges { target: sourcesLayout; Layout.preferredHeight: flow.height }
+            },
+            State {
+                name: "collapsed"
+                PropertyChanges { target: sourcesLayout; Layout.preferredHeight: 0 }
+            }
+        ]
+
+        transitions: [
+            Transition {
+                SequentialAnimation {
+                    PropertyAnimation {
+                        target: sourcesLayout
+                        property: "Layout.preferredHeight"
+                        duration: 300
+                        easing.type: Easing.InOutQuad
+                    }
+                }
+            }
+        ]
+
+        Flow {
+            id: flow
+            Layout.fillWidth: true
+            spacing: 10
+            visible: consolidatedSources.length !== 0
+            Repeater {
+                model: consolidatedSources
+
+                delegate: Rectangle {
+                    radius: 10
+                    color: ma.containsMouse ? theme.sourcesBackgroundHovered : theme.sourcesBackground
+                    width: 200
+                    height: 75
+
+                    MouseArea {
+                        id: ma
+                        enabled: modelData.path !== ""
+                        anchors.fill: parent
+                        hoverEnabled: true
+                        onClicked: function() {
+                            Qt.openUrlExternally(modelData.fileUri)
+                        }
+                    }
+
+                    Rectangle {
+                        id: debugTooltip
+                        anchors.right: parent.right
+                        anchors.bottom: parent.bottom
+                        width: 24
+                        height: 24
+                        color: "transparent"
+                        ToolTip {
+                            parent: debugTooltip
+                            visible: debugMouseArea.containsMouse
+                            text: modelData.text
+                            contentWidth: 900
+                            delay: 500
+                        }
+                        MouseArea {
+                            id: debugMouseArea
+                            anchors.fill: parent
+                            hoverEnabled: true
+                        }
+                    }
+
+                    ColumnLayout {
+                        anchors.left: parent.left
+                        anchors.top: parent.top
+                        anchors.margins: 10
+                        spacing: 0
+                        RowLayout {
+                            id: title
+                            spacing: 5
+                            Layout.maximumWidth: 180
+                            MyFileIcon {
+                                iconSize: 24
+                                fileName: modelData.file
+                                Layout.preferredWidth: iconSize
+                                Layout.preferredHeight: iconSize
+                            }
+                            Text {
+                                Layout.maximumWidth: 156
+                                text: modelData.collection !== "" ? modelData.collection : qsTr("LocalDocs")
+                                font.pixelSize: theme.fontSizeLarge
+                                font.bold: true
+                                color: theme.styledTextColor
+                                elide: Qt.ElideRight
+                            }
+                            Rectangle {
+                                Layout.fillWidth: true
+                                color: "transparent"
+                                height: 1
+                            }
+                        }
+                        Text {
+                            Layout.fillHeight: true
+                            Layout.maximumWidth: 180
+                            Layout.maximumHeight: 55 - title.height
+                            text: modelData.file
+                            color: theme.textColor
+                            font.pixelSize: theme.fontSizeSmall
+                            elide: Qt.ElideRight
+                            wrapMode: Text.WrapAnywhere
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    ConfirmationDialog {
+        id: editPromptDialog
+        dialogTitle: qsTr("Edit this prompt?")
+        description: qsTr("The existing response and all later messages will be permanently erased.")
+        onAccepted: {
+            const msg = currentChat.popPrompt(index);
+            if (msg !== null)
+                setInputBoxText(msg);
+        }
+    }
+
+    ConfirmationDialog {
+        id: redoResponseDialog
+        dialogTitle: qsTr("Redo this response?")
+        description: qsTr("The existing response and all later messages will be permanently erased.")
+        onAccepted: currentChat.regenerateResponse(index)
+    }
+
+    RowLayout {
+        id: buttonRow
+        Layout.row: 4
+        Layout.column: 1
+        Layout.maximumWidth: parent.width
+        Layout.fillWidth: false
+        Layout.alignment: Qt.AlignLeft | Qt.AlignTop
+        spacing: 3
+        visible: !isCurrentResponse || !currentChat.responseInProgress
+        enabled: opacity > 0
+        opacity: hoverArea.hovered
+        readonly property var canModify: !currentChat.isServer && currentChat.isModelLoaded && !currentChat.responseInProgress
+
+        Behavior on opacity {
+            OpacityAnimator { duration: 30 }
+        }
+
+        ChatMessageButton {
+            visible: parent.canModify && model.name === "Prompt: "
+            Layout.maximumWidth: 24
+            Layout.maximumHeight: 24
+            Layout.alignment: Qt.AlignVCenter
+            Layout.fillWidth: false
+            source: "qrc:/gpt4all/icons/edit.svg"
+            onClicked: {
+                if (inputBoxText === "")
+                    editPromptDialog.open();
+            }
+            name: qsTr("Edit")
+        }
+
+        ChatMessageButton {
+            visible: parent.canModify && model.name === "Response: "
+            Layout.maximumWidth: 24
+            Layout.maximumHeight: 24
+            Layout.alignment: Qt.AlignVCenter
+            Layout.fillWidth: false
+            name: qsTr("Redo")
+            source: "qrc:/gpt4all/icons/regenerate.svg"
+            onClicked: redoResponseDialog.open()
+        }
+
+        ChatMessageButton {
+            Layout.maximumWidth: 24
+            Layout.maximumHeight: 24
+            Layout.alignment: Qt.AlignVCenter
+            Layout.fillWidth: false
+            name: qsTr("Copy")
+            source: "qrc:/gpt4all/icons/copy.svg"
+            onClicked: {
+                myTextArea.selectAll();
+                myTextArea.copy();
+                myTextArea.deselect();
+            }
+        }
+
+        Item {
+            visible: name === "Response: " && MySettings.networkIsActive
+            Layout.alignment: Qt.AlignVCenter
+            Layout.preferredWidth: childrenRect.width
+            Layout.preferredHeight: childrenRect.height
+            Layout.fillWidth: false
+
+            ChatMessageButton {
+                id: thumbsUp
+                anchors.left: parent.left
+                anchors.verticalCenter: parent.verticalCenter
+                opacity: thumbsUpState || thumbsUpState == thumbsDownState ? 1.0 : 0.2
+                source: "qrc:/gpt4all/icons/thumbs_up.svg"
+                name: qsTr("Like response")
+                onClicked: {
+                    if (thumbsUpState && !thumbsDownState)
+                        return
+
+                    chatModel.updateNewResponse(index, "")
+                    chatModel.updateThumbsUpState(index, true)
+                    chatModel.updateThumbsDownState(index, false)
+                    Network.sendConversation(currentChat.id, getConversationJson());
+                }
+            }
+
+            ChatMessageButton {
+                id: thumbsDown
+                anchors.top: thumbsUp.top
+                anchors.topMargin: buttonRow.spacing
+                anchors.left: thumbsUp.right
+                anchors.leftMargin: buttonRow.spacing
+                checked: thumbsDownState
+                opacity: thumbsDownState || thumbsUpState == thumbsDownState ? 1.0 : 0.2
+                bgTransform: [
+                    Matrix4x4 {
+                        matrix: Qt.matrix4x4(-1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1)
+                    },
+                    Translate {
+                        x: thumbsDown.width
+                    }
+                ]
+                source: "qrc:/gpt4all/icons/thumbs_down.svg"
+                name: qsTr("Dislike response")
+                onClicked: {
+                    thumbsDownDialog.open()
+                }
+            }
+        }
+    }
+} // GridLayout
+
+} // Item
+
+GridLayout {
+    Layout.fillWidth: true
+    Layout.maximumWidth: parent.width
+
+    function shouldShowSuggestions() {
+        if (!isCurrentResponse)
+            return false;
+        if (MySettings.suggestionMode === 2) // Off
+            return false;
+        if (MySettings.suggestionMode === 0 && consolidatedSources.length === 0) // LocalDocs only
+            return false;
+        return currentChat.responseState === Chat.GeneratingQuestions || currentChat.generatedQuestions.length !== 0;
+    }
+
+    Item {
+        visible: parent.shouldShowSuggestions()
+        Layout.row: 5
+        Layout.column: 0
+        Layout.topMargin: 20
+        Layout.alignment: Qt.AlignVCenter | Qt.AlignRight
+        Layout.preferredWidth: 28
+        Layout.preferredHeight: 28
+        Image {
+            id: stack
+            sourceSize: Qt.size(28, 28)
+            fillMode: Image.PreserveAspectFit
+            mipmap: true
+            visible: false
+            source: "qrc:/gpt4all/icons/stack.svg"
+        }
+
+        ColorOverlay {
+            anchors.fill: stack
+            source: stack
+            color: theme.conversationHeader
+        }
+    }
+
+    Item {
+        visible: parent.shouldShowSuggestions()
+        Layout.row: 5
+        Layout.column: 1
+        Layout.topMargin: 20
+        Layout.fillWidth: true
+        Layout.preferredHeight: 38
+        RowLayout {
+            spacing: 5
+            anchors.left: parent.left
+            anchors.top: parent.top
+            anchors.bottom: parent.bottom
+
+            TextArea {
+                text: qsTr("Suggested follow-ups")
+                padding: 0
+                font.pixelSize: theme.fontSizeLarger
+                font.bold: true
+                color: theme.conversationHeader
+                enabled: false
+                focus: false
+                readOnly: true
+            }
+        }
+    }
+
+    ColumnLayout {
+        visible: parent.shouldShowSuggestions()
+        Layout.row: 6
+        Layout.column: 1
+        Layout.fillWidth: true
+        Layout.minimumHeight: 1
+        spacing: 10
+        Repeater {
+            model: currentChat.generatedQuestions
+            TextArea {
+                id: followUpText
+                Layout.fillWidth: true
+                Layout.alignment: Qt.AlignLeft
+                rightPadding: 40
+                topPadding: 10
+                leftPadding: 20
+                bottomPadding: 10
+                text: modelData
+                focus: false
+                readOnly: true
+                wrapMode: Text.WordWrap
+                hoverEnabled: !currentChat.responseInProgress
+                color: theme.textColor
+                font.pixelSize: theme.fontSizeLarge
+                background: Rectangle {
+                    color: hovered ? theme.sourcesBackgroundHovered : theme.sourcesBackground
+                    radius: 10
+                }
+                MouseArea {
+                    id: maFollowUp
+                    anchors.fill: parent
+                    enabled: !currentChat.responseInProgress
+                    onClicked: function() {
+                        var chat = window.currentChat
+                        var followup = modelData
+                        chat.stopGenerating()
+                        chat.newPromptResponsePair(followup)
+                    }
+                }
+                Item {
+                    anchors.right: parent.right
+                    anchors.verticalCenter: parent.verticalCenter
+                    width: 40
+                    height: 40
+                    visible: !currentChat.responseInProgress
+                    Image {
+                        id: plusImage
+                        anchors.verticalCenter: parent.verticalCenter
+                        sourceSize.width: 20
+                        sourceSize.height: 20
+                        mipmap: true
+                        visible: false
+                        source: "qrc:/gpt4all/icons/plus.svg"
+                    }
+
+                    ColorOverlay {
+                        anchors.fill: plusImage
+                        source: plusImage
+                        color: theme.styledTextColor
+                    }
+                }
+            }
+        }
+
+        Rectangle {
+            Layout.fillWidth: true
+            color: "transparent"
+            radius: 10
+            Layout.preferredHeight: currentChat.responseInProgress ? 40 : 0
+            clip: true
+            ColumnLayout {
+                id: followUpLayout
+                anchors.fill: parent
+                Rectangle {
+                    id: myRect1
+                    Layout.preferredWidth: 0
+                    Layout.minimumWidth: 0
+                    Layout.maximumWidth: parent.width
+                    height: 12
+                    color: theme.sourcesBackgroundHovered
+                }
+
+                Rectangle {
+                    id: myRect2
+                    Layout.preferredWidth: 0
+                    Layout.minimumWidth: 0
+                    Layout.maximumWidth: parent.width
+                    height: 12
+                    color: theme.sourcesBackgroundHovered
+                }
+
+                SequentialAnimation {
+                    id: followUpProgressAnimation
+                    ParallelAnimation {
+                        PropertyAnimation {
+                            target: myRect1
+                            property: "Layout.preferredWidth"
+                            from: 0
+                            to: followUpLayout.width
+                            duration: 1000
+                        }
+                        PropertyAnimation {
+                            target: myRect2
+                            property: "Layout.preferredWidth"
+                            from: 0
+                            to: followUpLayout.width / 2
+                            duration: 1000
+                        }
+                    }
+                    SequentialAnimation {
+                        loops: Animation.Infinite
+                        ParallelAnimation {
+                            PropertyAnimation {
+                                target: myRect1
+                                property: "opacity"
+                                from: 1
+                                to: 0.2
+                                duration: 1500
+                            }
+                            PropertyAnimation {
+                                target: myRect2
+                                property: "opacity"
+                                from: 1
+                                to: 0.2
+                                duration: 1500
+                            }
+                        }
+                        ParallelAnimation {
+                            PropertyAnimation {
+                                target: myRect1
+                                property: "opacity"
+                                from: 0.2
+                                to: 1
+                                duration: 1500
+                            }
+                            PropertyAnimation {
+                                target: myRect2
+                                property: "opacity"
+                                from: 0.2
+                                to: 1
+                                duration: 1500
+                            }
+                        }
+                    }
+                }
+
+                onVisibleChanged: {
+                    if (visible)
+                        followUpProgressAnimation.start();
+                }
+            }
+
+            Behavior on Layout.preferredHeight {
+                NumberAnimation {
+                    duration: 300
+                    easing.type: Easing.InOutQuad
+                }
+            }
+        }
+    }
+
+} // GridLayout
+
+} // ColumnLayout

gpt4all-chat/qml/ChatMessageButton.qml

@@ -0,0 +1,20 @@
+import QtQuick
+import QtQuick.Controls
+
+import gpt4all
+
+MyToolButton {
+    property string name
+
+    width: 24
+    height: 24
+    imageWidth: width
+    imageHeight: height
+    ToolTip {
+        visible: parent.hovered
+        y: parent.height * 1.5
+        text: name
+        delay: Qt.styleHints.mousePressAndHoldInterval
+    }
+    Accessible.name: name
+}

gpt4all-chat/qml/ConfirmationDialog.qml

@@ -0,0 +1,59 @@
+import QtCore
+import QtQuick
+import QtQuick.Controls
+import QtQuick.Controls.Basic
+import QtQuick.Layouts
+
+MyDialog {
+    id: confirmationDialog
+    anchors.centerIn: parent
+    modal: true
+    padding: 20
+    property alias dialogTitle: titleText.text
+    property alias description: descriptionText.text
+
+    Theme { id: theme }
+
+    contentItem: ColumnLayout {
+        Text {
+            id: titleText
+            Layout.alignment: Qt.AlignHCenter
+            textFormat: Text.StyledText
+            color: theme.textColor
+            font.pixelSize: theme.fontSizeLarger
+            font.bold: true
+        }
+
+        Text {
+            id: descriptionText
+            Layout.alignment: Qt.AlignHCenter
+            textFormat: Text.StyledText
+            color: theme.textColor
+            font.pixelSize: theme.fontSizeMedium
+        }
+    }
+
+    footer: DialogButtonBox {
+        id: dialogBox
+        padding: 20
+        alignment: Qt.AlignRight
+        spacing: 10
+        MySettingsButton {
+            text: qsTr("OK")
+            textColor: theme.mediumButtonText
+            backgroundColor: theme.mediumButtonBackground
+            backgroundColorHovered: theme.mediumButtonBackgroundHovered
+            DialogButtonBox.buttonRole: DialogButtonBox.AcceptRole
+        }
+        MySettingsButton {
+            text: qsTr("Cancel")
+            DialogButtonBox.buttonRole: DialogButtonBox.RejectRole
+        }
+        background: Rectangle {
+            color: "transparent"
+        }
+        Keys.onEnterPressed: confirmationDialog.accept()
+        Keys.onReturnPressed: confirmationDialog.accept()
+    }
+    Component.onCompleted: dialogBox.forceActiveFocus()
+}

gpt4all-chat/qml/HomeView.qml

@@ -281,7 +281,7 @@ Rectangle {
                 Layout.alignment: Qt.AlignCenter
                 text: qsTr("Subscribe to Newsletter")
                 imageSource: "qrc:/gpt4all/icons/email.svg"
-                onClicked: { Qt.openUrlExternally("https://nomic.ai/gpt4all") }
+                onClicked: { Qt.openUrlExternally("https://nomic.ai/gpt4all/#newsletter-form") }
             }
         }
     }

gpt4all-chat/qml/LocalDocsSettings.qml

@@ -10,7 +10,7 @@ import mysettings
 import network
 
 MySettingsTab {
-    onRestoreDefaultsClicked: {
+    onRestoreDefaults: {
         MySettings.restoreLocalDocsDefaults();
     }
 
@@ -176,6 +176,7 @@ MySettingsTab {
                     ListElement { text: qsTr("Application default") }
                     Component.onCompleted: {
                         MySettings.embeddingsDeviceList.forEach(d => append({"text": d}));
+                        deviceBox.updateModel();
                     }
                 }
                 Accessible.name: deviceLabel.text

gpt4all-chat/qml/ModelSettings.qml

@@ -8,10 +8,34 @@ import mysettings
 import chatlistmodel
 
 MySettingsTab {
-    onRestoreDefaultsClicked: {
+    onRestoreDefaults: {
         MySettings.restoreModelDefaults(root.currentModelInfo);
     }
     title: qsTr("Model")
+
+    ConfirmationDialog {
+        id: resetSystemMessageDialog
+        property var index: null
+        property bool resetClears: false
+        dialogTitle: qsTr("%1 system message?").arg(resetClears ? qsTr("Clear") : qsTr("Reset"))
+        description: qsTr("The system message will be %1.").arg(resetClears ? qsTr("removed") : qsTr("reset to the default"))
+        onAccepted: MySettings.resetModelSystemMessage(ModelList.modelInfo(index))
+        function show(index_, resetClears_) { index = index_; resetClears = resetClears_; open(); }
+    }
+
+    ConfirmationDialog {
+        id: resetChatTemplateDialog
+        property bool resetClears: false
+        property var index: null
+        dialogTitle: qsTr("%1 chat template?").arg(resetClears ? qsTr("Clear") : qsTr("Reset"))
+        description: qsTr("The chat template will be %1.").arg(resetClears ? qsTr("erased") : qsTr("reset to the default"))
+        onAccepted: {
+            MySettings.resetModelChatTemplate(ModelList.modelInfo(index));
+            templateTextArea.resetText();
+        }
+        function show(index_, resetClears_) { index = index_; resetClears = resetClears_; open(); }
+    }
+
     contentItem: GridLayout {
         id: root
         columns: 3
@@ -35,6 +59,7 @@ MySettingsTab {
 
         RowLayout {
             Layout.fillWidth: true
+            Layout.maximumWidth: parent.width
             Layout.row: 2
             Layout.column: 0
             Layout.columnSpan: 2
@@ -153,69 +178,154 @@ MySettingsTab {
             Layout.fillWidth: true
         }
 
-        MySettingsLabel {
-            visible: !root.currentModelInfo.isOnline
-            text: qsTr("System Prompt")
-            helpText: qsTr("Prefixed at the beginning of every conversation. Must contain the appropriate framing tokens.")
+        RowLayout {
             Layout.row: 7
-            Layout.column: 0
+            Layout.columnSpan: 2
             Layout.topMargin: 15
+            Layout.fillWidth: true
+            Layout.maximumWidth: parent.width
+            spacing: 10
+            MySettingsLabel {
+                id: systemMessageLabel
+                text: qsTr("System Message")
+                helpText: qsTr("A message to set the context or guide the behavior of the model. Leave blank for " +
+                               "none. NOTE: Since GPT4All 3.5, this should not contain control tokens.")
+                onReset: () => resetSystemMessageDialog.show(root.currentModelId, resetClears)
+                function updateResetButton() {
+                    const info = root.currentModelInfo;
+                    // NOTE: checks if the *override* is set, regardless of whether there is a default
+                    canReset = !!info.id && MySettings.isModelSystemMessageSet(info);
+                    resetClears = !info.defaultSystemMessage;
+                }
+                Component.onCompleted: updateResetButton()
+                Connections {
+                    target: root
+                    function onCurrentModelIdChanged() { systemMessageLabel.updateResetButton(); }
+                }
+                Connections {
+                    target: MySettings
+                    function onSystemMessageChanged(info)
+                    { if (info.id === root.currentModelId) systemMessageLabel.updateResetButton(); }
+                }
+            }
+            Label {
+                id: systemMessageLabelHelp
+                visible: systemMessageArea.errState !== "ok"
+                Layout.alignment: Qt.AlignBottom
+                Layout.fillWidth: true
+                Layout.rightMargin: 5
+                Layout.maximumHeight: systemMessageLabel.height
+                text: qsTr("System message is not " +
+                           "<a href=\"https://docs.gpt4all.io/gpt4all_desktop/chat_templates.html\">plain text</a>.")
+                color: systemMessageArea.errState === "error" ? theme.textErrorColor : theme.textWarningColor
+                font.pixelSize: theme.fontSizeLarger
+                font.bold: true
+                wrapMode: Text.Wrap
+                elide: Text.ElideRight
+                onLinkActivated: function(link) { Qt.openUrlExternally(link) }
+            }
         }
 
         Rectangle {
-            id: systemPrompt
-            visible: !root.currentModelInfo.isOnline
+            id: systemMessage
             Layout.row: 8
             Layout.column: 0
             Layout.columnSpan: 2
             Layout.fillWidth: true
             color: "transparent"
-            Layout.minimumHeight: Math.max(100, systemPromptArea.contentHeight + 20)
+            Layout.minimumHeight: Math.max(100, systemMessageArea.contentHeight + 20)
             MyTextArea {
-                id: systemPromptArea
+                id: systemMessageArea
                 anchors.fill: parent
-                text: root.currentModelInfo.systemPrompt
+                property bool isBeingReset: false
+                function resetText() {
+                    const info = root.currentModelInfo;
+                    isBeingReset = true;
+                    text = (info.id ? info.systemMessage.value : null) ?? "";
+                    isBeingReset = false;
+                }
+                Component.onCompleted: resetText()
                 Connections {
                     target: MySettings
-                    function onSystemPromptChanged() {
-                        systemPromptArea.text = root.currentModelInfo.systemPrompt;
-                    }
+                    function onSystemMessageChanged(info)
+                    { if (info.id === root.currentModelId) systemMessageArea.resetText(); }
                 }
                 Connections {
                     target: root
-                    function onCurrentModelInfoChanged() {
-                        systemPromptArea.text = root.currentModelInfo.systemPrompt;
-                    }
+                    function onCurrentModelIdChanged() { systemMessageArea.resetText(); }
                 }
+                // strict validation, because setModelSystemMessage clears isLegacy
+                readonly property var reLegacyCheck: (
+                    /(?:^|\s)(?:### *System\b|S(?:ystem|YSTEM):)|<\|(?:im_(?:start|end)|(?:start|end)_header_id|eot_id|SYSTEM_TOKEN)\|>|<<SYS>>/m
+                )
                 onTextChanged: {
-                    MySettings.setModelSystemPrompt(root.currentModelInfo, text)
+                    const info = root.currentModelInfo;
+                    if (!info.id) {
+                        errState = "ok";
+                    } else if (info.systemMessage.isLegacy && (isBeingReset || reLegacyCheck.test(text))) {
+                        errState = "error";
+                    } else
+                        errState = reLegacyCheck.test(text) ? "warning" : "ok";
+                    if (info.id && errState !== "error" && !isBeingReset)
+                        MySettings.setModelSystemMessage(info, text);
+                    systemMessageLabel.updateResetButton();
                 }
                 Accessible.role: Accessible.EditableText
+                Accessible.name: systemMessageLabel.text
+                Accessible.description: systemMessageLabelHelp.text
             }
         }
 
         RowLayout {
             Layout.row: 9
-            Layout.column: 0
             Layout.columnSpan: 2
             Layout.topMargin: 15
+            Layout.fillWidth: true
+            Layout.maximumWidth: parent.width
             spacing: 10
             MySettingsLabel {
-                id: promptTemplateLabel
-                text: qsTr("Prompt Template")
-                helpText: qsTr("The template that wraps every prompt.")
+                id: chatTemplateLabel
+                text: qsTr("Chat Template")
+                helpText: qsTr("This Jinja template turns the chat into input for the model.")
+                onReset: () => resetChatTemplateDialog.show(root.currentModelId, resetClears)
+                function updateResetButton() {
+                    const info = root.currentModelInfo;
+                    canReset = !!info.id && (
+                        MySettings.isModelChatTemplateSet(info)
+                        || templateTextArea.text !== (info.chatTemplate.value ?? "")
+                    );
+                    resetClears = !info.defaultChatTemplate;
+                }
+                Component.onCompleted: updateResetButton()
+                Connections {
+                    target: root
+                    function onCurrentModelIdChanged() { chatTemplateLabel.updateResetButton(); }
+                }
+                Connections {
+                    target: MySettings
+                    function onChatTemplateChanged(info)
+                    { if (info.id === root.currentModelId) chatTemplateLabel.updateResetButton(); }
+                }
             }
-            MySettingsLabel {
-                id: promptTemplateLabelHelp
-                text: qsTr("Must contain the string \"%1\" to be replaced with the user's input.")
-                color: theme.textErrorColor
-                visible: templateTextArea.text.indexOf("%1") === -1
-                wrapMode: TextArea.Wrap
+            Label {
+                id: chatTemplateLabelHelp
+                visible: templateTextArea.errState !== "ok"
+                Layout.alignment: Qt.AlignBottom
+                Layout.fillWidth: true
+                Layout.rightMargin: 5
+                Layout.maximumHeight: chatTemplateLabel.height
+                text: templateTextArea.errMsg
+                color: templateTextArea.errState === "error" ? theme.textErrorColor : theme.textWarningColor
+                font.pixelSize: theme.fontSizeLarger
+                font.bold: true
+                wrapMode: Text.Wrap
+                elide: Text.ElideRight
+                onLinkActivated: function(link) { Qt.openUrlExternally(link) }
             }
         }
 
         Rectangle {
-            id: promptTemplate
+            id: chatTemplate
             Layout.row: 10
             Layout.column: 0
             Layout.columnSpan: 2
@@ -226,27 +336,71 @@ MySettingsTab {
             MyTextArea {
                 id: templateTextArea
                 anchors.fill: parent
-                text: root.currentModelInfo.promptTemplate
+                font: fixedFont
+                property bool isBeingReset: false
+                property var errMsg: null
+                function resetText() {
+                    const info = root.currentModelInfo;
+                    isBeingReset = true;
+                    text = (info.id ? info.chatTemplate.value : null) ?? "";
+                    isBeingReset = false;
+                }
+                Component.onCompleted: resetText()
                 Connections {
                     target: MySettings
-                    function onPromptTemplateChanged() {
-                        templateTextArea.text = root.currentModelInfo.promptTemplate;
-                    }
+                    function onChatTemplateChanged() { templateTextArea.resetText(); }
                 }
                 Connections {
                     target: root
-                    function onCurrentModelInfoChanged() {
-                        templateTextArea.text = root.currentModelInfo.promptTemplate;
-                    }
+                    function onCurrentModelIdChanged() { templateTextArea.resetText(); }
+                }
+                function legacyCheck() {
+                    return /%[12]\b/.test(text) || !/\{%.*%\}.*\{\{.*\}\}.*\{%.*%\}/.test(text.replace(/\n/g, ''))
+                        || !/\bcontent\b/.test(text);
                 }
                 onTextChanged: {
-                    if (templateTextArea.text.indexOf("%1") !== -1) {
-                        MySettings.setModelPromptTemplate(root.currentModelInfo, text)
+                    const info = root.currentModelInfo;
+                    let jinjaError;
+                    if (!info.id) {
+                        errMsg = null;
+                        errState = "ok";
+                    } else if (info.chatTemplate.isLegacy && (isBeingReset || legacyCheck())) {
+                        errMsg = null;
+                        errState = "error";
+                    } else if (text === "" && !info.chatTemplate.isSet) {
+                        errMsg = qsTr("No <a href=\"https://docs.gpt4all.io/gpt4all_desktop/chat_templates.html\">" +
+                                      "chat template</a> configured.");
+                        errState = "error";
+                    } else if (/^\s*$/.test(text)) {
+                        errMsg = qsTr("The <a href=\"https://docs.gpt4all.io/gpt4all_desktop/chat_templates.html\">" +
+                                      "chat template</a> cannot be blank.");
+                        errState = "error";
+                    } else if ((jinjaError = MySettings.checkJinjaTemplateError(text)) !== null) {
+                        errMsg = qsTr("<a href=\"https://docs.gpt4all.io/gpt4all_desktop/chat_templates.html\">Syntax" +
+                                      " error</a>: %1").arg(jinjaError);
+                        errState = "error";
+                    } else if (legacyCheck()) {
+                        errMsg = qsTr("Chat template is not in " +
+                                      "<a href=\"https://docs.gpt4all.io/gpt4all_desktop/chat_templates.html\">" +
+                                      "Jinja format</a>.")
+                        errState = "warning";
+                    } else {
+                        errState = "ok";
+                    }
+                    if (info.id && errState !== "error" && !isBeingReset)
+                        MySettings.setModelChatTemplate(info, text);
+                    chatTemplateLabel.updateResetButton();
+                }
+                Keys.onPressed: event => {
+                    if (event.key === Qt.Key_Tab) {
+                        const a = templateTextArea;
+                        event.accepted = true;              // suppress tab
+                        a.insert(a.cursorPosition, '    '); // four spaces
                     }
                 }
                 Accessible.role: Accessible.EditableText
-                Accessible.name: promptTemplateLabel.text
-                Accessible.description: promptTemplateLabelHelp.text
+                Accessible.name: chatTemplateLabel.text
+                Accessible.description: chatTemplateLabelHelp.text
             }
         }
 

gpt4all-chat/qml/MyFileIcon.qml

@@ -0,0 +1,41 @@
+import QtCore
+import QtQuick
+import QtQuick.Controls
+import QtQuick.Controls.Basic
+import Qt5Compat.GraphicalEffects
+
+Item {
+    id: fileIcon
+    property real iconSize: 24
+    property string fileName: ""
+    implicitWidth: iconSize
+    implicitHeight: iconSize
+
+    Image {
+        id: fileImage
+        anchors.fill: parent
+        visible: false
+        sourceSize.width: iconSize
+        sourceSize.height: iconSize
+        mipmap: true
+        source: {
+            if (fileIcon.fileName.toLowerCase().endsWith(".txt"))
+                return "qrc:/gpt4all/icons/file-txt.svg"
+            else if (fileIcon.fileName.toLowerCase().endsWith(".pdf"))
+                return "qrc:/gpt4all/icons/file-pdf.svg"
+            else if (fileIcon.fileName.toLowerCase().endsWith(".md"))
+                return "qrc:/gpt4all/icons/file-md.svg"
+            else if (fileIcon.fileName.toLowerCase().endsWith(".xlsx"))
+                return "qrc:/gpt4all/icons/file-xls.svg"
+            else if (fileIcon.fileName.toLowerCase().endsWith(".docx"))
+                return "qrc:/gpt4all/icons/file-docx.svg"
+            else
+                return "qrc:/gpt4all/icons/file.svg"
+        }
+    }
+    ColorOverlay {
+        anchors.fill: fileImage
+        source: fileImage
+        color: theme.textColor
+    }
+}

gpt4all-chat/qml/MySettingsButton.qml

@@ -17,6 +17,7 @@ Button {
     property color  borderColor: "transparent"
     property real   fontPixelSize: theme.fontSizeLarge
     property string toolTip
+    property alias backgroundRadius: background.radius
 
     contentItem: Text {
         text: myButton.text
@@ -28,6 +29,7 @@ Button {
         Accessible.name: text
     }
     background: Rectangle {
+        id: background
         radius: 10
         border.width: borderWidth
         border.color: borderColor

gpt4all-chat/qml/MySettingsLabel.qml

@@ -17,13 +17,42 @@ ColumnLayout {
     property alias color: mainTextLabel.color
     property alias linkColor: mainTextLabel.linkColor
 
-    Label {
-        id: mainTextLabel
-        color: theme.settingsTitleTextColor
-        font.pixelSize: theme.fontSizeLarger
-        font.bold: true
-        onLinkActivated: function(link) {
-            root.linkActivated(link);
+    property var onReset: null
+    property alias canReset: resetButton.enabled
+    property bool resetClears: false
+
+    Item {
+        anchors.margins: 5
+        width: childrenRect.width
+        height: mainTextLabel.contentHeight
+
+        Label {
+            id: mainTextLabel
+            anchors.left: parent.left
+            anchors.top: parent.top
+            anchors.bottom: parent.bottom
+            color: theme.settingsTitleTextColor
+            font.pixelSize: theme.fontSizeLarger
+            font.bold: true
+            verticalAlignment: Text.AlignVCenter
+            onLinkActivated: function(link) {
+                root.linkActivated(link);
+            }
+        }
+
+        MySettingsButton {
+            id: resetButton
+            anchors.baseline: mainTextLabel.baseline
+            anchors.left: mainTextLabel.right
+            height: mainTextLabel.contentHeight
+            anchors.leftMargin: 10
+            padding: 2
+            leftPadding: 10
+            rightPadding: 10
+            backgroundRadius: 5
+            text: resetClears ? qsTr("Clear") : qsTr("Reset")
+            visible: root.onReset !== null
+            onClicked: root.onReset()
         }
     }
     Label {

gpt4all-chat/qml/MySettingsStack.qml

@@ -2,6 +2,7 @@ import QtCore
 import QtQuick
 import QtQuick.Controls
 import QtQuick.Controls.Basic
+import QtQuick.Controls.impl
 import QtQuick.Layouts
 import QtQuick.Dialogs
 import Qt.labs.folderlistmodel

gpt4all-chat/qml/MySettingsTab.qml

@@ -9,7 +9,7 @@ Item {
     property string title: ""
     property Item contentItem: null
     property bool showRestoreDefaultsButton: true
-    signal restoreDefaultsClicked
+    signal restoreDefaults
 
     onContentItemChanged: function() {
         if (contentItem) {
@@ -19,6 +19,13 @@ Item {
         }
     }
 
+    ConfirmationDialog {
+        id: restoreDefaultsDialog
+        dialogTitle: qsTr("Restore defaults?")
+        description: qsTr("This page of settings will be reset to the defaults.")
+        onAccepted: root.restoreDefaults()
+    }
+
     ScrollView {
         id: scrollView
         width: parent.width
@@ -47,6 +54,7 @@ Item {
             Column {
                 id: contentInner
                 Layout.fillWidth: true
+                Layout.maximumWidth: parent.width
             }
 
             Item {
@@ -63,9 +71,7 @@ Item {
                     Accessible.role: Accessible.Button
                     Accessible.name: text
                     Accessible.description: qsTr("Restores settings dialog to a default state")
-                    onClicked: {
-                        root.restoreDefaultsClicked();
-                    }
+                    onClicked: restoreDefaultsDialog.open()
                 }
             }
         }

gpt4all-chat/qml/MyTextArea.qml

@@ -5,18 +5,27 @@ import QtQuick.Controls.Basic
 
 TextArea {
     id: myTextArea
+
+    property string errState: "ok"  // one of "ok", "error", "warning"
+
     color: enabled ? theme.textColor : theme.mutedTextColor
     placeholderTextColor: theme.mutedTextColor
     font.pixelSize: theme.fontSizeLarge
     background: Rectangle {
         implicitWidth: 150
         color: theme.controlBackground
-        border.width: 1
-        border.color: theme.controlBorder
+        border.width: errState === "ok" ? 1 : 2
+        border.color: {
+            switch (errState) {
+                case "ok":      return theme.controlBorder;
+                case "warning": return theme.textWarningColor;
+                case "error":   return theme.textErrorColor;
+            }
+        }
         radius: 10
     }
     padding: 10
     wrapMode: TextArea.Wrap
 
     ToolTip.delay: Qt.styleHints.mousePressAndHoldInterval
-}
+}

gpt4all-chat/qml/MyToolButton.qml

@@ -16,6 +16,7 @@ Button {
     property alias fillMode: image.fillMode
     property alias imageWidth: image.sourceSize.width
     property alias imageHeight: image.sourceSize.height
+    property alias bgTransform: background.transform
     contentItem: Text {
         text: myButton.text
         horizontalAlignment: Text.AlignHCenter
@@ -26,6 +27,7 @@ Button {
     }
 
     background: Item {
+        id: background
         anchors.fill: parent
         Rectangle {
             anchors.fill: parent

gpt4all-chat/qml/SwitchModelDialog.qml

@@ -1,46 +0,0 @@
-import QtCore
-import QtQuick
-import QtQuick.Controls
-import QtQuick.Controls.Basic
-import QtQuick.Layouts
-import llm
-import mysettings
-
-MyDialog {
-    id: switchModelDialog
-    anchors.centerIn: parent
-    modal: true
-    padding: 20
-    property int index: -1
-
-    Theme {
-        id: theme
-    }
-
-    contentItem: Text {
-        textFormat: Text.StyledText
-        text: qsTr("<b>Warning:</b> changing the model will erase the current conversation. Do you wish to continue?")
-        color: theme.textColor
-        font.pixelSize: theme.fontSizeLarge
-    }
-
-    footer: DialogButtonBox {
-        id: dialogBox
-        padding: 20
-        alignment: Qt.AlignRight
-        spacing: 10
-        MySettingsButton {
-            text: qsTr("Continue")
-            Accessible.description: qsTr("Continue with model loading")
-            DialogButtonBox.buttonRole: DialogButtonBox.AcceptRole
-        }
-        MySettingsButton {
-            text: qsTr("Cancel")
-            Accessible.description: qsTr("Cancel")
-            DialogButtonBox.buttonRole: DialogButtonBox.RejectRole
-        }
-        background: Rectangle {
-            color: "transparent"
-        }
-    }
-}

gpt4all-chat/qml/Theme.qml

@@ -64,6 +64,9 @@ QtObject {
     property color green800: Qt.hsla(123/360, 0.17, 0.24)
     property color green900: Qt.hsla(124/360, 0.17, 0.20)
     property color green950: Qt.hsla(125/360, 0.22, 0.10)
+    property color green300_sat: Qt.hsla(122/360, 0.24, 0.73)
+    property color green400_sat: Qt.hsla(122/360, 0.23, 0.58)
+    property color green450_sat: Qt.hsla(122/360, 0.23, 0.52)
 
     // yellow
     property color yellow0: Qt.hsla(47/360, 0.90, 0.99)
@@ -99,6 +102,7 @@ QtObject {
     property color purple200: Qt.hsla(279/360, 1.0, 0.91)
     property color purple300: Qt.hsla(279/360, 1.0, 0.84)
     property color purple400: Qt.hsla(279/360, 1.0, 0.73)
+    property color purple450: Qt.hsla(279/360, 1.0, 0.68)
     property color purple500: Qt.hsla(279/360, 1.0, 0.63)
     property color purple600: Qt.hsla(279/360, 1.0, 0.53)
     property color purple700: Qt.hsla(279/360, 1.0, 0.47)
@@ -408,6 +412,39 @@ QtObject {
         }
     }
 
+    property color mediumButtonBackground: {
+        switch (MySettings.chatTheme) {
+            case MySettingsEnums.ChatTheme.LegacyDark:
+                return purple400
+            case MySettingsEnums.ChatTheme.Dark:
+                return green400_sat
+            default:
+                return green400_sat
+        }
+    }
+
+    property color mediumButtonBackgroundHovered: {
+        switch (MySettings.chatTheme) {
+            case MySettingsEnums.ChatTheme.LegacyDark:
+                return purple450
+            case MySettingsEnums.ChatTheme.Dark:
+                return green450_sat
+            default:
+                return green300_sat
+        }
+    }
+
+    property color mediumButtonText: {
+        switch (MySettings.chatTheme) {
+            case MySettingsEnums.ChatTheme.LegacyDark:
+                return textColor
+            case MySettingsEnums.ChatTheme.Dark:
+                return textColor
+            default:
+                return white
+        }
+    }
+
     property color darkButtonText: {
         switch (MySettings.chatTheme) {
             case MySettingsEnums.ChatTheme.LegacyDark:
@@ -922,16 +959,8 @@ QtObject {
         }
     }
 
-    property color textErrorColor: {
-        switch (MySettings.chatTheme) {
-            case MySettingsEnums.ChatTheme.LegacyDark:
-                return red400
-            case MySettingsEnums.ChatTheme.Dark:
-                return red400
-            default:
-                return red400
-        }
-    }
+    readonly property color textErrorColor:   red400
+    readonly property color textWarningColor: yellow400
 
     property color settingsTitleTextColor: {
         switch (MySettings.chatTheme) {

gpt4all-chat/src/chat.cpp

@@ -1,7 +1,6 @@
 #include "chat.h"
 
 #include "chatlistmodel.h"
-#include "mysettings.h"
 #include "network.h"
 #include "server.h"
 
@@ -11,7 +10,6 @@
 #include <QLatin1String>
 #include <QMap>
 #include <QString>
-#include <QStringList>
 #include <QVariant>
 #include <Qt>
 #include <QtLogging>
@@ -31,7 +29,7 @@ Chat::Chat(QObject *parent)
     connectLLM();
 }
 
-Chat::Chat(bool isServer, QObject *parent)
+Chat::Chat(server_tag_t, QObject *parent)
     : QObject(parent)
     , m_id(Network::globalInstance()->generateUniqueId())
     , m_name(tr("Server Chat"))
@@ -56,18 +54,18 @@ void Chat::connectLLM()
     // Should be in different threads
     connect(m_llmodel, &ChatLLM::modelLoadingPercentageChanged, this, &Chat::handleModelLoadingPercentageChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::responseChanged, this, &Chat::handleResponseChanged, Qt::QueuedConnection);
+    connect(m_llmodel, &ChatLLM::responseFailed, this, &Chat::handleResponseFailed, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::promptProcessing, this, &Chat::promptProcessing, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::generatingQuestions, this, &Chat::generatingQuestions, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::responseStopped, this, &Chat::responseStopped, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::modelLoadingError, this, &Chat::handleModelLoadingError, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::modelLoadingWarning, this, &Chat::modelLoadingWarning, Qt::QueuedConnection);
-    connect(m_llmodel, &ChatLLM::restoringFromTextChanged, this, &Chat::handleRestoringFromText, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::generatedNameChanged, this, &Chat::generatedNameChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::generatedQuestionFinished, this, &Chat::generatedQuestionFinished, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::reportSpeed, this, &Chat::handleTokenSpeedChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::loadedModelInfoChanged, this, &Chat::loadedModelInfoChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::databaseResultsChanged, this, &Chat::handleDatabaseResultsChanged, Qt::QueuedConnection);
-    connect(m_llmodel, &ChatLLM::modelInfoChanged, this, &Chat::handleModelInfoChanged, Qt::QueuedConnection);
+    connect(m_llmodel, &ChatLLM::modelInfoChanged, this, &Chat::handleModelChanged, Qt::QueuedConnection);
     connect(m_llmodel, &ChatLLM::trySwitchContextOfLoadedModelCompleted, this, &Chat::handleTrySwitchContextOfLoadedModelCompleted, Qt::QueuedConnection);
 
     connect(this, &Chat::promptRequested, m_llmodel, &ChatLLM::prompt, Qt::QueuedConnection);
@@ -75,11 +73,10 @@ void Chat::connectLLM()
     connect(this, &Chat::loadDefaultModelRequested, m_llmodel, &ChatLLM::loadDefaultModel, Qt::QueuedConnection);
     connect(this, &Chat::generateNameRequested, m_llmodel, &ChatLLM::generateName, Qt::QueuedConnection);
     connect(this, &Chat::regenerateResponseRequested, m_llmodel, &ChatLLM::regenerateResponse, Qt::QueuedConnection);
-    connect(this, &Chat::resetResponseRequested, m_llmodel, &ChatLLM::resetResponse, Qt::QueuedConnection);
-    connect(this, &Chat::resetContextRequested, m_llmodel, &ChatLLM::resetContext, Qt::QueuedConnection);
-    connect(this, &Chat::processSystemPromptRequested, m_llmodel, &ChatLLM::processSystemPrompt, Qt::QueuedConnection);
 
     connect(this, &Chat::collectionListChanged, m_collectionModel, &LocalDocsCollectionsModel::setCollections);
+
+    connect(ModelList::globalInstance(), &ModelList::modelInfoChanged, this, &Chat::handleModelInfoChanged);
 }
 
 void Chat::reset()
@@ -87,25 +84,15 @@ void Chat::reset()
     stopGenerating();
     // Erase our current on disk representation as we're completely resetting the chat along with id
     ChatListModel::globalInstance()->removeChatFile(this);
-    emit resetContextRequested();
     m_id = Network::globalInstance()->generateUniqueId();
     emit idChanged(m_id);
     // NOTE: We deliberately do no reset the name or creation date to indicate that this was originally
     // an older chat that was reset for another purpose. Resetting this data will lead to the chat
     // name label changing back to 'New Chat' and showing up in the chat model list as a 'New Chat'
     // further down in the list. This might surprise the user. In the future, we might get rid of
-    // the "reset context" button in the UI. Right now, by changing the model in the combobox dropdown
-    // we effectively do a reset context. We *have* to do this right now when switching between different
-    // types of models. The only way to get rid of that would be a very long recalculate where we rebuild
-    // the context if we switch between different types of models. Probably the right way to fix this
-    // is to allow switching models but throwing up a dialog warning users if we switch between types
-    // of models that a long recalculation will ensue.
+    // the "reset context" button in the UI.
     m_chatModel->clear();
-}
-
-void Chat::processSystemPrompt()
-{
-    emit processSystemPromptRequested();
+    m_needsSave = true;
 }
 
 void Chat::resetResponseState()
@@ -131,7 +118,13 @@ void Chat::newPromptResponsePair(const QString &prompt, const QList<QUrl> &attac
         Q_ASSERT(url.isLocalFile());
         const QString localFilePath = url.toLocalFile();
         const QFileInfo info(localFilePath);
-        Q_ASSERT(info.suffix() == "xlsx"); // We only support excel right now
+
+        Q_ASSERT(
+            info.suffix().toLower() == "xlsx" ||
+            info.suffix().toLower() == "txt" ||
+            info.suffix().toLower() == "md" ||
+            info.suffix().toLower() == "rst"
+        );
 
         PromptAttachment attached;
         attached.url = url;
@@ -153,23 +146,29 @@ void Chat::newPromptResponsePair(const QString &prompt, const QList<QUrl> &attac
     if (!attachedContexts.isEmpty())
         promptPlusAttached = attachedContexts.join("\n\n") + "\n\n" + prompt;
 
-    newPromptResponsePairInternal(prompt, attachments);
-    emit resetResponseRequested();
+    resetResponseState();
+    if (int count = m_chatModel->count())
+        m_chatModel->updateCurrentResponse(count - 1, false);
+    m_chatModel->appendPrompt(prompt, attachments);
+    m_chatModel->appendResponse();
 
-    this->prompt(promptPlusAttached);
+    emit promptRequested(m_collections);
+    m_needsSave = true;
 }
 
-void Chat::prompt(const QString &prompt)
+void Chat::regenerateResponse(int index)
 {
     resetResponseState();
-    emit promptRequested(m_collections, prompt);
+    emit regenerateResponseRequested(index);
+    m_needsSave = true;
 }
 
-void Chat::regenerateResponse()
+QVariant Chat::popPrompt(int index)
 {
-    const int index = m_chatModel->count() - 1;
-    m_chatModel->updateSources(index, QList<ResultInfo>());
-    emit regenerateResponseRequested();
+    auto content = m_llmodel->popPrompt(index);
+    m_needsSave = true;
+    if (content) return *content;
+    return QVariant::fromValue(nullptr);
 }
 
 void Chat::stopGenerating()
@@ -177,11 +176,6 @@ void Chat::stopGenerating()
     m_llmodel->stopGenerating();
 }
 
-QString Chat::response() const
-{
-    return m_response;
-}
-
 Chat::ResponseState Chat::responseState() const
 {
     return m_responseState;
@@ -194,10 +188,16 @@ void Chat::handleResponseChanged(const QString &response)
         emit responseStateChanged();
     }
 
-    m_response = response;
     const int index = m_chatModel->count() - 1;
-    m_chatModel->updateValue(index, this->response());
-    emit responseChanged();
+    m_chatModel->updateValue(index, response);
+}
+
+void Chat::handleResponseFailed(const QString &error)
+{
+    const int index = m_chatModel->count() - 1;
+    m_chatModel->updateValue(index, error);
+    m_chatModel->setError();
+    responseStopped(0);
 }
 
 void Chat::handleModelLoadingPercentageChanged(float loadingPercentage)
@@ -224,7 +224,7 @@ void Chat::handleModelLoadingPercentageChanged(float loadingPercentage)
 void Chat::promptProcessing()
 {
     m_responseState = !databaseResults().isEmpty() ? Chat::LocalDocsProcessing : Chat::PromptProcessing;
-     emit responseStateChanged();
+    emit responseStateChanged();
 }
 
 void Chat::generatingQuestions()
@@ -237,7 +237,6 @@ void Chat::responseStopped(qint64 promptResponseMs)
 {
     m_tokenSpeed = QString();
     emit tokenSpeedChanged();
-    emit responseChanged();
 
     m_responseInProgress = false;
     m_responseState = Chat::ResponseStopped;
@@ -261,33 +260,16 @@ ModelInfo Chat::modelInfo() const
 
 void Chat::setModelInfo(const ModelInfo &modelInfo)
 {
-    if (m_modelInfo == modelInfo && isModelLoaded())
+    if (m_modelInfo != modelInfo) {
+        m_modelInfo = modelInfo;
+        m_needsSave = true;
+    } else if (isModelLoaded())
         return;
 
-    m_modelInfo = modelInfo;
     emit modelInfoChanged();
     emit modelChangeRequested(modelInfo);
 }
 
-// the server needs to block until response is reset, so it calls resetResponse on its own m_llmThread
-void Chat::serverNewPromptResponsePair(const QString &prompt, const QList<PromptAttachment> &attachments)
-{
-    newPromptResponsePairInternal(prompt, attachments);
-}
-
-void Chat::newPromptResponsePairInternal(const QString &prompt, const QList<PromptAttachment> &attachments)
-{
-    resetResponseState();
-    m_chatModel->updateCurrentResponse(m_chatModel->count() - 1, false);
-    m_chatModel->appendPrompt("Prompt: ", prompt, attachments);
-    m_chatModel->appendResponse("Response: ");
-}
-
-bool Chat::restoringFromText() const
-{
-    return m_llmodel->restoringFromText();
-}
-
 void Chat::unloadAndDeleteLater()
 {
     if (!isModelLoaded()) {
@@ -343,18 +325,14 @@ void Chat::generatedNameChanged(const QString &name)
     int wordCount = qMin(7, words.size());
     m_name = words.mid(0, wordCount).join(' ');
     emit nameChanged();
+    m_needsSave = true;
 }
 
 void Chat::generatedQuestionFinished(const QString &question)
 {
     m_generatedQuestions << question;
     emit generatedQuestionsChanged();
-}
-
-void Chat::handleRestoringFromText()
-{
-    Network::globalInstance()->trackChatEvent("recalc_context", { {"length", m_chatModel->count()} });
-    emit restoringFromTextChanged();
+    m_needsSave = true;
 }
 
 void Chat::handleModelLoadingError(const QString &error)
@@ -391,17 +369,26 @@ QString Chat::fallbackReason() const
 void Chat::handleDatabaseResultsChanged(const QList<ResultInfo> &results)
 {
     m_databaseResults = results;
-    const int index = m_chatModel->count() - 1;
-    m_chatModel->updateSources(index, m_databaseResults);
+    m_needsSave = true;
 }
 
+// we need to notify listeners of the modelInfo property when its properties are updated,
+// since it's a gadget and can't do that on its own
 void Chat::handleModelInfoChanged(const ModelInfo &modelInfo)
+{
+    if (!m_modelInfo.id().isNull() && modelInfo.id() == m_modelInfo.id())
+        emit modelInfoChanged();
+}
+
+// react if a new model is loaded
+void Chat::handleModelChanged(const ModelInfo &modelInfo)
 {
     if (m_modelInfo == modelInfo)
         return;
 
     m_modelInfo = modelInfo;
     emit modelInfoChanged();
+    m_needsSave = true;
 }
 
 void Chat::handleTrySwitchContextOfLoadedModelCompleted(int value)
@@ -416,17 +403,14 @@ bool Chat::serialize(QDataStream &stream, int version) const
     stream << m_id;
     stream << m_name;
     stream << m_userName;
-    if (version > 4)
+    if (version >= 5)
         stream << m_modelInfo.id();
     else
         stream << m_modelInfo.filename();
-    if (version > 2)
+    if (version >= 3)
         stream << m_collections;
 
-    const bool serializeKV = MySettings::globalInstance()->saveChatsContext();
-    if (version > 5)
-        stream << serializeKV;
-    if (!m_llmodel->serialize(stream, version, serializeKV))
+    if (!m_llmodel->serialize(stream, version))
         return false;
     if (!m_chatModel->serialize(stream, version))
         return false;
@@ -445,7 +429,7 @@ bool Chat::deserialize(QDataStream &stream, int version)
 
     QString modelId;
     stream >> modelId;
-    if (version > 4) {
+    if (version >= 5) {
         if (ModelList::globalInstance()->contains(modelId))
             m_modelInfo = ModelList::globalInstance()->modelInfo(modelId);
     } else {
@@ -455,25 +439,23 @@ bool Chat::deserialize(QDataStream &stream, int version)
     if (!m_modelInfo.id().isEmpty())
         emit modelInfoChanged();
 
-    bool discardKV = m_modelInfo.id().isEmpty();
-
-    if (version > 2) {
+    if (version >= 3) {
         stream >> m_collections;
         emit collectionListChanged(m_collections);
     }
 
-    bool deserializeKV = true;
-    if (version > 5)
-        stream >> deserializeKV;
-
     m_llmodel->setModelInfo(m_modelInfo);
-    if (!m_llmodel->deserialize(stream, version, deserializeKV, discardKV))
+    if (!m_llmodel->deserialize(stream, version))
         return false;
     if (!m_chatModel->deserialize(stream, version))
         return false;
 
     emit chatModelChanged();
-    return stream.status() == QDataStream::Ok;
+    if (stream.status() != QDataStream::Ok)
+        return false;
+
+    m_needsSave = false;
+    return true;
 }
 
 QList<QString> Chat::collectionList() const
@@ -493,6 +475,7 @@ void Chat::addCollection(const QString &collection)
 
     m_collections.append(collection);
     emit collectionListChanged(m_collections);
+    m_needsSave = true;
 }
 
 void Chat::removeCollection(const QString &collection)
@@ -502,4 +485,5 @@ void Chat::removeCollection(const QString &collection)
 
     m_collections.removeAll(collection);
     emit collectionListChanged(m_collections);
+    m_needsSave = true;
 }

gpt4all-chat/src/chat.h

@@ -12,6 +12,8 @@
 #include <QObject>
 #include <QQmlEngine>
 #include <QString>
+#include <QStringList> // IWYU pragma: keep
+#include <QStringView>
 #include <QtGlobal>
 
 class QDataStream;
@@ -25,15 +27,13 @@ class Chat : public QObject
     Q_PROPERTY(bool isModelLoaded READ isModelLoaded NOTIFY isModelLoadedChanged)
     Q_PROPERTY(bool isCurrentlyLoading READ isCurrentlyLoading NOTIFY isCurrentlyLoadingChanged)
     Q_PROPERTY(float modelLoadingPercentage READ modelLoadingPercentage NOTIFY modelLoadingPercentageChanged)
-    Q_PROPERTY(QString response READ response NOTIFY responseChanged)
     Q_PROPERTY(ModelInfo modelInfo READ modelInfo WRITE setModelInfo NOTIFY modelInfoChanged)
     Q_PROPERTY(bool responseInProgress READ responseInProgress NOTIFY responseInProgressChanged)
-    Q_PROPERTY(bool restoringFromText READ restoringFromText NOTIFY restoringFromTextChanged)
     Q_PROPERTY(bool isServer READ isServer NOTIFY isServerChanged)
     Q_PROPERTY(ResponseState responseState READ responseState NOTIFY responseStateChanged)
     Q_PROPERTY(QList<QString> collectionList READ collectionList NOTIFY collectionListChanged)
     Q_PROPERTY(QString modelLoadingError READ modelLoadingError NOTIFY modelLoadingErrorChanged)
-    Q_PROPERTY(QString tokenSpeed READ tokenSpeed NOTIFY tokenSpeedChanged);
+    Q_PROPERTY(QString tokenSpeed READ tokenSpeed NOTIFY tokenSpeedChanged)
     Q_PROPERTY(QString deviceBackend READ deviceBackend NOTIFY loadedModelInfoChanged)
     Q_PROPERTY(QString device READ device NOTIFY loadedModelInfoChanged)
     Q_PROPERTY(QString fallbackReason READ fallbackReason NOTIFY loadedModelInfoChanged)
@@ -45,6 +45,10 @@ class Chat : public QObject
     QML_UNCREATABLE("Only creatable from c++!")
 
 public:
+    // tag for constructing a server chat
+    struct server_tag_t { explicit server_tag_t() = default; };
+    static inline constexpr server_tag_t server_tag = server_tag_t();
+
     enum ResponseState {
         ResponseStopped,
         LocalDocsRetrieval,
@@ -56,7 +60,7 @@ public:
     Q_ENUM(ResponseState)
 
     explicit Chat(QObject *parent = nullptr);
-    explicit Chat(bool isServer, QObject *parent = nullptr);
+    explicit Chat(server_tag_t, QObject *parent = nullptr);
     virtual ~Chat();
     void destroy() { m_llmodel->destroy(); }
     void connectLLM();
@@ -67,29 +71,27 @@ public:
     {
         m_userName = name;
         emit nameChanged();
+        m_needsSave = true;
     }
     ChatModel *chatModel() { return m_chatModel; }
 
     bool isNewChat() const { return m_name == tr("New Chat") && !m_chatModel->count(); }
 
     Q_INVOKABLE void reset();
-    Q_INVOKABLE void processSystemPrompt();
     bool  isModelLoaded()          const { return m_modelLoadingPercentage == 1.0f; }
     bool  isCurrentlyLoading()     const { return m_modelLoadingPercentage > 0.0f && m_modelLoadingPercentage < 1.0f; }
     float modelLoadingPercentage() const { return m_modelLoadingPercentage; }
     Q_INVOKABLE void newPromptResponsePair(const QString &prompt, const QList<QUrl> &attachedUrls = {});
-    Q_INVOKABLE void prompt(const QString &prompt);
-    Q_INVOKABLE void regenerateResponse();
+    Q_INVOKABLE void regenerateResponse(int index);
+    Q_INVOKABLE QVariant popPrompt(int index);
     Q_INVOKABLE void stopGenerating();
 
     QList<ResultInfo> databaseResults() const { return m_databaseResults; }
 
-    QString response() const;
     bool responseInProgress() const { return m_responseInProgress; }
     ResponseState responseState() const;
     ModelInfo modelInfo() const;
     void setModelInfo(const ModelInfo &modelInfo);
-    bool restoringFromText() const;
 
     Q_INVOKABLE void unloadModel();
     Q_INVOKABLE void reloadModel();
@@ -110,7 +112,6 @@ public:
     Q_INVOKABLE bool hasCollection(const QString &collection) const;
     Q_INVOKABLE void addCollection(const QString &collection);
     Q_INVOKABLE void removeCollection(const QString &collection);
-    void resetResponseState();
 
     QString modelLoadingError() const { return m_modelLoadingError; }
 
@@ -124,8 +125,11 @@ public:
 
     QList<QString> generatedQuestions() const { return m_generatedQuestions; }
 
+    bool needsSave() const { return m_needsSave; }
+    void setNeedsSave(bool n) { m_needsSave = n; }
+
 public Q_SLOTS:
-    void serverNewPromptResponsePair(const QString &prompt, const QList<PromptAttachment> &attachments = {});
+    void resetResponseState();
 
 Q_SIGNALS:
     void idChanged(const QString &id);
@@ -135,17 +139,14 @@ Q_SIGNALS:
     void isCurrentlyLoadingChanged();
     void modelLoadingPercentageChanged();
     void modelLoadingWarning(const QString &warning);
-    void responseChanged();
     void responseInProgressChanged();
     void responseStateChanged();
-    void promptRequested(const QList<QString> &collectionList, const QString &prompt);
-    void regenerateResponseRequested();
+    void promptRequested(const QStringList &enabledCollections);
+    void regenerateResponseRequested(int index);
     void resetResponseRequested();
     void resetContextRequested();
-    void processSystemPromptRequested();
     void modelChangeRequested(const ModelInfo &modelInfo);
     void modelInfoChanged();
-    void restoringFromTextChanged();
     void loadDefaultModelRequested();
     void generateNameRequested();
     void modelLoadingErrorChanged();
@@ -161,22 +162,20 @@ Q_SIGNALS:
 
 private Q_SLOTS:
     void handleResponseChanged(const QString &response);
+    void handleResponseFailed(const QString &error);
     void handleModelLoadingPercentageChanged(float);
     void promptProcessing();
     void generatingQuestions();
     void responseStopped(qint64 promptResponseMs);
     void generatedNameChanged(const QString &name);
     void generatedQuestionFinished(const QString &question);
-    void handleRestoringFromText();
     void handleModelLoadingError(const QString &error);
     void handleTokenSpeedChanged(const QString &tokenSpeed);
     void handleDatabaseResultsChanged(const QList<ResultInfo> &results);
     void handleModelInfoChanged(const ModelInfo &modelInfo);
+    void handleModelChanged(const ModelInfo &modelInfo);
     void handleTrySwitchContextOfLoadedModelCompleted(int value);
 
-private:
-    void newPromptResponsePairInternal(const QString &prompt, const QList<PromptAttachment> &attachments);
-
 private:
     QString m_id;
     QString m_name;
@@ -187,7 +186,6 @@ private:
     QString m_tokenSpeed;
     QString m_device;
     QString m_fallbackReason;
-    QString m_response;
     QList<QString> m_collections;
     QList<QString> m_generatedQuestions;
     ChatModel *m_chatModel;
@@ -203,6 +201,10 @@ private:
     bool m_firstResponse = true;
     int m_trySwitchContextInProgress = 0;
     bool m_isCurrentlyLoading = false;
+    // True if we need to serialize the chat to disk, because of one of two reasons:
+    // - The chat was freshly created during this launch.
+    // - The chat was changed after loading it from disk.
+    bool m_needsSave = true;
 };
 
 #endif // CHAT_H

gpt4all-chat/src/chatapi.cpp

@@ -1,10 +1,10 @@
 #include "chatapi.h"
 
-#include <gpt4all-backend/llmodel.h>
+#include "utils.h"
 
 #include <QCoreApplication>
-#include <QGuiApplication>
 #include <QDebug>
+#include <QGuiApplication>
 #include <QJsonArray>
 #include <QJsonDocument>
 #include <QJsonObject>
@@ -13,12 +13,17 @@
 #include <QNetworkRequest>
 #include <QThread>
 #include <QUrl>
+#include <QUtf8StringView>
 #include <QVariant>
+#include <QXmlStreamReader>
 #include <Qt>
 #include <QtGlobal>
 #include <QtLogging>
 
+#include <expected>
+#include <functional>
 #include <iostream>
+#include <utility>
 
 using namespace Qt::Literals::StringLiterals;
 
@@ -51,7 +56,6 @@ bool ChatAPI::loadModel(const std::string &modelPath, int n_ctx, int ngl)
 void ChatAPI::setThreadCount(int32_t n_threads)
 {
     Q_UNUSED(n_threads);
-    qt_noop();
 }
 
 int32_t ChatAPI::threadCount() const
@@ -68,89 +72,119 @@ bool ChatAPI::isModelLoaded() const
     return true;
 }
 
-// All three of the state virtual functions are handled custom inside of chatllm save/restore
-size_t ChatAPI::stateSize() const
+static auto parsePrompt(QXmlStreamReader &xml) -> std::expected<QJsonArray, QString>
 {
-    throw std::logic_error("not implemented");
+    QJsonArray messages;
+
+    auto xmlError = [&xml] {
+        return std::unexpected(u"%1:%2: %3"_s.arg(xml.lineNumber()).arg(xml.columnNumber()).arg(xml.errorString()));
+    };
+
+    if (xml.hasError())
+        return xmlError();
+    if (xml.atEnd())
+        return messages;
+
+    // skip header
+    bool foundElement = false;
+    do {
+        switch (xml.readNext()) {
+            using enum QXmlStreamReader::TokenType;
+        case Invalid:
+            return xmlError();
+        case EndDocument:
+            return messages;
+        default:
+            foundElement = true;
+        case StartDocument:
+        case Comment:
+        case DTD:
+        case ProcessingInstruction:
+            ;
+        }
+    } while (!foundElement);
+
+    // document body loop
+    bool foundRoot = false;
+    for (;;) {
+        switch (xml.tokenType()) {
+            using enum QXmlStreamReader::TokenType;
+        case StartElement:
+            {
+                auto name = xml.name();
+                if (!foundRoot) {
+                    if (name != "chat"_L1)
+                        return std::unexpected(u"unexpected tag: %1"_s.arg(name));
+                    foundRoot = true;
+                } else {
+                    if (name != "user"_L1 && name != "assistant"_L1 && name != "system"_L1)
+                        return std::unexpected(u"unknown role: %1"_s.arg(name));
+                    auto content = xml.readElementText();
+                    if (xml.tokenType() != EndElement)
+                        return xmlError();
+                    messages << makeJsonObject({
+                        { "role"_L1,    name.toString().trimmed() },
+                        { "content"_L1, content                   },
+                    });
+                }
+                break;
+            }
+        case Characters:
+            if (!xml.isWhitespace())
+                return std::unexpected(u"unexpected text: %1"_s.arg(xml.text()));
+        case Comment:
+        case ProcessingInstruction:
+        case EndElement:
+            break;
+        case EndDocument:
+            return messages;
+        case Invalid:
+            return xmlError();
+        default:
+            return std::unexpected(u"unexpected token: %1"_s.arg(xml.tokenString()));
+        }
+        xml.readNext();
+    }
 }
 
-size_t ChatAPI::saveState(std::span<uint8_t> dest) const
-{
-    Q_UNUSED(dest);
-    throw std::logic_error("not implemented");
-}
+void ChatAPI::prompt(
+    std::string_view        prompt,
+    const PromptCallback   &promptCallback,
+    const ResponseCallback &responseCallback,
+    const PromptContext    &promptCtx
+) {
+    Q_UNUSED(promptCallback)
 
-size_t ChatAPI::restoreState(std::span<const uint8_t> src)
-{
-    Q_UNUSED(src);
-    throw std::logic_error("not implemented");
-}
-
-void ChatAPI::prompt(const std::string &prompt,
-                     const std::string &promptTemplate,
-                     std::function<bool(int32_t)> promptCallback,
-                     std::function<bool(int32_t, const std::string&)> responseCallback,
-                     bool allowContextShift,
-                     PromptContext &promptCtx,
-                     bool special,
-                     std::optional<std::string_view> fakeReply) {
-
-    Q_UNUSED(promptCallback);
-    Q_UNUSED(allowContextShift);
-    Q_UNUSED(special);
-
-    if (!isModelLoaded()) {
-        std::cerr << "ChatAPI ERROR: prompt won't work with an unloaded model!\n";
-        return;
-    }
-
-    if (!promptCtx.n_past) { m_queuedPrompts.clear(); }
-    Q_ASSERT(promptCtx.n_past <= m_context.size());
-    m_context.resize(promptCtx.n_past);
-
-    // FIXME(cebtenzzre): We're assuming people don't try to use %2 with ChatGPT. What would that even mean?
-    m_queuedPrompts << QString::fromStdString(promptTemplate).arg(QString::fromStdString(prompt));
-
-    if (!promptCtx.n_predict && !fakeReply) {
-        return; // response explicitly suppressed, queue prompt for later
-    }
-
-    QString formattedPrompt = m_queuedPrompts.join("");
-    m_queuedPrompts.clear();
-
-    if (fakeReply) {
-        promptCtx.n_past += 1;
-        m_context.append(formattedPrompt);
-        m_context.append(QString::fromUtf8(fakeReply->data(), fakeReply->size()));
-        return;
-    }
+    if (!isModelLoaded())
+        throw std::invalid_argument("Attempted to prompt an unloaded model.");
+    if (!promptCtx.n_predict)
+        return; // nothing requested
 
     // FIXME: We don't set the max_tokens on purpose because in order to do so safely without encountering
     // an error we need to be able to count the tokens in our prompt. The only way to do this is to use
-    // the OpenAI tiktokken library or to implement our own tokenization function that matches precisely
+    // the OpenAI tiktoken library or to implement our own tokenization function that matches precisely
     // the tokenization used by the OpenAI model we're calling. OpenAI has not introduced any means of
     // using the REST API to count tokens in a prompt.
-    QJsonObject root;
-    root.insert("model", m_modelName);
-    root.insert("stream", true);
-    root.insert("temperature", promptCtx.temp);
-    root.insert("top_p", promptCtx.top_p);
+    auto root = makeJsonObject({
+        { "model"_L1,       m_modelName     },
+        { "stream"_L1,      true            },
+        { "temperature"_L1, promptCtx.temp  },
+        { "top_p"_L1,       promptCtx.top_p },
+    });
 
     // conversation history
-    QJsonArray messages;
-    for (int i = 0; i < m_context.count(); ++i) {
-        QJsonObject message;
-        message.insert("role", i % 2 == 0 ? "user" : "assistant");
-        message.insert("content", m_context.at(i));
-        messages.append(message);
+    {
+        QUtf8StringView promptUtf8(prompt);
+        QXmlStreamReader xml(promptUtf8);
+        auto messages = parsePrompt(xml);
+        if (!messages) {
+            auto error = fmt::format("Failed to parse API model prompt: {}", messages.error());
+            qDebug().noquote() << "ChatAPI ERROR:" << error << "Prompt:\n\n" << promptUtf8 << '\n';
+            throw std::invalid_argument(error);
+        }
+        root.insert("messages"_L1, *messages);
     }
 
-    QJsonObject promptObject;
-    promptObject.insert("role", "user");
-    promptObject.insert("content", formattedPrompt);
-    messages.append(promptObject);
-    root.insert("messages", messages);
-
     QJsonDocument doc(root);
 
 #if defined(DEBUG)
@@ -167,12 +201,9 @@ void ChatAPI::prompt(const std::string &prompt,
     connect(&worker, &ChatAPIWorker::finished, &workerThread, &QThread::quit, Qt::DirectConnection);
     connect(this, &ChatAPI::request, &worker, &ChatAPIWorker::request, Qt::QueuedConnection);
     workerThread.start();
-    emit request(m_apiKey, &promptCtx, doc.toJson(QJsonDocument::Compact));
+    emit request(m_apiKey, doc.toJson(QJsonDocument::Compact));
     workerThread.wait();
 
-    promptCtx.n_past += 1;
-    m_context.append(formattedPrompt);
-    m_context.append(worker.currentResponse());
     m_responseCallback = nullptr;
 
 #if defined(DEBUG)
@@ -190,12 +221,8 @@ bool ChatAPI::callResponse(int32_t token, const std::string& string)
     return m_responseCallback(token, string);
 }
 
-void ChatAPIWorker::request(const QString &apiKey,
-                            LLModel::PromptContext *promptCtx,
-                            const QByteArray &array)
+void ChatAPIWorker::request(const QString &apiKey, const QByteArray &array)
 {
-    m_ctx = promptCtx;
-
     QUrl apiUrl(m_chat->url());
     const QString authorization = u"Bearer %1"_s.arg(apiKey).trimmed();
     QNetworkRequest request(apiUrl);
@@ -302,7 +329,6 @@ void ChatAPIWorker::handleReadyRead()
         const QJsonObject choice = choices.first().toObject();
         const QJsonObject delta = choice.value("delta").toObject();
         const QString content = delta.value("content").toString();
-        Q_ASSERT(m_ctx);
         m_currentResponse += content;
         if (!m_chat->callResponse(0, content.toStdString())) {
             reply->abort();

gpt4all-chat/src/chatapi.h

@@ -3,19 +3,18 @@
 
 #include <gpt4all-backend/llmodel.h>
 
-#include <QByteArray>
+#include <QByteArray> // IWYU pragma: keep
 #include <QNetworkReply>
 #include <QObject>
 #include <QString>
-#include <QStringList>
-#include <QList>
 
 #include <cstddef>
 #include <cstdint>
-#include <functional>
+#include <span>
 #include <stdexcept>
 #include <string>
 #include <string_view>
+#include <unordered_map>
 #include <vector>
 
 class QNetworkAccessManager;
@@ -26,16 +25,13 @@ class ChatAPIWorker : public QObject {
 public:
     ChatAPIWorker(ChatAPI *chatAPI)
         : QObject(nullptr)
-        , m_ctx(nullptr)
         , m_networkManager(nullptr)
         , m_chat(chatAPI) {}
     virtual ~ChatAPIWorker() {}
 
     QString currentResponse() const { return m_currentResponse; }
 
-    void request(const QString &apiKey,
-                 LLModel::PromptContext *promptCtx,
-                 const QByteArray &array);
+    void request(const QString &apiKey, const QByteArray &array);
 
 Q_SIGNALS:
     void finished();
@@ -47,7 +43,6 @@ private Q_SLOTS:
 
 private:
     ChatAPI *m_chat;
-    LLModel::PromptContext *m_ctx;
     QNetworkAccessManager *m_networkManager;
     QString m_currentResponse;
 };
@@ -63,17 +58,23 @@ public:
     bool loadModel(const std::string &modelPath, int n_ctx, int ngl) override;
     bool isModelLoaded() const override;
     size_t requiredMem(const std::string &modelPath, int n_ctx, int ngl) override;
-    size_t stateSize() const override;
-    size_t saveState(std::span<uint8_t> dest) const override;
-    size_t restoreState(std::span<const uint8_t> src) override;
-    void prompt(const std::string &prompt,
-                const std::string &promptTemplate,
-                std::function<bool(int32_t)> promptCallback,
-                std::function<bool(int32_t, const std::string&)> responseCallback,
-                bool allowContextShift,
-                PromptContext &ctx,
-                bool special,
-                std::optional<std::string_view> fakeReply) override;
+
+    // All three of the state virtual functions are handled custom inside of chatllm save/restore
+    size_t stateSize() const override
+    { throwNotImplemented(); }
+    size_t saveState(std::span<uint8_t> stateOut, std::vector<Token> &inputTokensOut) const override
+    { Q_UNUSED(stateOut); Q_UNUSED(inputTokensOut); throwNotImplemented(); }
+    size_t restoreState(std::span<const uint8_t> state, std::span<const Token> inputTokens) override
+    { Q_UNUSED(state); Q_UNUSED(inputTokens); throwNotImplemented(); }
+
+    void prompt(std::string_view        prompt,
+                const PromptCallback   &promptCallback,
+                const ResponseCallback &responseCallback,
+                const PromptContext    &ctx) override;
+
+    [[noreturn]]
+    int32_t countPromptTokens(std::string_view prompt) const override
+    { Q_UNUSED(prompt); throwNotImplemented(); }
 
     void setThreadCount(int32_t n_threads) override;
     int32_t threadCount() const override;
@@ -83,84 +84,87 @@ public:
     void setRequestURL(const QString &requestURL) { m_requestURL = requestURL; }
     QString url() const { return m_requestURL; }
 
-    QList<QString> context() const { return m_context; }
-    void setContext(const QList<QString> &context) { m_context = context; }
-
     bool callResponse(int32_t token, const std::string &string);
 
+    [[noreturn]]
+    int32_t contextLength() const override
+    { throwNotImplemented(); }
+
+    auto specialTokens() -> std::unordered_map<std::string, std::string> const override
+    { return {}; }
+
 Q_SIGNALS:
-    void request(const QString &apiKey,
-                 LLModel::PromptContext *ctx,
-                 const QByteArray &array);
+    void request(const QString &apiKey, const QByteArray &array);
 
 protected:
     // We have to implement these as they are pure virtual in base class, but we don't actually use
     // them as they are only called from the default implementation of 'prompt' which we override and
     // completely replace
 
-    std::vector<Token> tokenize(PromptContext &ctx, std::string_view str, bool special) override
-    {
-        (void)ctx;
-        (void)str;
-        (void)special;
-        throw std::logic_error("not implemented");
-    }
+    [[noreturn]]
+    static void throwNotImplemented() { throw std::logic_error("not implemented"); }
 
+    [[noreturn]]
+    std::vector<Token> tokenize(std::string_view str) const override
+    { Q_UNUSED(str); throwNotImplemented(); }
+
+    [[noreturn]]
     bool isSpecialToken(Token id) const override
-    {
-        (void)id;
-        throw std::logic_error("not implemented");
-    }
+    { Q_UNUSED(id); throwNotImplemented(); }
 
+    [[noreturn]]
     std::string tokenToString(Token id) const override
-    {
-        (void)id;
-        throw std::logic_error("not implemented");
-    }
+    { Q_UNUSED(id); throwNotImplemented(); }
 
-    void initSampler(PromptContext &ctx) override
-    {
-        (void)ctx;
-        throw std::logic_error("not implemented");
-    }
+    [[noreturn]]
+    void initSampler(const PromptContext &ctx) override
+    { Q_UNUSED(ctx); throwNotImplemented(); }
 
-    Token sampleToken() const override { throw std::logic_error("not implemented"); }
+    [[noreturn]]
+    Token sampleToken() const override
+    { throwNotImplemented(); }
 
-    bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override
-    {
-        (void)ctx;
-        (void)tokens;
-        throw std::logic_error("not implemented");
-    }
+    [[noreturn]]
+    bool evalTokens(int32_t nPast, std::span<const Token> tokens) const override
+    { Q_UNUSED(nPast); Q_UNUSED(tokens); throwNotImplemented(); }
 
-    void shiftContext(PromptContext &promptCtx) override
-    {
-        (void)promptCtx;
-        throw std::logic_error("not implemented");
-    }
+    [[noreturn]]
+    void shiftContext(const PromptContext &promptCtx, int32_t *nPast) override
+    { Q_UNUSED(promptCtx); Q_UNUSED(nPast); throwNotImplemented(); }
 
-    int32_t contextLength() const override
-    {
-        throw std::logic_error("not implemented");
-    }
+    [[noreturn]]
+    int32_t inputLength() const override
+    { throwNotImplemented(); }
 
+    [[noreturn]]
+    int32_t computeModelInputPosition(std::span<const Token> input) const override
+    { Q_UNUSED(input); throwNotImplemented(); }
+
+    [[noreturn]]
+    void setModelInputPosition(int32_t pos) override
+    { Q_UNUSED(pos); throwNotImplemented(); }
+
+    [[noreturn]]
+    void appendInputToken(Token tok) override
+    { Q_UNUSED(tok); throwNotImplemented(); }
+
+    [[noreturn]]
     const std::vector<Token> &endTokens() const override
-    {
-        throw std::logic_error("not implemented");
-    }
+    { throwNotImplemented(); }
 
+    [[noreturn]]
     bool shouldAddBOS() const override
-    {
-        throw std::logic_error("not implemented");
-    }
+    { throwNotImplemented(); }
+
+    [[noreturn]]
+    std::span<const Token> inputTokens() const override
+    { throwNotImplemented(); }
 
 private:
-    std::function<bool(int32_t, const std::string&)> m_responseCallback;
-    QString m_modelName;
-    QString m_apiKey;
-    QString m_requestURL;
-    QList<QString> m_context;
-    QStringList m_queuedPrompts;
+    ResponseCallback m_responseCallback;
+    QString          m_modelName;
+    QString          m_apiKey;
+    QString          m_requestURL;
 };
 
 #endif // CHATAPI_H

gpt4all-chat/src/chatlistmodel.cpp

@@ -17,9 +17,10 @@
 #include <Qt>
 
 #include <algorithm>
+#include <memory>
 
-#define CHAT_FORMAT_MAGIC 0xF5D553CC
-#define CHAT_FORMAT_VERSION 10
+static constexpr quint32 CHAT_FORMAT_MAGIC   = 0xF5D553CC;
+static constexpr qint32  CHAT_FORMAT_VERSION = 11;
 
 class MyChatListModel: public ChatListModel { };
 Q_GLOBAL_STATIC(MyChatListModel, chatListModelInstance)
@@ -51,6 +52,10 @@ void ChatListModel::loadChats()
     connect(thread, &ChatsRestoreThread::finished, thread, &QObject::deleteLater);
     thread->start();
 
+    ChatSaver *saver = new ChatSaver;
+    connect(this, &ChatListModel::requestSaveChats, saver, &ChatSaver::saveChats, Qt::QueuedConnection);
+    connect(saver, &ChatSaver::saveChatsFinished, this, &ChatListModel::saveChatsFinished, Qt::QueuedConnection);
+
     connect(MySettings::globalInstance(), &MySettings::serverChatChanged, this, &ChatListModel::handleServerEnabledChanged);
 }
 
@@ -88,9 +93,6 @@ void ChatListModel::saveChats()
         return;
     }
 
-    ChatSaver *saver = new ChatSaver;
-    connect(this, &ChatListModel::requestSaveChats, saver, &ChatSaver::saveChats, Qt::QueuedConnection);
-    connect(saver, &ChatSaver::saveChatsFinished, this, &ChatListModel::saveChatsFinished, Qt::QueuedConnection);
     emit requestSaveChats(toSave);
 }
 
@@ -99,7 +101,12 @@ void ChatSaver::saveChats(const QVector<Chat *> &chats)
     QElapsedTimer timer;
     timer.start();
     const QString savePath = MySettings::globalInstance()->modelPath();
+    qsizetype nSavedChats = 0;
     for (Chat *chat : chats) {
+        if (!chat->needsSave())
+            continue;
+        ++nSavedChats;
+
         QString fileName = "gpt4all-" + chat->id() + ".chat";
         QString filePath = savePath + "/" + fileName;
         QFile originalFile(filePath);
@@ -112,8 +119,8 @@ void ChatSaver::saveChats(const QVector<Chat *> &chats)
         }
         QDataStream out(&tempFile);
 
-        out << (quint32)CHAT_FORMAT_MAGIC;
-        out << (qint32)CHAT_FORMAT_VERSION;
+        out << CHAT_FORMAT_MAGIC;
+        out << CHAT_FORMAT_VERSION;
         out.setVersion(QDataStream::Qt_6_2);
 
         qDebug() << "serializing chat" << fileName;
@@ -123,13 +130,14 @@ void ChatSaver::saveChats(const QVector<Chat *> &chats)
             continue;
         }
 
+        chat->setNeedsSave(false);
         if (originalFile.exists())
             originalFile.remove();
         tempFile.rename(filePath);
     }
 
     qint64 elapsedTime = timer.elapsed();
-    qDebug() << "serializing chats took:" << elapsedTime << "ms";
+    qDebug() << "serializing chats took" << elapsedTime << "ms, saved" << nSavedChats << "/" << chats.size() << "chats";
     emit saveChatsFinished();
 }
 
@@ -194,11 +202,16 @@ void ChatsRestoreThread::run()
             qint32 version;
             in >> version;
             if (version < 1) {
-                qWarning() << "ERROR: Chat file has non supported version:" << file.fileName();
+                qWarning() << "WARNING: Chat file version" << version << "is not supported:" << file.fileName();
+                continue;
+            }
+            if (version > CHAT_FORMAT_VERSION) {
+                qWarning().nospace() << "WARNING: Chat file is from a future version (have " << version << " want "
+                                     << CHAT_FORMAT_VERSION << "): " << file.fileName();
                 continue;
             }
 
-            if (version <= 1)
+            if (version < 2)
                 in.setVersion(QDataStream::Qt_6_2);
 
             FileInfo info;
@@ -239,18 +252,21 @@ void ChatsRestoreThread::run()
                 continue;
             }
 
-            if (version <= 1)
+            if (version < 2)
                 in.setVersion(QDataStream::Qt_6_2);
         }
 
         qDebug() << "deserializing chat" << f.file;
 
-        Chat *chat = new Chat;
+        auto chat = std::make_unique<Chat>();
         chat->moveToThread(qGuiApp->thread());
-        if (!chat->deserialize(in, version)) {
+        bool ok = chat->deserialize(in, version);
+        if (!ok) {
             qWarning() << "ERROR: Couldn't deserialize chat from file:" << file.fileName();
+        } else if (!in.atEnd()) {
+            qWarning().nospace() << "error loading chat from " << file.fileName() << ": extra data at end of file";
         } else {
-            emit chatRestored(chat);
+            emit chatRestored(chat.release());
         }
         if (f.oldFile)
            file.remove(); // No longer storing in this directory

gpt4all-chat/src/chatlistmodel.h

@@ -33,7 +33,6 @@ class ChatSaver : public QObject
     Q_OBJECT
 public:
     explicit ChatSaver();
-    void stop();
 
 Q_SIGNALS:
     void saveChatsFinished();
@@ -147,7 +146,7 @@ public:
         if (m_serverChat)
             return;
 
-        m_serverChat = new Chat(true /*isServer*/, this);
+        m_serverChat = new Chat(Chat::server_tag, this);
         beginInsertRows(QModelIndex(), m_chats.size(), m_chats.size());
         m_chats.append(m_serverChat);
         endInsertRows();
@@ -238,7 +237,6 @@ public Q_SLOTS:
 Q_SIGNALS:
     void countChanged();
     void currentChatChanged();
-    void chatsSavedFinished();
     void requestSaveChats(const QVector<Chat*> &);
     void saveChatsFinished();
 

gpt4all-chat/src/chatllm.h

@@ -9,31 +9,84 @@
 #include <QByteArray>
 #include <QElapsedTimer>
 #include <QFileInfo>
-#include <QList>
+#include <QList>      // IWYU pragma: keep
 #include <QObject>
 #include <QPointer>
 #include <QString>
+#include <QStringList> // IWYU pragma: keep
+#include <QStringView>
 #include <QThread>
-#include <QVariantMap>
+#include <QVariantMap> // IWYU pragma: keep
 #include <QtGlobal>
 
 #include <atomic>
 #include <cstdint>
 #include <memory>
 #include <optional>
+#include <span>
 #include <string>
+#include <variant>
 
 using namespace Qt::Literals::StringLiterals;
 
 class QDataStream;
+struct ChatItem;
 
 // NOTE: values serialized to disk, do not change or reuse
-enum LLModelType {
-    GPTJ_  = 0, // no longer used
-    LLAMA_ = 1,
-    API_   = 2,
-    BERT_  = 3, // no longer used
+enum class LLModelTypeV0 { // chat versions 2-5
+    MPT       = 0,
+    GPTJ      = 1,
+    LLAMA     = 2,
+    CHATGPT   = 3,
+    REPLIT    = 4,
+    FALCON    = 5,
+    BERT      = 6, // not used
+    STARCODER = 7,
 };
+enum class LLModelTypeV1 { // since chat version 6 (v2.5.0)
+    GPTJ      = 0, // not for new chats
+    LLAMA     = 1,
+    API       = 2,
+    BERT      = 3, // not used
+    // none of the below are used in new chats
+    REPLIT    = 4,
+    FALCON    = 5,
+    MPT       = 6,
+    STARCODER = 7,
+    NONE      = -1, // no state
+};
+
+inline LLModelTypeV1 parseLLModelTypeV1(int type)
+{
+    switch (LLModelTypeV1(type)) {
+    case LLModelTypeV1::GPTJ:
+    case LLModelTypeV1::LLAMA:
+    case LLModelTypeV1::API:
+    // case LLModelTypeV1::BERT: -- not used
+    case LLModelTypeV1::REPLIT:
+    case LLModelTypeV1::FALCON:
+    case LLModelTypeV1::MPT:
+    case LLModelTypeV1::STARCODER:
+        return LLModelTypeV1(type);
+    default:
+        return LLModelTypeV1::NONE;
+    }
+}
+
+inline LLModelTypeV1 parseLLModelTypeV0(int v0)
+{
+    switch (LLModelTypeV0(v0)) {
+    case LLModelTypeV0::MPT:       return LLModelTypeV1::MPT;
+    case LLModelTypeV0::GPTJ:      return LLModelTypeV1::GPTJ;
+    case LLModelTypeV0::LLAMA:     return LLModelTypeV1::LLAMA;
+    case LLModelTypeV0::CHATGPT:   return LLModelTypeV1::API;
+    case LLModelTypeV0::REPLIT:    return LLModelTypeV1::REPLIT;
+    case LLModelTypeV0::FALCON:    return LLModelTypeV1::FALCON;
+    // case LLModelTypeV0::BERT: -- not used
+    case LLModelTypeV0::STARCODER: return LLModelTypeV1::STARCODER;
+    default:                       return LLModelTypeV1::NONE;
+    }
+}
 
 class ChatLLM;
 class ChatModel;
@@ -93,7 +146,6 @@ class Chat;
 class ChatLLM : public QObject
 {
     Q_OBJECT
-    Q_PROPERTY(bool restoringFromText READ restoringFromText NOTIFY restoringFromTextChanged)
     Q_PROPERTY(QString deviceBackend READ deviceBackend NOTIFY loadedModelInfoChanged)
     Q_PROPERTY(QString device READ device NOTIFY loadedModelInfoChanged)
     Q_PROPERTY(QString fallbackReason READ fallbackReason NOTIFY loadedModelInfoChanged)
@@ -101,12 +153,14 @@ public:
     ChatLLM(Chat *parent, bool isServer = false);
     virtual ~ChatLLM();
 
-    void destroy();
     static void destroyStore();
+    static std::optional<std::string> checkJinjaTemplateError(const std::string &source);
+
+    void destroy();
     bool isModelLoaded() const;
-    void regenerateResponse();
-    void resetResponse();
-    void resetContext();
+    void regenerateResponse(int index);
+    // used to implement edit functionality
+    std::optional<QString> popPrompt(int index);
 
     void stopGenerating() { m_stopGenerating = true; }
 
@@ -116,13 +170,9 @@ public:
     void setForceUnloadModel(bool b) { m_forceUnloadModel = b; }
     void setMarkedForDeletion(bool b) { m_markedForDeletion = b; }
 
-    QString response(bool trim = true) const;
-
     ModelInfo modelInfo() const;
     void setModelInfo(const ModelInfo &info);
 
-    bool restoringFromText() const { return m_restoringFromText; }
-
     void acquireModel();
     void resetModel();
 
@@ -147,13 +197,11 @@ public:
         return m_llModelInfo.fallbackReason.value_or(u""_s);
     }
 
-    QString generatedName() const { return QString::fromStdString(m_nameResponse); }
-
-    bool serialize(QDataStream &stream, int version, bool serializeKV);
-    bool deserialize(QDataStream &stream, int version, bool deserializeKV, bool discardKV);
+    bool serialize(QDataStream &stream, int version);
+    bool deserialize(QDataStream &stream, int version);
 
 public Q_SLOTS:
-    bool prompt(const QList<QString> &collectionList, const QString &prompt);
+    void prompt(const QStringList &enabledCollections);
     bool loadDefaultModel();
     void trySwitchContextOfLoadedModel(const ModelInfo &modelInfo);
     bool loadModel(const ModelInfo &modelInfo);
@@ -161,22 +209,19 @@ public Q_SLOTS:
     void unloadModel();
     void reloadModel();
     void generateName();
-    void generateQuestions(qint64 elapsed);
     void handleChatIdChanged(const QString &id);
     void handleShouldBeLoadedChanged();
     void handleThreadStarted();
     void handleForceMetalChanged(bool forceMetal);
     void handleDeviceChanged();
-    void processSystemPrompt();
-    void processRestoreStateFromText();
 
 Q_SIGNALS:
-    void restoringFromTextChanged();
     void loadedModelInfoChanged();
     void modelLoadingPercentageChanged(float);
     void modelLoadingError(const QString &error);
     void modelLoadingWarning(const QString &warning);
     void responseChanged(const QString &response);
+    void responseFailed(const QString &error);
     void promptProcessing();
     void generatingQuestions();
     void responseStopped(qint64 promptResponseMs);
@@ -195,55 +240,50 @@ Q_SIGNALS:
     void modelInfoChanged(const ModelInfo &modelInfo);
 
 protected:
-    bool promptInternal(const QList<QString> &collectionList, const QString &prompt, const QString &promptTemplate,
-        int32_t n_predict, int32_t top_k, float top_p, float min_p, float temp, int32_t n_batch, float repeat_penalty,
-        int32_t repeat_penalty_tokens, std::optional<QString> fakeReply = {});
-    bool handlePrompt(int32_t token);
-    bool handleResponse(int32_t token, const std::string &response);
-    bool handleNamePrompt(int32_t token);
-    bool handleNameResponse(int32_t token, const std::string &response);
-    bool handleSystemPrompt(int32_t token);
-    bool handleSystemResponse(int32_t token, const std::string &response);
-    bool handleRestoreStateFromTextPrompt(int32_t token);
-    bool handleRestoreStateFromTextResponse(int32_t token, const std::string &response);
-    bool handleQuestionPrompt(int32_t token);
-    bool handleQuestionResponse(int32_t token, const std::string &response);
-    void saveState();
-    void restoreState();
+    struct PromptResult {
+        QByteArray response;       // raw UTF-8
+        int        promptTokens;   // note: counts *entire* history, even if cached
+        int        responseTokens;
+    };
 
-protected:
-    LLModel::PromptContext m_ctx;
-    quint32 m_promptTokens;
-    quint32 m_promptResponseTokens;
+    struct ChatPromptResult : PromptResult {
+        QList<ResultInfo> databaseResults;
+    };
+
+    ChatPromptResult promptInternalChat(const QStringList &enabledCollections, const LLModel::PromptContext &ctx);
+    // passing a string_view directly skips templating and uses the raw string
+    PromptResult promptInternal(const std::variant<std::span<const ChatItem>, std::string_view> &prompt,
+                                const LLModel::PromptContext &ctx,
+                                bool usedLocalDocs);
 
 private:
     bool loadNewModel(const ModelInfo &modelInfo, QVariantMap &modelLoadProps);
 
-    std::string m_response;
-    std::string m_trimmedResponse;
-    std::string m_nameResponse;
-    QString m_questionResponse;
+    std::vector<ChatItem> forkConversation(const QString &prompt) const;
+
+    // Applies the Jinja template. Query mode returns only the last message without special tokens.
+    // Returns a (# of messages, rendered prompt) pair.
+    std::string applyJinjaTemplate(std::span<const ChatItem> items) const;
+
+    void generateQuestions(qint64 elapsed);
+
+protected:
+    QPointer<ChatModel> m_chatModel;
+
+private:
+    const Chat *m_chat;
     LLModelInfo m_llModelInfo;
-    LLModelType m_llModelType;
+    LLModelTypeV1 m_llModelType = LLModelTypeV1::NONE;
     ModelInfo m_modelInfo;
     TokenTimer *m_timer;
-    QByteArray m_state;
     QThread m_llmThread;
     std::atomic<bool> m_stopGenerating;
     std::atomic<bool> m_shouldBeLoaded;
-    std::atomic<bool> m_restoringFromText; // status indication
     std::atomic<bool> m_forceUnloadModel;
     std::atomic<bool> m_markedForDeletion;
     bool m_isServer;
     bool m_forceMetal;
     bool m_reloadingToChangeVariant;
-    bool m_processedSystemPrompt;
-    bool m_restoreStateFromText;
-    // m_pristineLoadedState is set if saveSate is unnecessary, either because:
-    // - an unload was queued during LLModel::restoreState()
-    // - the chat will be restored from text and hasn't been interacted with yet
-    bool m_pristineLoadedState = false;
-    QPointer<ChatModel> m_chatModel;
 };
 
 #endif // CHATLLM_H

gpt4all-chat/src/chatmodel.h

@@ -2,8 +2,11 @@
 #define CHATMODEL_H
 
 #include "database.h"
+#include "utils.h"
 #include "xlsxtomd.h"
 
+#include <fmt/format.h>
+
 #include <QAbstractListModel>
 #include <QBuffer>
 #include <QByteArray>
@@ -18,6 +21,15 @@
 #include <Qt>
 #include <QtGlobal>
 
+#include <iterator>
+#include <ranges>
+#include <span>
+#include <utility>
+
+using namespace Qt::Literals::StringLiterals;
+namespace ranges = std::ranges;
+
+
 struct PromptAttachment {
     Q_GADGET
     Q_PROPERTY(QUrl url MEMBER url)
@@ -40,12 +52,17 @@ public:
 
     QString processedContent() const
     {
+        const QString localFilePath = url.toLocalFile();
+        const QFileInfo info(localFilePath);
+        if (info.suffix().toLower() != "xlsx")
+            return u"## Attached: %1\n\n%2"_s.arg(file(), content);
+
         QBuffer buffer;
         buffer.setData(content);
         buffer.open(QIODevice::ReadOnly);
         const QString md = XLSXToMD::toMarkdown(&buffer);
         buffer.close();
-        return md;
+        return u"## Attached: %1\n\n%2"_s.arg(file(), md);
     }
 
     bool operator==(const PromptAttachment &other) const { return url == other.url; }
@@ -55,69 +72,145 @@ Q_DECLARE_METATYPE(PromptAttachment)
 struct ChatItem
 {
     Q_GADGET
-    Q_PROPERTY(int id MEMBER id)
-    Q_PROPERTY(QString name MEMBER name)
+    Q_PROPERTY(QString name  MEMBER name )
     Q_PROPERTY(QString value MEMBER value)
-    Q_PROPERTY(QString newResponse MEMBER newResponse)
-    Q_PROPERTY(bool currentResponse MEMBER currentResponse)
-    Q_PROPERTY(bool stopped MEMBER stopped)
-    Q_PROPERTY(bool thumbsUpState MEMBER thumbsUpState)
-    Q_PROPERTY(bool thumbsDownState MEMBER thumbsDownState)
-    Q_PROPERTY(QList<ResultInfo> sources MEMBER sources)
-    Q_PROPERTY(QList<ResultInfo> consolidatedSources MEMBER consolidatedSources)
-    Q_PROPERTY(QList<PromptAttachment> promptAttachments MEMBER promptAttachments);
-    Q_PROPERTY(QString promptPlusAttachments READ promptPlusAttachments);
+
+    // prompts
+    Q_PROPERTY(QList<PromptAttachment> promptAttachments MEMBER promptAttachments)
+    Q_PROPERTY(QString                 bakedPrompt       READ   bakedPrompt      )
+
+    // responses
+    Q_PROPERTY(bool isCurrentResponse MEMBER isCurrentResponse)
+    Q_PROPERTY(bool isError           MEMBER isError          )
+
+    // responses (DataLake)
+    Q_PROPERTY(QString newResponse     MEMBER newResponse    )
+    Q_PROPERTY(bool    stopped         MEMBER stopped        )
+    Q_PROPERTY(bool    thumbsUpState   MEMBER thumbsUpState  )
+    Q_PROPERTY(bool    thumbsDownState MEMBER thumbsDownState)
 
 public:
-    QString promptPlusAttachments() const
-    {
-        QStringList attachedContexts;
-        for (auto attached : promptAttachments)
-            attachedContexts << attached.processedContent();
+    enum class Type { System, Prompt, Response };
 
-        QString promptPlus = value;
-        if (!attachedContexts.isEmpty())
-            promptPlus = attachedContexts.join("\n\n") + "\n\n" + value;
-        return promptPlus;
+    // tags for constructing ChatItems
+    struct prompt_tag_t { explicit prompt_tag_t() = default; };
+    static inline constexpr prompt_tag_t prompt_tag = prompt_tag_t();
+    struct response_tag_t { explicit response_tag_t() = default; };
+    static inline constexpr response_tag_t response_tag = response_tag_t();
+    struct system_tag_t { explicit system_tag_t() = default; };
+    static inline constexpr system_tag_t system_tag = system_tag_t();
+
+    // FIXME(jared): This should not be necessary. QML should see null or undefined if it
+    // tries to access something invalid.
+    ChatItem() = default;
+
+    // NOTE: system messages are currently never stored in the model or serialized
+    ChatItem(system_tag_t, const QString &value)
+        : name(u"System: "_s), value(value) {}
+
+    ChatItem(prompt_tag_t, const QString &value, const QList<PromptAttachment> &attachments = {})
+        : name(u"Prompt: "_s), value(value), promptAttachments(attachments) {}
+
+    ChatItem(response_tag_t, bool isCurrentResponse = true)
+        : name(u"Response: "_s), isCurrentResponse(isCurrentResponse) {}
+
+    Type type() const
+    {
+        if (name == u"System: "_s)
+            return Type::System;
+        if (name == u"Prompt: "_s)
+            return Type::Prompt;
+        if (name == u"Response: "_s)
+            return Type::Response;
+        throw std::invalid_argument(fmt::format("Chat item has unknown label: {:?}", name));
+    }
+
+    // used with version 0 Jinja templates
+    QString bakedPrompt() const
+    {
+        if (type() != Type::Prompt)
+            throw std::logic_error("bakedPrompt() called on non-prompt item");
+        QStringList parts;
+        if (!sources.isEmpty()) {
+            parts << u"### Context:\n"_s;
+            for (auto &source : std::as_const(sources))
+                parts << u"Collection: "_s << source.collection
+                      << u"\nPath: "_s     << source.path
+                      << u"\nExcerpt: "_s  << source.text << u"\n\n"_s;
+        }
+        for (auto &attached : std::as_const(promptAttachments))
+            parts << attached.processedContent() << u"\n\n"_s;
+        parts << value;
+        return parts.join(QString());
     }
 
     // TODO: Maybe we should include the model name here as well as timestamp?
-    int id = 0;
     QString name;
     QString value;
-    QString newResponse;
-    QList<ResultInfo> sources;
-    QList<ResultInfo> consolidatedSources;
+
+    // prompts
+    QList<ResultInfo>       sources;
+    QList<ResultInfo>       consolidatedSources;
     QList<PromptAttachment> promptAttachments;
-    bool currentResponse = false;
-    bool stopped = false;
-    bool thumbsUpState = false;
-    bool thumbsDownState = false;
+
+    // responses
+    bool isCurrentResponse = false;
+    bool isError           = false;
+
+    // responses (DataLake)
+    QString newResponse;
+    bool    stopped         = false;
+    bool    thumbsUpState   = false;
+    bool    thumbsDownState = false;
 };
 Q_DECLARE_METATYPE(ChatItem)
 
-using ChatModelIterator = QList<ChatItem>::const_iterator;
+class ChatModelAccessor : public ranges::subrange<QList<ChatItem>::const_iterator> {
+private:
+    using Super = ranges::subrange<QList<ChatItem>::const_iterator>;
+
+public:
+    template <typename... T>
+    ChatModelAccessor(QMutex &mutex, T &&...args)
+        : Super(std::forward<T>(args)...), m_lock(&mutex) {}
+
+private:
+    QMutexLocker<QMutex> m_lock;
+};
 
 class ChatModel : public QAbstractListModel
 {
     Q_OBJECT
     Q_PROPERTY(int count READ count NOTIFY countChanged)
+    Q_PROPERTY(bool hasError READ hasError NOTIFY hasErrorChanged)
 
 public:
-    explicit ChatModel(QObject *parent = nullptr) : QAbstractListModel(parent) {}
+    explicit ChatModel(QObject *parent = nullptr)
+        : QAbstractListModel(parent) {}
 
+    // FIXME(jared): can't this start at Qt::UserRole (no +1)?
     enum Roles {
-        IdRole = Qt::UserRole + 1,
-        NameRole,
+        NameRole = Qt::UserRole + 1,
         ValueRole,
+
+        // prompts and responses
+        PeerRole,
+
+        // prompts
+        PromptAttachmentsRole,
+
+        // responses
+        // NOTE: sources are stored on the *prompts*, but in the model, they are only on the *responses*!
+        SourcesRole,
+        ConsolidatedSourcesRole,
+        IsCurrentResponseRole,
+        IsErrorRole,
+
+        // responses (DataLake)
         NewResponseRole,
-        CurrentResponseRole,
         StoppedRole,
         ThumbsUpStateRole,
         ThumbsDownStateRole,
-        SourcesRole,
-        ConsolidatedSourcesRole,
-        PromptAttachmentsRole
     };
 
     int rowCount(const QModelIndex &parent = QModelIndex()) const override
@@ -127,36 +220,96 @@ public:
         return m_chatItems.size();
     }
 
+    /* a "peer" is a bidirectional 1:1 link between a prompt and the response that would cite its LocalDocs
+     * sources. Return std::nullopt if there is none, which is possible for e.g. server chats. */
+    auto getPeerUnlocked(QList<ChatItem>::const_iterator item) const
+        -> std::optional<QList<ChatItem>::const_iterator>
+    {
+        switch (item->type()) {
+            using enum ChatItem::Type;
+        case Prompt:
+            {
+                auto peer = std::next(item);
+                if (peer < m_chatItems.cend() && peer->type() == Response)
+                    return peer;
+                break;
+            }
+        case Response:
+            {
+                if (item > m_chatItems.cbegin()) {
+                    if (auto peer = std::prev(item); peer->type() == Prompt)
+                        return peer;
+                }
+                break;
+            }
+        default:
+            throw std::invalid_argument("getPeer() called on item that is not a prompt or response");
+        }
+        return std::nullopt;
+    }
+
+    auto getPeerUnlocked(int index) const -> std::optional<int>
+    {
+        return getPeerUnlocked(m_chatItems.cbegin() + index)
+            .transform([&](auto &&i) { return i - m_chatItems.cbegin(); } );
+    }
+
     QVariant data(const QModelIndex &index, int role = Qt::DisplayRole) const override
     {
         QMutexLocker locker(&m_mutex);
         if (!index.isValid() || index.row() < 0 || index.row() >= m_chatItems.size())
             return QVariant();
 
-        const ChatItem &item = m_chatItems.at(index.row());
+        auto item = m_chatItems.cbegin() + index.row();
         switch (role) {
-            case IdRole:
-                return item.id;
             case NameRole:
-                return item.name;
+                return item->name;
             case ValueRole:
-                return item.value;
-            case NewResponseRole:
-                return item.newResponse;
-            case CurrentResponseRole:
-                return item.currentResponse;
-            case StoppedRole:
-                return item.stopped;
-            case ThumbsUpStateRole:
-                return item.thumbsUpState;
-            case ThumbsDownStateRole:
-                return item.thumbsDownState;
-            case SourcesRole:
-                return QVariant::fromValue(item.sources);
-            case ConsolidatedSourcesRole:
-                return QVariant::fromValue(item.consolidatedSources);
+                return item->value;
+            case PeerRole:
+                switch (item->type()) {
+                    using enum ChatItem::Type;
+                case Prompt:
+                case Response:
+                    {
+                        auto peer = getPeerUnlocked(item);
+                        return peer ? QVariant::fromValue(**peer) : QVariant::fromValue(nullptr);
+                    }
+                default:
+                    return QVariant();
+                }
             case PromptAttachmentsRole:
-                return QVariant::fromValue(item.promptAttachments);
+                return QVariant::fromValue(item->promptAttachments);
+            case SourcesRole:
+                {
+                    QList<ResultInfo> data;
+                    if (item->type() == ChatItem::Type::Response) {
+                        if (auto prompt = getPeerUnlocked(item))
+                            data = (*prompt)->consolidatedSources;
+                    }
+                    return QVariant::fromValue(data);
+                }
+            case ConsolidatedSourcesRole:
+                {
+                    QList<ResultInfo> data;
+                    if (item->type() == ChatItem::Type::Response) {
+                        if (auto prompt = getPeerUnlocked(item))
+                            data = (*prompt)->sources;
+                    }
+                    return QVariant::fromValue(data);
+                }
+            case IsCurrentResponseRole:
+                return item->isCurrentResponse;
+            case NewResponseRole:
+                return item->newResponse;
+            case StoppedRole:
+                return item->stopped;
+            case ThumbsUpStateRole:
+                return item->thumbsUpState;
+            case ThumbsDownStateRole:
+                return item->thumbsDownState;
+            case IsErrorRole:
+                return item->type() == ChatItem::Type::Response && item->isError;
         }
 
         return QVariant();
@@ -164,56 +317,119 @@ public:
 
     QHash<int, QByteArray> roleNames() const override
     {
-        QHash<int, QByteArray> roles;
-        roles[IdRole] = "id";
-        roles[NameRole] = "name";
-        roles[ValueRole] = "value";
-        roles[NewResponseRole] = "newResponse";
-        roles[CurrentResponseRole] = "currentResponse";
-        roles[StoppedRole] = "stopped";
-        roles[ThumbsUpStateRole] = "thumbsUpState";
-        roles[ThumbsDownStateRole] = "thumbsDownState";
-        roles[SourcesRole] = "sources";
-        roles[ConsolidatedSourcesRole] = "consolidatedSources";
-        roles[PromptAttachmentsRole] = "promptAttachments";
-        return roles;
+        return {
+            { NameRole,                "name"                },
+            { ValueRole,               "value"               },
+            { PeerRole,                "peer"                },
+            { PromptAttachmentsRole,   "promptAttachments"   },
+            { SourcesRole,             "sources"             },
+            { ConsolidatedSourcesRole, "consolidatedSources" },
+            { IsCurrentResponseRole,   "isCurrentResponse"   },
+            { IsErrorRole,             "isError"             },
+            { NewResponseRole,         "newResponse"         },
+            { StoppedRole,             "stopped"             },
+            { ThumbsUpStateRole,       "thumbsUpState"       },
+            { ThumbsDownStateRole,     "thumbsDownState"     },
+        };
     }
 
-    void appendPrompt(const QString &name, const QString &value, const QList<PromptAttachment> &attachments)
+    void appendPrompt(const QString &value, const QList<PromptAttachment> &attachments = {})
     {
-        ChatItem item;
-        item.name = name;
-        item.value = value;
-        item.promptAttachments << attachments;
+        qsizetype count;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (hasErrorUnlocked())
+                throw std::logic_error("cannot append to a failed chat");
+            count = m_chatItems.count();
+        }
 
-        m_mutex.lock();
-        const int count = m_chatItems.count();
-        m_mutex.unlock();
         beginInsertRows(QModelIndex(), count, count);
         {
             QMutexLocker locker(&m_mutex);
-            m_chatItems.append(item);
+            m_chatItems.emplace_back(ChatItem::prompt_tag, value, attachments);
         }
         endInsertRows();
         emit countChanged();
     }
 
-    void appendResponse(const QString &name)
+    void appendResponse()
     {
-        m_mutex.lock();
-        const int count = m_chatItems.count();
-        m_mutex.unlock();
-        ChatItem item;
-        item.id = count; // This is only relevant for responses
-        item.name = name;
-        item.currentResponse = true;
+        qsizetype count;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (hasErrorUnlocked())
+                throw std::logic_error("cannot append to a failed chat");
+            count = m_chatItems.count();
+        }
+
+        int promptIndex = 0;
         beginInsertRows(QModelIndex(), count, count);
         {
             QMutexLocker locker(&m_mutex);
-            m_chatItems.append(item);
+            m_chatItems.emplace_back(ChatItem::response_tag, promptIndex);
+            if (auto pi = getPeerUnlocked(m_chatItems.size() - 1))
+                promptIndex = *pi;
         }
         endInsertRows();
         emit countChanged();
+        if (promptIndex >= 0)
+            emit dataChanged(createIndex(promptIndex, 0), createIndex(promptIndex, 0), {PeerRole});
+    }
+
+    // Used by Server to append a new conversation to the chat log.
+    void appendResponseWithHistory(std::span<const ChatItem> history)
+    {
+        if (history.empty())
+            throw std::invalid_argument("at least one message is required");
+
+        m_mutex.lock();
+        qsizetype startIndex = m_chatItems.count();
+        m_mutex.unlock();
+
+        qsizetype nNewItems = history.size() + 1;
+        qsizetype endIndex  = startIndex + nNewItems;
+        beginInsertRows(QModelIndex(), startIndex, endIndex - 1 /*inclusive*/);
+        bool hadError;
+        {
+            QMutexLocker locker(&m_mutex);
+            hadError = hasErrorUnlocked();
+            m_chatItems.reserve(m_chatItems.count() + nNewItems);
+            for (auto &item : history)
+                m_chatItems << item;
+            m_chatItems.emplace_back(ChatItem::response_tag);
+        }
+        endInsertRows();
+        emit countChanged();
+        // Server can add messages when there is an error because each call is a new conversation
+        if (hadError)
+            emit hasErrorChanged(false);
+    }
+
+    void truncate(qsizetype size)
+    {
+        qsizetype oldSize;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (size >= (oldSize = m_chatItems.size()))
+                return;
+            if (size && m_chatItems.at(size - 1).type() != ChatItem::Type::Response)
+                throw std::invalid_argument(
+                    fmt::format("chat model truncated to {} items would not end in a response", size)
+                );
+        }
+
+        bool oldHasError;
+        beginRemoveRows(QModelIndex(), size, oldSize - 1 /*inclusive*/);
+        {
+            QMutexLocker locker(&m_mutex);
+            oldHasError = hasErrorUnlocked();
+            Q_ASSERT(size < m_chatItems.size());
+            m_chatItems.resize(size);
+        }
+        endRemoveRows();
+        emit countChanged();
+        if (oldHasError)
+            emit hasErrorChanged(false);
     }
 
     Q_INVOKABLE void clear()
@@ -223,13 +439,17 @@ public:
             if (m_chatItems.isEmpty()) return;
         }
 
+        bool oldHasError;
         beginResetModel();
         {
             QMutexLocker locker(&m_mutex);
+            oldHasError = hasErrorUnlocked();
             m_chatItems.clear();
         }
         endResetModel();
         emit countChanged();
+        if (oldHasError)
+            emit hasErrorChanged(false);
     }
 
     Q_INVOKABLE ChatItem get(int index)
@@ -247,13 +467,13 @@ public:
             if (index < 0 || index >= m_chatItems.size()) return;
 
             ChatItem &item = m_chatItems[index];
-            if (item.currentResponse != b) {
-                item.currentResponse = b;
+            if (item.isCurrentResponse != b) {
+                item.isCurrentResponse = b;
                 changed = true;
             }
         }
 
-        if (changed) emit dataChanged(createIndex(index, 0), createIndex(index, 0), {CurrentResponseRole});
+        if (changed) emit dataChanged(createIndex(index, 0), createIndex(index, 0), {IsCurrentResponseRole});
     }
 
     Q_INVOKABLE void updateStopped(int index, bool b)
@@ -306,16 +526,23 @@ public:
 
     Q_INVOKABLE void updateSources(int index, const QList<ResultInfo> &sources)
     {
+        int responseIndex = -1;
         {
             QMutexLocker locker(&m_mutex);
             if (index < 0 || index >= m_chatItems.size()) return;
 
-            ChatItem &item = m_chatItems[index];
-            item.sources = sources;
-            item.consolidatedSources = consolidateSources(sources);
+            auto promptItem = m_chatItems.begin() + index;
+            if (promptItem->type() != ChatItem::Type::Prompt)
+                throw std::invalid_argument(fmt::format("item at index {} is not a prompt", index));
+            if (auto peer = getPeerUnlocked(promptItem))
+                responseIndex = *peer - m_chatItems.cbegin();
+            promptItem->sources = sources;
+            promptItem->consolidatedSources = consolidateSources(sources);
+        }
+        if (responseIndex >= 0) {
+            emit dataChanged(createIndex(responseIndex, 0), createIndex(responseIndex, 0), {SourcesRole});
+            emit dataChanged(createIndex(responseIndex, 0), createIndex(responseIndex, 0), {ConsolidatedSourcesRole});
         }
-        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {SourcesRole});
-        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {ConsolidatedSourcesRole});
     }
 
     Q_INVOKABLE void updateThumbsUpState(int index, bool b)
@@ -366,27 +593,70 @@ public:
         if (changed) emit dataChanged(createIndex(index, 0), createIndex(index, 0), {NewResponseRole});
     }
 
-    int count() const { QMutexLocker locker(&m_mutex); return m_chatItems.size(); }
+    Q_INVOKABLE void setError(bool value = true)
+    {
+        qsizetype index;
+        {
+            QMutexLocker locker(&m_mutex);
 
-    ChatModelIterator begin() const { return m_chatItems.begin(); }
-    ChatModelIterator end() const { return m_chatItems.end(); }
-    void lock() { m_mutex.lock(); }
-    void unlock() { m_mutex.unlock(); }
+            if (m_chatItems.isEmpty() || m_chatItems.cend()[-1].type() != ChatItem::Type::Response)
+                throw std::logic_error("can only set error on a chat that ends with a response");
+
+            index = m_chatItems.count() - 1;
+            auto &last = m_chatItems.back();
+            if (last.isError == value)
+                return; // already set
+            last.isError = value;
+        }
+        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {IsErrorRole});
+        emit hasErrorChanged(value);
+    }
+
+    qsizetype count() const { QMutexLocker locker(&m_mutex); return m_chatItems.size(); }
+
+    ChatModelAccessor chatItems() const { return {m_mutex, std::as_const(m_chatItems)}; }
+
+    bool hasError() const { QMutexLocker locker(&m_mutex); return hasErrorUnlocked(); }
 
     bool serialize(QDataStream &stream, int version) const
     {
         QMutexLocker locker(&m_mutex);
         stream << int(m_chatItems.size());
-        for (const auto &c : m_chatItems) {
-            stream << c.id;
+        for (auto itemIt = m_chatItems.cbegin(); itemIt < m_chatItems.cend(); ++itemIt) {
+            auto c = *itemIt; // NB: copies
+            if (version < 11) {
+                // move sources from their prompt to the next response
+                switch (c.type()) {
+                    using enum ChatItem::Type;
+                case Prompt:
+                    c.sources.clear();
+                    c.consolidatedSources.clear();
+                    break;
+                case Response:
+                    // note: we drop sources for responseless prompts
+                    if (auto peer = getPeerUnlocked(itemIt)) {
+                        c.sources             = (*peer)->sources;
+                        c.consolidatedSources = (*peer)->consolidatedSources;
+                    }
+                default:
+                    ;
+                }
+            }
+
+            // FIXME: This 'id' should be eliminated the next time we bump serialization version.
+            // (Jared) This was apparently never used.
+            int id = 0;
+            stream << id;
             stream << c.name;
             stream << c.value;
             stream << c.newResponse;
-            stream << c.currentResponse;
+            stream << c.isCurrentResponse;
             stream << c.stopped;
             stream << c.thumbsUpState;
             stream << c.thumbsDownState;
-            if (version > 7) {
+            if (version >= 11 && c.type() == ChatItem::Type::Response)
+                stream << c.isError;
+            if (version >= 8) {
                 stream << c.sources.size();
                 for (const ResultInfo &info : c.sources) {
                     Q_ASSERT(!info.file.isEmpty());
@@ -401,7 +671,7 @@ public:
                     stream << info.from;
                     stream << info.to;
                 }
-            } else if (version > 2) {
+            } else if (version >= 3) {
                 QList<QString> references;
                 QList<QString> referencesContext;
                 int validReferenceNumber = 1;
@@ -451,12 +721,24 @@ public:
 
     bool deserialize(QDataStream &stream, int version)
     {
+        clear(); // reset to known state
+
         int size;
         stream >> size;
+        int lastPromptIndex = -1;
+        QList<ChatItem> chatItems;
         for (int i = 0; i < size; ++i) {
             ChatItem c;
-            stream >> c.id;
+            // FIXME: see comment in serialization about id
+            int id;
+            stream >> id;
             stream >> c.name;
+            try {
+                c.type(); // check name
+            } catch (const std::exception &e) {
+                qWarning() << "ChatModel ERROR:" << e.what();
+                return false;
+            }
             stream >> c.value;
             if (version < 10) {
                 // This is deprecated and no longer used
@@ -464,11 +746,13 @@ public:
                 stream >> prompt;
             }
             stream >> c.newResponse;
-            stream >> c.currentResponse;
+            stream >> c.isCurrentResponse;
             stream >> c.stopped;
             stream >> c.thumbsUpState;
             stream >> c.thumbsDownState;
-            if (version > 7) {
+            if (version >= 11 && c.type() == ChatItem::Type::Response)
+                stream >> c.isError;
+            if (version >= 8) {
                 qsizetype count;
                 stream >> count;
                 QList<ResultInfo> sources;
@@ -488,7 +772,7 @@ public:
                 }
                 c.sources = sources;
                 c.consolidatedSources = consolidateSources(sources);
-            } else if (version > 2) {
+            } else if (version >= 3) {
                 QString references;
                 QList<QString> referencesContext;
                 stream >> references;
@@ -584,23 +868,53 @@ public:
                 }
                 c.promptAttachments = attachments;
             }
-            m_mutex.lock();
-            const int count = m_chatItems.size();
-            m_mutex.unlock();
-            beginInsertRows(QModelIndex(), count, count);
-            {
-                QMutexLocker locker(&m_mutex);
-                m_chatItems.append(c);
+
+            if (version < 11 && c.type() == ChatItem::Type::Response) {
+                // move sources from the response to their last prompt
+                if (lastPromptIndex >= 0) {
+                    auto &prompt = chatItems[lastPromptIndex];
+                    prompt.sources             = std::move(c.sources            );
+                    prompt.consolidatedSources = std::move(c.consolidatedSources);
+                    lastPromptIndex = -1;
+                } else {
+                    // drop sources for promptless responses
+                    c.sources.clear();
+                    c.consolidatedSources.clear();
+                }
             }
-            endInsertRows();
+
+            chatItems << c;
+            if (c.type() == ChatItem::Type::Prompt)
+                lastPromptIndex = chatItems.size() - 1;
         }
+
+        bool hasError;
+        beginInsertRows(QModelIndex(), 0, chatItems.size() - 1 /*inclusive*/);
+        {
+            QMutexLocker locker(&m_mutex);
+            m_chatItems = chatItems;
+            hasError = hasErrorUnlocked();
+        }
+        endInsertRows();
         emit countChanged();
+        if (hasError)
+            emit hasErrorChanged(true);
         return stream.status() == QDataStream::Ok;
     }
 
 Q_SIGNALS:
     void countChanged();
     void valueChanged(int index, const QString &value);
+    void hasErrorChanged(bool value);
+
+private:
+    bool hasErrorUnlocked() const
+    {
+        if (m_chatItems.isEmpty())
+            return false;
+        auto &last = m_chatItems.back();
+        return last.type() == ChatItem::Type::Response && last.isError;
+    }
 
 private:
     mutable QMutex m_mutex;

gpt4all-chat/src/chatviewtextprocessor.cpp

@@ -738,7 +738,7 @@ void SyntaxHighlighter::highlightBlock(const QString &text)
     case Java:
         rules = javaHighlightingRules(); break;
     case Go:
-        rules = javaHighlightingRules(); break;
+        rules = goHighlightingRules(); break;
     case Json:
         rules = jsonHighlightingRules(); break;
     case Latex:
@@ -967,8 +967,6 @@ void ChatViewTextProcessor::handleCodeBlocks()
         cursor.setPosition(matchesCode[index].capturedEnd(), QTextCursor::KeepAnchor);
         cursor.removeSelectedText();
 
-        int startPos = cursor.position();
-
         QTextFrameFormat frameFormat = frameFormatBase;
         QString capturedText = matchesCode[index].captured(1);
         QString codeLanguage;
@@ -1004,7 +1002,7 @@ void ChatViewTextProcessor::handleCodeBlocks()
         QTextFrame *mainFrame = cursor.currentFrame();
         cursor.setCharFormat(textFormat);
 
-        QTextFrame *frame = cursor.insertFrame(frameFormat);
+        cursor.insertFrame(frameFormat);
         QTextTable *table = cursor.insertTable(codeLanguage.isEmpty() ? 1 : 2, 1, tableFormat);
 
         if (!codeLanguage.isEmpty()) {
@@ -1016,7 +1014,6 @@ void ChatViewTextProcessor::handleCodeBlocks()
             headerCursor.insertText(codeLanguage);
             QTextTableCell copy = headerTable->cellAt(0, 1);
             QTextCursor copyCursor = copy.firstCursorPosition();
-            int startPos = copyCursor.position();
             CodeCopy newCopy;
             newCopy.text = lines.join("\n");
             newCopy.startPos = copyCursor.position();

gpt4all-chat/src/database.cpp

@@ -32,6 +32,7 @@
 #include <stdexcept>
 
 using namespace Qt::Literals::StringLiterals;
+namespace ranges = std::ranges;
 namespace us = unum::usearch;
 
 //#define DEBUG
@@ -175,6 +176,14 @@ static const QString INSERT_CHUNK_FTS_SQL = uR"(
             values(?, ?, ?, ?, ?, ?, ?);
 )"_s;
 
+static const QString SELECT_CHUNKED_DOCUMENTS_SQL[] = {
+    uR"(
+        select distinct document_id from chunks;
+    )"_s, uR"(
+        select distinct document_id from chunks_fts;
+    )"_s,
+};
+
 static const QString DELETE_CHUNKS_SQL[] = {
     uR"(
         delete from embeddings
@@ -224,58 +233,16 @@ static const QString SELECT_COUNT_CHUNKS_SQL = uR"(
 )"_s;
 
 static const QString SELECT_CHUNKS_FTS_SQL = uR"(
-    select id, bm25(chunks_fts) as score
-    from chunks_fts
+    select fts.id, bm25(chunks_fts) as score
+    from chunks_fts fts
+    join documents d on fts.document_id = d.id
+    join collection_items ci on d.folder_id = ci.folder_id
+    join collections co on ci.collection_id = co.id
     where chunks_fts match ?
-    order by score limit %1;
+    and co.name in ('%1')
+    order by score limit %2;
 )"_s;
 
-static bool addChunk(QSqlQuery &q, int document_id, const QString &chunk_text, const QString &file,
-                     const QString &title, const QString &author, const QString &subject, const QString &keywords,
-                     int page, int from, int to, int words, int *chunk_id)
-{
-    if (!q.prepare(INSERT_CHUNK_SQL))
-        return false;
-    q.addBindValue(document_id);
-    q.addBindValue(chunk_text);
-    q.addBindValue(file);
-    q.addBindValue(title);
-    q.addBindValue(author);
-    q.addBindValue(subject);
-    q.addBindValue(keywords);
-    q.addBindValue(page);
-    q.addBindValue(from);
-    q.addBindValue(to);
-    q.addBindValue(words);
-    if (!q.exec() || !q.next())
-        return false;
-    *chunk_id = q.value(0).toInt();
-
-    if (!q.prepare(INSERT_CHUNK_FTS_SQL))
-        return false;
-    q.addBindValue(document_id);
-    q.addBindValue(chunk_text);
-    q.addBindValue(file);
-    q.addBindValue(title);
-    q.addBindValue(author);
-    q.addBindValue(subject);
-    q.addBindValue(keywords);
-    if (!q.exec())
-        return false;
-    return true;
-}
-
-static bool removeChunksByDocumentId(QSqlQuery &q, int document_id)
-{
-    for (const auto &cmd: DELETE_CHUNKS_SQL) {
-        if (!q.prepare(cmd))
-            return false;
-        q.addBindValue(document_id);
-        if (!q.exec())
-            return false;
-    }
-    return true;
-}
 
 #define NAMED_PAIR(name, typea, a, typeb, b) \
     struct name { typea a; typeb b; }; \
@@ -296,10 +263,12 @@ static bool selectAllUncompletedChunks(QSqlQuery &q, QHash<IncompleteChunk, QStr
     while (q.next()) {
         QString collection = q.value(0).toString();
         IncompleteChunk ic {
-            /*embedding_model =*/ q.value(1).toString(),
-            /*chunk_id        =*/ q.value(2).toInt(),
-            /*folder_id       =*/ q.value(3).toInt(),
-            /*text            =*/ q.value(4).toString(),
+            /*EmbeddingKey*/ {
+                .embedding_model = q.value(1).toString(),
+                .chunk_id        = q.value(2).toInt(),
+            },
+            /*folder_id =*/ q.value(3).toInt(),
+            /*text      =*/ q.value(4).toString(),
         };
         chunks[ic] << collection;
     }
@@ -321,7 +290,7 @@ static bool selectCountChunks(QSqlQuery &q, int folder_id, int &count)
     return true;
 }
 
-static bool selectChunk(QSqlQuery &q, const QList<int> &chunk_ids, int retrievalSize)
+static bool selectChunk(QSqlQuery &q, const QList<int> &chunk_ids)
 {
     QString chunk_ids_str = QString::number(chunk_ids[0]);
     for (size_t i = 1; i < chunk_ids.size(); ++i)
@@ -338,10 +307,6 @@ static const QString INSERT_COLLECTION_SQL = uR"(
         returning id;
     )"_s;
 
-static const QString DELETE_COLLECTION_SQL = uR"(
-    delete from collections where name = ? and folder_id = ?;
-    )"_s;
-
 static const QString SELECT_FOLDERS_FROM_COLLECTIONS_SQL = uR"(
     select f.id, f.path
     from collections c
@@ -385,6 +350,14 @@ static const QString UPDATE_LAST_UPDATE_TIME_SQL = uR"(
     update collections set last_update_time = ? where id = ?;
 )"_s;
 
+static const QString FTS_INTEGRITY_SQL = uR"(
+    insert into chunks_fts(chunks_fts, rank) values('integrity-check', 1);
+)"_s;
+
+static const QString FTS_REBUILD_SQL = uR"(
+    insert into chunks_fts(chunks_fts) values('rebuild');
+)"_s;
+
 static bool addCollection(QSqlQuery &q, const QString &collection_name, const QDateTime &start_update,
                           const QDateTime &last_update, const QString &embedding_model, CollectionItem &item)
 {
@@ -402,15 +375,6 @@ static bool addCollection(QSqlQuery &q, const QString &collection_name, const QD
     return true;
 }
 
-static bool removeCollection(QSqlQuery &q, const QString &collection_name, int folder_id)
-{
-    if (!q.prepare(DELETE_COLLECTION_SQL))
-        return false;
-    q.addBindValue(collection_name);
-    q.addBindValue(folder_id);
-    return q.exec();
-}
-
 static bool selectFoldersFromCollection(QSqlQuery &q, const QString &collection_name, QList<QPair<int, QString>> *folders)
 {
     if (!q.prepare(SELECT_FOLDERS_FROM_COLLECTIONS_SQL))
@@ -543,10 +507,6 @@ static const QString GET_FOLDER_EMBEDDING_MODEL_SQL = uR"(
     where ci.folder_id = ?;
     )"_s;
 
-static const QString SELECT_ALL_FOLDERPATHS_SQL = uR"(
-    select path from folders;
-    )"_s;
-
 static const QString FOLDER_REMOVE_ALL_DOCS_SQL[] = {
     uR"(
         delete from embeddings
@@ -621,29 +581,6 @@ static bool sqlGetFolderEmbeddingModel(QSqlQuery &q, int id, QString &embedding_
     return true;
 }
 
-static bool selectAllFolderPaths(QSqlQuery &q, QList<QString> *folder_paths)
-{
-    if (!q.prepare(SELECT_ALL_FOLDERPATHS_SQL))
-        return false;
-    if (!q.exec())
-        return false;
-    while (q.next())
-        folder_paths->append(q.value(0).toString());
-    return true;
-}
-
-static bool sqlRemoveDocsByFolderPath(QSqlQuery &q, const QString &path)
-{
-    for (const auto &cmd: FOLDER_REMOVE_ALL_DOCS_SQL) {
-        if (!q.prepare(cmd))
-            return false;
-        q.addBindValue(path);
-        if (!q.exec())
-            return false;
-    }
-    return true;
-}
-
 static const QString INSERT_COLLECTION_ITEM_SQL = uR"(
     insert into collection_items(collection_id, folder_id)
     values(?, ?)
@@ -887,6 +824,79 @@ void Database::rollback()
     Q_ASSERT(ok);
 }
 
+bool Database::refreshDocumentIdCache(QSqlQuery &q)
+{
+    m_documentIdCache.clear();
+    for (const auto &cmd: SELECT_CHUNKED_DOCUMENTS_SQL) {
+        if (!q.exec(cmd))
+            return false;
+        while (q.next())
+            m_documentIdCache << q.value(0).toInt();
+    }
+    return true;
+}
+
+bool Database::addChunk(QSqlQuery &q, int document_id, const QString &chunk_text, const QString &file,
+                        const QString &title, const QString &author, const QString &subject, const QString &keywords,
+                        int page, int from, int to, int words, int *chunk_id)
+{
+    if (!q.prepare(INSERT_CHUNK_SQL))
+        return false;
+    q.addBindValue(document_id);
+    q.addBindValue(chunk_text);
+    q.addBindValue(file);
+    q.addBindValue(title);
+    q.addBindValue(author);
+    q.addBindValue(subject);
+    q.addBindValue(keywords);
+    q.addBindValue(page);
+    q.addBindValue(from);
+    q.addBindValue(to);
+    q.addBindValue(words);
+    if (!q.exec() || !q.next())
+        return false;
+    *chunk_id = q.value(0).toInt();
+
+    if (!q.prepare(INSERT_CHUNK_FTS_SQL))
+        return false;
+    q.addBindValue(document_id);
+    q.addBindValue(chunk_text);
+    q.addBindValue(file);
+    q.addBindValue(title);
+    q.addBindValue(author);
+    q.addBindValue(subject);
+    q.addBindValue(keywords);
+    if (!q.exec())
+        return false;
+    m_documentIdCache << document_id;
+    return true;
+}
+
+bool Database::removeChunksByDocumentId(QSqlQuery &q, int document_id)
+{
+    for (const auto &cmd: DELETE_CHUNKS_SQL) {
+        if (!q.prepare(cmd))
+            return false;
+        q.addBindValue(document_id);
+        if (!q.exec())
+            return false;
+    }
+    m_documentIdCache.remove(document_id);
+    return true;
+}
+
+bool Database::sqlRemoveDocsByFolderPath(QSqlQuery &q, const QString &path)
+{
+    for (const auto &cmd: FOLDER_REMOVE_ALL_DOCS_SQL) {
+        if (!q.prepare(cmd))
+            return false;
+        q.addBindValue(path);
+        if (!q.exec())
+            return false;
+    }
+    return refreshDocumentIdCache(q);
+}
+
 bool Database::hasContent()
 {
     return m_db.tables().contains("chunks", Qt::CaseInsensitive);
@@ -1091,9 +1101,12 @@ static void handleDocumentError(const QString &errorMessage, int document_id, co
 
 class DocumentReader {
 public:
+    struct Metadata { QString title, author, subject, keywords; };
+
     static std::unique_ptr<DocumentReader> fromDocument(const DocumentInfo &info);
 
     const DocumentInfo           &doc     () const { return *m_info; }
+    const Metadata               &metadata() const { return m_metadata; }
     const std::optional<QString> &word    () const { return m_word; }
     const std::optional<QString> &nextWord()       { m_word = advance(); return m_word; }
     virtual std::optional<ChunkStreamer::Status> getError() const { return std::nullopt; }
@@ -1105,11 +1118,16 @@ protected:
     explicit DocumentReader(const DocumentInfo &info)
         : m_info(&info) {}
 
-    void postInit() { m_word = advance(); }
+    void postInit(Metadata &&metadata = {})
+    {
+        m_metadata = std::move(metadata);
+        m_word = advance();
+    }
 
     virtual std::optional<QString> advance() = 0;
 
     const DocumentInfo     *m_info;
+    Metadata                m_metadata;
     std::optional<QString>  m_word;
 };
 
@@ -1123,7 +1141,13 @@ public:
         QString path = info.file.canonicalFilePath();
         if (m_doc.load(path) != QPdfDocument::Error::None)
             throw std::runtime_error(fmt::format("Failed to load PDF: {}", path));
-        postInit();
+        Metadata metadata {
+            .title    = m_doc.metaData(QPdfDocument::MetaDataField::Title   ).toString(),
+            .author   = m_doc.metaData(QPdfDocument::MetaDataField::Author  ).toString(),
+            .subject  = m_doc.metaData(QPdfDocument::MetaDataField::Subject ).toString(),
+            .keywords = m_doc.metaData(QPdfDocument::MetaDataField::Keywords).toString(),
+        };
+        postInit(std::move(metadata));
     }
 
     int page() const override { return m_currentPage; }
@@ -1162,6 +1186,7 @@ public:
 
         m_paragraph = &m_doc.paragraphs();
         m_run       = &m_paragraph->runs();
+        // TODO(jared): metadata for Word documents?
         postInit();
     }
 
@@ -1183,11 +1208,14 @@ protected:
         qsizetype wordEnd = wordStart + 1;
         while (wordEnd >= m_buffer.size() || !m_buffer[wordEnd].isSpace()) {
             if (wordEnd >= m_buffer.size() && !fillBuffer())
-                return std::nullopt;
+                break;
             if (!m_buffer[wordEnd].isSpace())
                 ++wordEnd;
         }
 
+        if (wordStart == wordEnd)
+            return std::nullopt;
+
         auto size = wordEnd - wordStart;
         QString word = std::move(m_buffer);
         m_buffer = word.sliced(wordStart + size);
@@ -1195,7 +1223,6 @@ protected:
             word.resize(size);
         else
             word = word.sliced(wordStart, size);
-
         return word;
     }
 
@@ -1207,18 +1234,30 @@ protected:
                 // try next paragraph
                 if (!m_paragraph->has_next())
                     return false;
+
                 m_paragraph->next();
                 m_buffer += u'\n';
             }
+
+            bool foundText = false;
             auto &run = m_run->get_node();
-            const char *text = run.child("w:t").text().get();
-            if (!*text && run.child("w:tab"))
-                text = "\t";
-            m_run->next();
-            if (*text) {
-                m_buffer += QUtf8StringView(text);
-                return true;
+            for (auto node = run.first_child(); node; node = node.next_sibling()) {
+                std::string node_name = node.name();
+                if (node_name == "w:t") {
+                    const char *text = node.text().get();
+                    if (*text) {
+                        foundText = true;
+                        m_buffer += QUtf8StringView(text);
+                    }
+                } else if (node_name == "w:br") {
+                    m_buffer += u'\n';
+                } else if (node_name == "w:tab") {
+                    m_buffer += u'\t';
+                }
             }
+
+            m_run->next();
+            if (foundText) return true;
         }
     }
 
@@ -1244,9 +1283,13 @@ public:
 protected:
     std::optional<QString> advance() override
     {
+        if (getError())
+            return std::nullopt;
         while (!m_stream.atEnd()) {
             QString word;
             m_stream >> word;
+            if (getError())
+                return std::nullopt;
             if (!word.isEmpty())
                 return word;
         }
@@ -1255,9 +1298,11 @@ protected:
 
     std::optional<ChunkStreamer::Status> getError() const override
     {
-        if (!m_file.error())
-            return std::nullopt;
-        return m_file.binarySeen() ? ChunkStreamer::Status::BINARY_SEEN : ChunkStreamer::Status::ERROR;
+        if (m_file.binarySeen())
+            return ChunkStreamer::Status::BINARY_SEEN;
+        if (m_file.error())
+            return ChunkStreamer::Status::ERROR;
+        return std::nullopt;
     }
 
     BinaryDetectingFile m_file;
@@ -1280,9 +1325,7 @@ ChunkStreamer::ChunkStreamer(Database *database)
 
 ChunkStreamer::~ChunkStreamer() = default;
 
-void ChunkStreamer::setDocument(const DocumentInfo &doc, int documentId, const QString &embeddingModel,
-                                const QString &title, const QString &author, const QString &subject,
-                                const QString &keywords)
+void ChunkStreamer::setDocument(const DocumentInfo &doc, int documentId, const QString &embeddingModel)
 {
     auto docKey = doc.key();
     if (!m_docKey || *m_docKey != docKey) {
@@ -1290,20 +1333,28 @@ void ChunkStreamer::setDocument(const DocumentInfo &doc, int documentId, const Q
         m_reader         = DocumentReader::fromDocument(doc);
         m_documentId     = documentId;
         m_embeddingModel = embeddingModel;
-        m_title          = title;
-        m_author         = author;
-        m_subject        = subject;
-        m_keywords       = keywords;
         m_chunk.clear();
         m_page = 0;
 
         // make sure the document doesn't already have any chunks
-        QSqlQuery q(m_database->m_db);
-        if (!removeChunksByDocumentId(q, documentId))
-            handleDocumentError("ERROR: Cannot remove chunks of document", documentId, doc.file.canonicalPath(), q.lastError());
+        if (m_database->m_documentIdCache.contains(documentId)) {
+            QSqlQuery q(m_database->m_db);
+            if (!m_database->removeChunksByDocumentId(q, documentId))
+                handleDocumentError("ERROR: Cannot remove chunks of document", documentId, doc.file.canonicalPath(), q.lastError());
+        }
     }
 }
 
+std::optional<DocumentInfo::key_type> ChunkStreamer::currentDocKey() const
+{
+    return m_docKey;
+}
+
+void ChunkStreamer::reset()
+{
+    m_docKey.reset();
+}
+
 ChunkStreamer::Status ChunkStreamer::step()
 {
     // TODO: implement line_from/line_to
@@ -1316,10 +1367,9 @@ ChunkStreamer::Status ChunkStreamer::step()
     Status retval;
 
     for (;;) {
-        if (auto error = m_reader->getError())
-            return *error;
-        if (m_database->scanQueueInterrupted()) {
-            retval = Status::INTERRUPTED;
+        if (auto error = m_reader->getError()) {
+            m_docKey.reset(); // done processing
+            retval = *error;
             break;
         }
 
@@ -1338,7 +1388,7 @@ ChunkStreamer::Status ChunkStreamer::step()
             }
         }
 
-        if (!word || m_chunk.length() >= maxChunkSize + 1) { // +1 for leading space
+        if (!word || m_chunk.length() >= maxChunkSize + 1) { // +1 for trailing space
             if (!m_chunk.isEmpty()) {
                 int nThisChunkWords = 0;
                 auto chunk = m_chunk; // copy
@@ -1346,39 +1396,49 @@ ChunkStreamer::Status ChunkStreamer::step()
                 // handle overlength chunks
                 if (m_chunk.length() > maxChunkSize + 1) {
                     // find the final space
-                    qsizetype lastSpace = chunk.lastIndexOf(u' ', -2);
+                    qsizetype chunkEnd = chunk.lastIndexOf(u' ', -2);
 
-                    if (lastSpace < 0) {
+                    qsizetype spaceSize;
+                    if (chunkEnd >= 0) {
                         // slice off the last word
+                        spaceSize = 1;
                         Q_ASSERT(m_nChunkWords >= 1);
-                        lastSpace = maxChunkSize;
+                        // one word left
                         nThisChunkWords = m_nChunkWords - 1;
                         m_nChunkWords = 1;
                     } else {
                         // slice the overlong word
+                        spaceSize = 0;
+                        chunkEnd = maxChunkSize;
+                        // partial word left, don't count it
                         nThisChunkWords = m_nChunkWords;
                         m_nChunkWords = 0;
                     }
-                    // save the extra part
-                    m_chunk = chunk.sliced(lastSpace + 1);
-                    // slice
-                    chunk.truncate(lastSpace + 1);
-                    Q_ASSERT(chunk.length() <= maxChunkSize + 1);
+                    // save the second part, excluding space if any
+                    m_chunk = chunk.sliced(chunkEnd + spaceSize);
+                    // consume the first part
+                    chunk.truncate(chunkEnd);
                 } else {
                     nThisChunkWords = m_nChunkWords;
                     m_nChunkWords = 0;
+                    // there is no second part
+                    m_chunk.clear();
+                    // consume the whole chunk, excluding space
+                    chunk.chop(1);
                 }
+                Q_ASSERT(chunk.length() <= maxChunkSize);
 
                 QSqlQuery q(m_database->m_db);
                 int chunkId = 0;
-                if (!addChunk(q,
+                auto &metadata = m_reader->metadata();
+                if (!m_database->addChunk(q,
                     m_documentId,
-                    chunk.chopped(1), // strip trailing space
-                    m_reader->doc().file.canonicalFilePath(),
-                    m_title,
-                    m_author,
-                    m_subject,
-                    m_keywords,
+                    chunk,
+                    m_reader->doc().file.fileName(), // basename
+                    metadata.title,
+                    metadata.author,
+                    metadata.subject,
+                    metadata.keywords,
                     m_page,
                     line_from,
                     line_to,
@@ -1397,15 +1457,19 @@ ChunkStreamer::Status ChunkStreamer::step()
                 toEmbed.chunk = chunk;
                 m_database->appendChunk(toEmbed);
                 ++nChunks;
-
-                m_chunk.clear();
             }
 
             if (!word) {
                 retval = Status::DOC_COMPLETE;
+                m_docKey.reset(); // done processing
                 break;
             }
         }
+
+        if (m_database->scanQueueInterrupted()) {
+            retval = Status::INTERRUPTED;
+            break;
+        }
     }
 
     if (nChunks) {
@@ -1466,8 +1530,22 @@ void Database::handleEmbeddingsGenerated(const QVector<EmbeddingResult> &embeddi
     for (const auto &[key, stat]: std::as_const(stats).asKeyValueRange()) {
         if (!m_collectionMap.contains(key.folder_id)) continue;
         CollectionItem item = guiCollectionItem(key.folder_id);
-        item.currentEmbeddingsToIndex -= stat.nAdded + stat.nSkipped;
-        item.totalEmbeddingsToIndex -= stat.nSkipped;
+        Q_ASSERT(item.currentEmbeddingsToIndex >= stat.nAdded + stat.nSkipped);
+        if (item.currentEmbeddingsToIndex < stat.nAdded + stat.nSkipped) {
+            qWarning() << "Database ERROR: underflow in current embeddings to index statistics";
+            item.currentEmbeddingsToIndex = 0;
+        } else {
+            item.currentEmbeddingsToIndex -= stat.nAdded + stat.nSkipped;
+        }
+
+        Q_ASSERT(item.totalEmbeddingsToIndex >= stat.nSkipped);
+        if (item.totalEmbeddingsToIndex < stat.nSkipped) {
+            qWarning() << "Database ERROR: underflow in total embeddings to index statistics";
+            item.totalEmbeddingsToIndex = 0;
+        } else {
+            item.totalEmbeddingsToIndex -= stat.nSkipped;
+        }
+
         if (!stat.lastFile.isNull())
             item.fileCurrentlyProcessing = stat.lastFile;
 
@@ -1530,8 +1608,14 @@ DocumentInfo Database::dequeueDocument()
 
 void Database::removeFolderFromDocumentQueue(int folder_id)
 {
-    if (auto it = m_docsToScan.find(folder_id); it != m_docsToScan.end())
-        m_docsToScan.erase(it);
+    if (auto queueIt = m_docsToScan.find(folder_id); queueIt != m_docsToScan.end()) {
+        if (auto key = m_chunkStreamer.currentDocKey()) {
+            if (ranges::any_of(queueIt->second, [&key](auto &d) { return d.key() == key; }))
+                m_chunkStreamer.reset(); // done with this document
+        }
+        // remove folder from queue
+        m_docsToScan.erase(queueIt);
+    }
 }
 
 void Database::enqueueDocumentInternal(DocumentInfo &&info, bool prepend)
@@ -1563,13 +1647,16 @@ bool Database::scanQueueInterrupted() const
 
 void Database::scanQueueBatch()
 {
-    m_scanDurationTimer.start();
-
     transaction();
 
-    // scan for up to 100ms or until we run out of documents
-    while (!m_docsToScan.empty() && !scanQueueInterrupted())
+    m_scanDurationTimer.start();
+
+    // scan for up to the maximum scan duration or until we run out of documents
+    while (!m_docsToScan.empty()) {
         scanQueue();
+        if (scanQueueInterrupted())
+            break;
+    }
 
     commit();
 
@@ -1655,22 +1742,8 @@ void Database::scanQueue()
     Q_ASSERT(document_id != -1);
 
     {
-        QString title, author, subject, keywords;
-        if (info.isPdf()) {
-            QPdfDocument doc;
-            if (doc.load(document_path) != QPdfDocument::Error::None) {
-                qWarning() << "ERROR: Could not load pdf" << document_id << document_path;;
-                return updateFolderToIndex(folder_id, countForFolder);
-            }
-            title    = doc.metaData(QPdfDocument::MetaDataField::Title).toString();
-            author   = doc.metaData(QPdfDocument::MetaDataField::Author).toString();
-            subject  = doc.metaData(QPdfDocument::MetaDataField::Subject).toString();
-            keywords = doc.metaData(QPdfDocument::MetaDataField::Keywords).toString();
-            // TODO(jared): metadata for Word documents?
-        }
-
         try {
-            m_chunkStreamer.setDocument(info, document_id, embedding_model, title, author, subject, keywords);
+            m_chunkStreamer.setDocument(info, document_id, embedding_model);
         } catch (const std::runtime_error &e) {
             qWarning() << "LocalDocs ERROR:" << e.what();
             goto dequeue;
@@ -1702,7 +1775,13 @@ void Database::scanQueue()
 
 dequeue:
     auto item = guiCollectionItem(folder_id);
-    item.currentBytesToIndex -= info.file.size();
+    Q_ASSERT(item.currentBytesToIndex >= info.file.size());
+    if (item.currentBytesToIndex < info.file.size()) {
+        qWarning() << "Database ERROR: underflow in current bytes to index statistics";
+        item.currentBytesToIndex = 0;
+    } else {
+        item.currentBytesToIndex -= info.file.size();
+    }
     updateGuiForCollectionItem(item);
     return updateFolderToIndex(folder_id, countForFolder);
 }
@@ -1756,7 +1835,13 @@ void Database::start()
         m_databaseValid = false;
     } else {
         cleanDB();
-        addCurrentFolders();
+        ftsIntegrityCheck();
+        QSqlQuery q(m_db);
+        if (!refreshDocumentIdCache(q)) {
+            m_databaseValid = false;
+        } else {
+            addCurrentFolders();
+        }
     }
 
     if (!m_databaseValid)
@@ -2264,7 +2349,7 @@ QList<int> Database::searchBM25(const QString &query, const QList<QString> &coll
     QList<BM25Query> bm25Queries = queriesForFTS5(query);
 
     QSqlQuery sqlQuery(m_db);
-    sqlQuery.prepare(SELECT_CHUNKS_FTS_SQL.arg(k));
+    sqlQuery.prepare(SELECT_CHUNKS_FTS_SQL.arg(collections.join("', '"), QString::number(k)));
 
     QList<SearchResult> results;
     for (auto &bm25Query : std::as_const(bm25Queries)) {
@@ -2282,11 +2367,13 @@ QList<int> Database::searchBM25(const QString &query, const QList<QString> &coll
         }
     }
 
-    do {
-        const int chunkId = sqlQuery.value(0).toInt();
-        const float score = sqlQuery.value(1).toFloat();
-        results.append({chunkId, score});
-    } while (sqlQuery.next());
+    if (sqlQuery.at() != QSql::AfterLastRow) {
+        do {
+            const int chunkId = sqlQuery.value(0).toInt();
+            const float score = sqlQuery.value(1).toFloat();
+            results.append({chunkId, score});
+        } while (sqlQuery.next());
+    }
 
     k = qMin(k, results.size());
     std::partial_sort(
@@ -2419,7 +2506,7 @@ void Database::retrieveFromDB(const QList<QString> &collections, const QString &
         return;
 
     QSqlQuery q(m_db);
-    if (!selectChunk(q, searchResults, retrievalSize)) {
+    if (!selectChunk(q, searchResults)) {
         qDebug() << "ERROR: selecting chunks:" << q.lastError();
         return;
     }
@@ -2460,6 +2547,26 @@ void Database::retrieveFromDB(const QList<QString> &collections, const QString &
             results->append(tempResults.value(id));
 }
 
+bool Database::ftsIntegrityCheck()
+{
+    QSqlQuery q(m_db);
+
+    // Returns an error executing sql if it the integrity check fails
+    // See: https://www.sqlite.org/fts5.html#the_integrity_check_command
+    const bool success = q.exec(FTS_INTEGRITY_SQL);
+    if (!success && q.lastError().nativeErrorCode() != "267" /*SQLITE_CORRUPT_VTAB from sqlite header*/) {
+        qWarning() << "ERROR: Cannot prepare sql for fts integrity check" << q.lastError();
+        return false;
+    }
+
+    if (!success && !q.exec(FTS_REBUILD_SQL)) {
+        qWarning() << "ERROR: Cannot exec sql for fts rebuild" << q.lastError();
+        return false;
+    }
+
+    return true;
+}
+
 // FIXME This is very slow and non-interruptible and when we close the application and we're
 // cleaning a large table this can cause the app to take forever to shut down. This would ideally be
 // interruptible and we'd continue 'cleaning' when we restart
@@ -2510,7 +2617,7 @@ bool Database::cleanDB()
         int document_id = q.value(0).toInt();
         QString document_path = q.value(1).toString();
         QFileInfo info(document_path);
-        if (info.exists() && info.isReadable() && m_scannedFileExtensions.contains(info.suffix()))
+        if (info.exists() && info.isReadable() && m_scannedFileExtensions.contains(info.suffix(), Qt::CaseInsensitive))
             continue;
 
 #if defined(DEBUG)

gpt4all-chat/src/database.h

@@ -41,10 +41,20 @@ class QTimer;
 
 /* Version 0: GPT4All v2.4.3, full-text search
  * Version 1: GPT4All v2.5.3, embeddings in hsnwlib
- * Version 2: GPT4All v3.0.0, embeddings in sqlite */
+ * Version 2: GPT4All v3.0.0, embeddings in sqlite
+ * Version 3: GPT4All v3.4.0, hybrid search
+ */
 
 // minimum supported version
 static const int LOCALDOCS_MIN_VER = 1;
+
+// FIXME: (Adam) The next time we bump the version we should add triggers to manage the fts external
+// content table as recommended in the official documentation to keep the fts index in sync
+// See: https://www.sqlite.org/fts5.html#external_content_tables
+
+// FIXME: (Adam) The fts virtual table should include the chunk_id explicitly instead of relying upon
+// the id of the two tables to be in sync
+
 // current version
 static const int LOCALDOCS_VERSION = 3;
 
@@ -161,8 +171,9 @@ public:
     explicit ChunkStreamer(Database *database);
     ~ChunkStreamer();
 
-    void setDocument(const DocumentInfo &doc, int documentId, const QString &embeddingModel, const QString &title,
-                     const QString &author, const QString &subject, const QString &keywords);
+    void setDocument(const DocumentInfo &doc, int documentId, const QString &embeddingModel);
+    std::optional<DocumentInfo::key_type> currentDocKey() const;
+    void reset();
 
     Status step();
 
@@ -176,7 +187,6 @@ private:
     QString                                m_author;
     QString                                m_subject;
     QString                                m_keywords;
-    bool                                   m_atStart;
 
     // working state
     QString                                m_chunk; // has a trailing space for convenience
@@ -225,6 +235,12 @@ private:
     void commit();
     void rollback();
 
+    bool addChunk(QSqlQuery &q, int document_id, const QString &chunk_text, const QString &file,
+                  const QString &title, const QString &author, const QString &subject, const QString &keywords,
+                  int page, int from, int to, int words, int *chunk_id);
+    bool refreshDocumentIdCache(QSqlQuery &q);
+    bool removeChunksByDocumentId(QSqlQuery &q, int document_id);
+    bool sqlRemoveDocsByFolderPath(QSqlQuery &q, const QString &path);
     bool hasContent();
     // not found -> 0, , exists and has content -> 1, error -> -1
     int openDatabase(const QString &modelPath, bool create = true, int ver = LOCALDOCS_VERSION);
@@ -245,6 +261,7 @@ private:
     void enqueueDocumentInternal(DocumentInfo &&info, bool prepend = false);
     void enqueueDocuments(int folder_id, std::list<DocumentInfo> &&infos);
     void scanQueue();
+    bool ftsIntegrityCheck();
     bool cleanDB();
     void addFolderToWatch(const QString &path);
     void removeFolderFromWatch(const QString &path);
@@ -294,6 +311,7 @@ private:
     QHash<int, CollectionItem> m_collectionMap; // used only for tracking indexing/embedding progress
     std::atomic<bool> m_databaseValid;
     ChunkStreamer m_chunkStreamer;
+    QSet<int> m_documentIdCache; // cached list of documents with chunks for fast lookup
 
     friend class ChunkStreamer;
 };

gpt4all-chat/src/download.cpp

@@ -58,11 +58,6 @@ Download::Download()
     m_startTime = QDateTime::currentDateTime();
 }
 
-static bool operator==(const ReleaseInfo& lhs, const ReleaseInfo& rhs)
-{
-    return lhs.version == rhs.version;
-}
-
 std::strong_ordering Download::compareAppVersions(const QString &a, const QString &b)
 {
     static QRegularExpression versionRegex(R"(^(\d+(?:\.\d+){0,2})(-.+)?$)");

gpt4all-chat/src/jinja_helpers.cpp

@@ -0,0 +1,111 @@
+#include "jinja_helpers.h"
+
+#include "utils.h"
+
+#include <fmt/format.h>
+
+#include <QString>
+#include <QUrl>
+
+#include <memory>
+#include <vector>
+
+using namespace std::literals::string_view_literals;
+
+
+JinjaResultInfo::~JinjaResultInfo() = default;
+
+const JinjaFieldMap<ResultInfo> JinjaResultInfo::s_fields = {
+    { "collection", [](auto &s) { return s.collection.toStdString(); } },
+    { "path",       [](auto &s) { return s.path      .toStdString(); } },
+    { "file",       [](auto &s) { return s.file      .toStdString(); } },
+    { "title",      [](auto &s) { return s.title     .toStdString(); } },
+    { "author",     [](auto &s) { return s.author    .toStdString(); } },
+    { "date",       [](auto &s) { return s.date      .toStdString(); } },
+    { "text",       [](auto &s) { return s.text      .toStdString(); } },
+    { "page",       [](auto &s) { return s.page;                     } },
+    { "file_uri",   [](auto &s) { return s.fileUri() .toStdString(); } },
+};
+
+JinjaPromptAttachment::~JinjaPromptAttachment() = default;
+
+const JinjaFieldMap<PromptAttachment> JinjaPromptAttachment::s_fields = {
+    { "url",               [](auto &s) { return s.url.toString()    .toStdString(); } },
+    { "file",              [](auto &s) { return s.file()            .toStdString(); } },
+    { "processed_content", [](auto &s) { return s.processedContent().toStdString(); } },
+};
+
+std::vector<std::string> JinjaMessage::GetKeys() const
+{
+    std::vector<std::string> result;
+    auto &keys = this->keys();
+    result.reserve(keys.size());
+    result.assign(keys.begin(), keys.end());
+    return result;
+}
+
+auto JinjaMessage::keys() const -> const std::unordered_set<std::string_view> &
+{
+    static const std::unordered_set<std::string_view> baseKeys
+        { "role", "content" };
+    static const std::unordered_set<std::string_view> userKeys
+        { "role", "content", "sources", "prompt_attachments" };
+    switch (m_item->type()) {
+        using enum ChatItem::Type;
+    case System:
+    case Response:
+        return baseKeys;
+    case Prompt:
+        return userKeys;
+    }
+    Q_UNREACHABLE();
+}
+
+bool operator==(const JinjaMessage &a, const JinjaMessage &b)
+{
+    if (a.m_item == b.m_item)
+        return true;
+    const auto &[ia, ib] = std::tie(*a.m_item, *b.m_item);
+    auto type = ia.type();
+    if (type != ib.type() || ia.value != ib.value)
+        return false;
+
+    switch (type) {
+        using enum ChatItem::Type;
+    case System:
+    case Response:
+        return true;
+    case Prompt:
+        return ia.sources == ib.sources && ia.promptAttachments == ib.promptAttachments;
+    }
+    Q_UNREACHABLE();
+}
+
+const JinjaFieldMap<JinjaMessage> JinjaMessage::s_fields = {
+    { "role", [](auto &m) {
+        switch (m.item().type()) {
+            using enum ChatItem::Type;
+            case System:   return "system"sv;
+            case Prompt:   return "user"sv;
+            case Response: return "assistant"sv;
+        }
+        Q_UNREACHABLE();
+    } },
+    { "content", [](auto &m) {
+        if (m.version() == 0 && m.item().type() == ChatItem::Type::Prompt)
+            return m.item().bakedPrompt().toStdString();
+        return m.item().value.toStdString();
+    } },
+    { "sources", [](auto &m) {
+        auto sources = m.item().sources | views::transform([](auto &r) {
+            return jinja2::GenericMap([map = std::make_shared<JinjaResultInfo>(r)] { return map.get(); });
+        });
+        return jinja2::ValuesList(sources.begin(), sources.end());
+    } },
+    { "prompt_attachments", [](auto &m) {
+        auto attachments = m.item().promptAttachments | views::transform([](auto &pa) {
+            return jinja2::GenericMap([map = std::make_shared<JinjaPromptAttachment>(pa)] { return map.get(); });
+        });
+        return jinja2::ValuesList(attachments.begin(), attachments.end());
+    } },
+};

gpt4all-chat/src/jinja_helpers.h

@@ -0,0 +1,116 @@
+#pragma once
+
+#include "chatmodel.h"
+#include "database.h"
+
+#include <jinja2cpp/value.h>
+
+#include <functional>
+#include <ranges>
+#include <string>
+#include <string_view>
+#include <unordered_map>
+#include <unordered_set>
+
+#include <QtGlobal>
+
+namespace views = std::views;
+
+
+template <typename T>
+using JinjaFieldMap = std::unordered_map<std::string_view, std::function<jinja2::Value (const T &)>>;
+
+template <typename Derived>
+class JinjaComparable : public jinja2::IMapItemAccessor {
+public:
+    JinjaComparable() = default;
+
+    bool IsEqual(const jinja2::IComparable &other) const override;
+
+private:
+    Q_DISABLE_COPY_MOVE(JinjaComparable)
+};
+
+template <typename Derived>
+class JinjaHelper : public JinjaComparable<Derived> {
+public:
+    size_t GetSize() const override
+    { return Derived::s_fields.size(); }
+
+    bool HasValue(const std::string &name) const override
+    { return Derived::s_fields.contains(name); }
+
+    jinja2::Value GetValueByName(const std::string &name) const override;
+
+    std::vector<std::string> GetKeys() const override
+    { auto keys = views::elements<0>(Derived::s_fields); return { keys.begin(), keys.end() }; }
+};
+
+class JinjaResultInfo : public JinjaHelper<JinjaResultInfo> {
+public:
+    explicit JinjaResultInfo(const ResultInfo &source) noexcept
+        : m_source(&source) {}
+
+    ~JinjaResultInfo() override;
+
+    const ResultInfo &value() const { return *m_source; }
+
+    friend bool operator==(const JinjaResultInfo &a, const JinjaResultInfo &b)
+    { return a.m_source == b.m_source || *a.m_source == *b.m_source; }
+
+private:
+    static const JinjaFieldMap<ResultInfo> s_fields;
+    const ResultInfo *m_source;
+
+    friend class JinjaHelper<JinjaResultInfo>;
+};
+
+class JinjaPromptAttachment : public JinjaHelper<JinjaPromptAttachment> {
+public:
+    explicit JinjaPromptAttachment(const PromptAttachment &attachment) noexcept
+        : m_attachment(&attachment) {}
+
+    ~JinjaPromptAttachment() override;
+
+    const PromptAttachment &value() const { return *m_attachment; }
+
+    friend bool operator==(const JinjaPromptAttachment &a, const JinjaPromptAttachment &b)
+    { return a.m_attachment == b.m_attachment || *a.m_attachment == *b.m_attachment; }
+
+private:
+    static const JinjaFieldMap<PromptAttachment> s_fields;
+    const PromptAttachment *m_attachment;
+
+    friend class JinjaHelper<JinjaPromptAttachment>;
+};
+
+class JinjaMessage : public JinjaHelper<JinjaMessage> {
+public:
+    explicit JinjaMessage(uint version, const ChatItem &item) noexcept
+        : m_version(version), m_item(&item) {}
+
+    const JinjaMessage &value  () const { return *this;     }
+    uint                version() const { return m_version; }
+    const ChatItem      &item  () const { return *m_item;   }
+
+    size_t GetSize() const override { return keys().size(); }
+    bool HasValue(const std::string &name) const override { return keys().contains(name); }
+
+    jinja2::Value GetValueByName(const std::string &name) const override
+    { return HasValue(name) ? JinjaHelper::GetValueByName(name) : jinja2::EmptyValue(); }
+
+    std::vector<std::string> GetKeys() const override;
+
+private:
+    auto keys() const -> const std::unordered_set<std::string_view> &;
+
+private:
+    static const JinjaFieldMap<JinjaMessage> s_fields;
+    uint            m_version;
+    const ChatItem *m_item;
+
+    friend class JinjaHelper<JinjaMessage>;
+    friend bool operator==(const JinjaMessage &a, const JinjaMessage &b);
+};
+
+#include "jinja_helpers.inl"

gpt4all-chat/src/jinja_helpers.inl

@@ -0,0 +1,17 @@
+template <typename D>
+bool JinjaComparable<D>::IsEqual(const jinja2::IComparable &other) const
+{
+    if (auto *omsg = dynamic_cast<const D *>(&other))
+        return *static_cast<const D *>(this) == *omsg;
+    return false;
+}
+
+template <typename D>
+jinja2::Value JinjaHelper<D>::GetValueByName(const std::string &name) const
+{
+    if (auto it = D::s_fields.find(name); it != D::s_fields.end()) {
+        auto [_, func] = *it;
+        return func(static_cast<const D *>(this)->value());
+    }
+    return jinja2::EmptyValue();
+}

gpt4all-chat/src/llm.cpp

@@ -19,6 +19,10 @@
 #   include "network.h"
 #endif
 
+#ifdef Q_OS_MAC
+#include "macosdock.h"
+#endif
+
 using namespace Qt::Literals::StringLiterals;
 
 class MyLLM: public LLM { };
@@ -105,3 +109,21 @@ bool LLM::isNetworkOnline() const
     auto * netinfo = QNetworkInformation::instance();
     return !netinfo || netinfo->reachability() == QNetworkInformation::Reachability::Online;
 }
+
+void LLM::showDockIcon() const
+{
+#ifdef Q_OS_MAC
+    MacOSDock::showIcon();
+#else
+    qt_noop();
+#endif
+}
+
+void LLM::hideDockIcon() const
+{
+#ifdef Q_OS_MAC
+    MacOSDock::hideIcon();
+#else
+    qt_noop();
+#endif
+}

gpt4all-chat/src/llm.h

@@ -23,6 +23,9 @@ public:
     Q_INVOKABLE QString systemTotalRAMInGBString() const;
     Q_INVOKABLE bool isNetworkOnline() const;
 
+    Q_INVOKABLE void showDockIcon() const;
+    Q_INVOKABLE void hideDockIcon() const;
+
 Q_SIGNALS:
     void isNetworkOnlineChanged();
 

gpt4all-chat/src/localdocsmodel.cpp

@@ -20,7 +20,6 @@ LocalDocsCollectionsModel::LocalDocsCollectionsModel(QObject *parent)
     connect(this, &LocalDocsCollectionsModel::rowsInserted, this, &LocalDocsCollectionsModel::countChanged);
     connect(this, &LocalDocsCollectionsModel::rowsRemoved, this, &LocalDocsCollectionsModel::countChanged);
     connect(this, &LocalDocsCollectionsModel::modelReset, this, &LocalDocsCollectionsModel::countChanged);
-    connect(this, &LocalDocsCollectionsModel::layoutChanged, this, &LocalDocsCollectionsModel::countChanged);
 }
 
 bool LocalDocsCollectionsModel::filterAcceptsRow(int sourceRow,
@@ -67,7 +66,6 @@ LocalDocsModel::LocalDocsModel(QObject *parent)
     connect(this, &LocalDocsModel::rowsInserted, this, &LocalDocsModel::countChanged);
     connect(this, &LocalDocsModel::rowsRemoved, this, &LocalDocsModel::countChanged);
     connect(this, &LocalDocsModel::modelReset, this, &LocalDocsModel::countChanged);
-    connect(this, &LocalDocsModel::layoutChanged, this, &LocalDocsModel::countChanged);
 }
 
 int LocalDocsModel::rowCount(const QModelIndex &parent) const

gpt4all-chat/src/macosdock.h

@@ -0,0 +1,9 @@
+#ifndef MACOSDOCK_H
+#define MACOSDOCK_H
+
+struct MacOSDock {
+static void showIcon();
+static void hideIcon();
+};
+
+#endif // MACOSDOCK_H

gpt4all-chat/src/macosdock.mm

@@ -0,0 +1,13 @@
+#include "macosdock.h"
+
+#include <Cocoa/Cocoa.h>
+
+void MacOSDock::showIcon()
+{
+    [[NSApplication sharedApplication] setActivationPolicy:NSApplicationActivationPolicyRegular];
+}
+
+void MacOSDock::hideIcon()
+{
+    [[NSApplication sharedApplication] setActivationPolicy:NSApplicationActivationPolicyProhibited];
+}

gpt4all-chat/src/main.cpp

@@ -12,12 +12,16 @@
 #include <singleapplication.h>
 
 #include <QCoreApplication>
+#include <QFont>
+#include <QFontDatabase>
 #include <QObject>
 #include <QQmlApplicationEngine>
+#include <QQmlContext>
 #include <QQuickWindow>
 #include <QSettings>
 #include <QString>
 #include <QUrl>
+#include <QVariant>
 #include <Qt>
 
 #ifdef Q_OS_LINUX
@@ -26,6 +30,8 @@
 
 #ifdef Q_OS_WINDOWS
 #   include <windows.h>
+#else
+#   include <signal.h>
 #endif
 
 using namespace Qt::Literals::StringLiterals;
@@ -42,6 +48,7 @@ static void raiseWindow(QWindow *window)
 
     SetForegroundWindow(hwnd);
 #else
+    LLM::globalInstance()->showDockIcon();
     window->show();
     window->raise();
     window->requestActivate();
@@ -88,18 +95,22 @@ int main(int argc, char *argv[])
 
     // Set the local and language translation before the qml engine has even been started. This will
     // use the default system locale unless the user has explicitly set it to use a different one.
-    MySettings::globalInstance()->setLanguageAndLocale();
+    auto *mySettings = MySettings::globalInstance();
+    mySettings->setLanguageAndLocale();
 
     QQmlApplicationEngine engine;
 
     // Add a connection here from MySettings::languageAndLocaleChanged signal to a lambda slot where I can call
     // engine.uiLanguage property
-    QObject::connect(MySettings::globalInstance(), &MySettings::languageAndLocaleChanged, [&engine]() {
+    QObject::connect(mySettings, &MySettings::languageAndLocaleChanged, [&engine]() {
         engine.setUiLanguage(MySettings::globalInstance()->languageAndLocale());
     });
 
-    qmlRegisterSingletonInstance("mysettings", 1, 0, "MySettings", MySettings::globalInstance());
-    qmlRegisterSingletonInstance("modellist", 1, 0, "ModelList", ModelList::globalInstance());
+    auto *modelList = ModelList::globalInstance();
+    QObject::connect(modelList, &ModelList::dataChanged, mySettings, &MySettings::onModelInfoChanged);
+
+    qmlRegisterSingletonInstance("mysettings", 1, 0, "MySettings", mySettings);
+    qmlRegisterSingletonInstance("modellist", 1, 0, "ModelList", modelList);
     qmlRegisterSingletonInstance("chatlistmodel", 1, 0, "ChatListModel", ChatListModel::globalInstance());
     qmlRegisterSingletonInstance("llm", 1, 0, "LLM", LLM::globalInstance());
     qmlRegisterSingletonInstance("download", 1, 0, "Download", Download::globalInstance());
@@ -107,6 +118,11 @@ int main(int argc, char *argv[])
     qmlRegisterSingletonInstance("localdocs", 1, 0, "LocalDocs", LocalDocs::globalInstance());
     qmlRegisterUncreatableMetaObject(MySettingsEnums::staticMetaObject, "mysettingsenums", 1, 0, "MySettingsEnums", "Error: only enums");
 
+    {
+        auto fixedFont = QFontDatabase::systemFont(QFontDatabase::FixedFont);
+        engine.rootContext()->setContextProperty("fixedFont", fixedFont);
+    }
+
     const QUrl url(u"qrc:/gpt4all/main.qml"_s);
 
     QObject::connect(&engine, &QQmlApplicationEngine::objectCreated,
@@ -130,6 +146,17 @@ int main(int argc, char *argv[])
     }
 #endif
 
+#ifndef Q_OS_WINDOWS
+    // handle signals gracefully
+    struct sigaction sa;
+    sa.sa_handler = [](int s) { QCoreApplication::exit(s == SIGINT ? 0 : 1); };
+    sa.sa_flags   = SA_RESETHAND;
+    sigemptyset(&sa.sa_mask);
+    sigaction(SIGINT,  &sa, nullptr);
+    sigaction(SIGTERM, &sa, nullptr);
+    sigaction(SIGHUP,  &sa, nullptr);
+#endif
+
     int res = app.exec();
 
     // Make sure ChatLLM threads are joined before global destructors run.

gpt4all-chat/src/modellist.cpp

@@ -27,6 +27,7 @@
 #include <QSettings>
 #include <QSslConfiguration>
 #include <QSslSocket>
+#include <QStandardPaths>
 #include <QStringList>
 #include <QTextStream>
 #include <QTimer>
@@ -36,6 +37,7 @@
 #include <algorithm>
 #include <cstddef>
 #include <iterator>
+#include <optional>
 #include <string>
 #include <utility>
 
@@ -43,6 +45,8 @@ using namespace Qt::Literals::StringLiterals;
 
 //#define USE_LOCAL_MODELSJSON
 
+#define MODELS_JSON_VERSION "3"
+
 static const QStringList FILENAME_BLACKLIST { u"gpt4all-nomic-embed-text-v1.rmodel"_s };
 
 QString ModelInfo::id() const
@@ -312,26 +316,44 @@ void ModelInfo::setRepeatPenaltyTokens(int t)
     m_repeatPenaltyTokens = t;
 }
 
-QString ModelInfo::promptTemplate() const
+QVariant ModelInfo::defaultChatTemplate() const
 {
-    return MySettings::globalInstance()->modelPromptTemplate(*this);
+    auto res = m_chatTemplate.or_else([this]() -> std::optional<QString> {
+        if (!installed || isOnline)
+            return std::nullopt;
+        if (!m_modelChatTemplate) {
+            auto path = (dirpath + filename()).toUtf8();
+            auto res = LLModel::Implementation::chatTemplate(path.constData());
+            if (res) {
+                m_modelChatTemplate = QString::fromStdString(*res);
+            } else {
+                qWarning().nospace() << "failed to get chat template for " << filename() << ": " << res.error().c_str();
+                m_modelChatTemplate = QString(); // do not retry
+            }
+        }
+        if (m_modelChatTemplate->isNull())
+            return std::nullopt;
+        return m_modelChatTemplate;
+    });
+
+    if (res)
+        return std::move(*res);
+    return QVariant::fromValue(nullptr);
 }
 
-void ModelInfo::setPromptTemplate(const QString &t)
+auto ModelInfo::chatTemplate() const -> UpgradeableSetting
 {
-    if (shouldSaveMetadata()) MySettings::globalInstance()->setModelPromptTemplate(*this, t, true /*force*/);
-    m_promptTemplate = t;
+    return MySettings::globalInstance()->modelChatTemplate(*this);
 }
 
-QString ModelInfo::systemPrompt() const
+QString ModelInfo::defaultSystemMessage() const
 {
-    return MySettings::globalInstance()->modelSystemPrompt(*this);
+    return m_systemMessage;
 }
 
-void ModelInfo::setSystemPrompt(const QString &p)
+auto ModelInfo::systemMessage() const -> UpgradeableSetting
 {
-    if (shouldSaveMetadata()) MySettings::globalInstance()->setModelSystemPrompt(*this, p, true /*force*/);
-    m_systemPrompt = p;
+    return MySettings::globalInstance()->modelSystemMessage(*this);
 }
 
 QString ModelInfo::chatNamePrompt() const
@@ -356,39 +378,41 @@ void ModelInfo::setSuggestedFollowUpPrompt(const QString &p)
     m_suggestedFollowUpPrompt = p;
 }
 
+// FIXME(jared): this should not be used for model settings that have meaningful defaults, such as temperature
 bool ModelInfo::shouldSaveMetadata() const
 {
     return installed && (isClone() || isDiscovered() || description() == "" /*indicates sideloaded*/);
 }
 
-QVariantMap ModelInfo::getFields() const
+QVariant ModelInfo::getField(QLatin1StringView name) const
 {
-    return {
-        { "filename",            m_filename },
-        { "description",         m_description },
-        { "url",                 m_url },
-        { "quant",               m_quant },
-        { "type",                m_type },
-        { "isClone",             m_isClone },
-        { "isDiscovered",        m_isDiscovered },
-        { "likes",               m_likes },
-        { "downloads",           m_downloads },
-        { "recency",             m_recency },
-        { "temperature",         m_temperature },
-        { "topP",                m_topP },
-        { "minP",                m_minP },
-        { "topK",                m_topK },
-        { "maxLength",           m_maxLength },
-        { "promptBatchSize",     m_promptBatchSize },
-        { "contextLength",       m_contextLength },
-        { "gpuLayers",           m_gpuLayers },
-        { "repeatPenalty",       m_repeatPenalty },
-        { "repeatPenaltyTokens", m_repeatPenaltyTokens },
-        { "promptTemplate",      m_promptTemplate },
-        { "systemPrompt",        m_systemPrompt },
-        { "chatNamePrompt",      m_chatNamePrompt },
-        { "suggestedFollowUpPrompt", m_suggestedFollowUpPrompt },
+    static const std::unordered_map<QLatin1StringView, QVariant(*)(const ModelInfo &)> s_fields = {
+        { "filename"_L1,                [](auto &i) -> QVariant { return i.m_filename;                } },
+        { "description"_L1,             [](auto &i) -> QVariant { return i.m_description;             } },
+        { "url"_L1,                     [](auto &i) -> QVariant { return i.m_url;                     } },
+        { "quant"_L1,                   [](auto &i) -> QVariant { return i.m_quant;                   } },
+        { "type"_L1,                    [](auto &i) -> QVariant { return i.m_type;                    } },
+        { "isClone"_L1,                 [](auto &i) -> QVariant { return i.m_isClone;                 } },
+        { "isDiscovered"_L1,            [](auto &i) -> QVariant { return i.m_isDiscovered;            } },
+        { "likes"_L1,                   [](auto &i) -> QVariant { return i.m_likes;                   } },
+        { "downloads"_L1,               [](auto &i) -> QVariant { return i.m_downloads;               } },
+        { "recency"_L1,                 [](auto &i) -> QVariant { return i.m_recency;                 } },
+        { "temperature"_L1,             [](auto &i) -> QVariant { return i.m_temperature;             } },
+        { "topP"_L1,                    [](auto &i) -> QVariant { return i.m_topP;                    } },
+        { "minP"_L1,                    [](auto &i) -> QVariant { return i.m_minP;                    } },
+        { "topK"_L1,                    [](auto &i) -> QVariant { return i.m_topK;                    } },
+        { "maxLength"_L1,               [](auto &i) -> QVariant { return i.m_maxLength;               } },
+        { "promptBatchSize"_L1,         [](auto &i) -> QVariant { return i.m_promptBatchSize;         } },
+        { "contextLength"_L1,           [](auto &i) -> QVariant { return i.m_contextLength;           } },
+        { "gpuLayers"_L1,               [](auto &i) -> QVariant { return i.m_gpuLayers;               } },
+        { "repeatPenalty"_L1,           [](auto &i) -> QVariant { return i.m_repeatPenalty;           } },
+        { "repeatPenaltyTokens"_L1,     [](auto &i) -> QVariant { return i.m_repeatPenaltyTokens;     } },
+        { "chatTemplate"_L1,            [](auto &i) -> QVariant { return i.defaultChatTemplate();     } },
+        { "systemMessage"_L1,           [](auto &i) -> QVariant { return i.m_systemMessage;           } },
+        { "chatNamePrompt"_L1,          [](auto &i) -> QVariant { return i.m_chatNamePrompt;          } },
+        { "suggestedFollowUpPrompt"_L1, [](auto &i) -> QVariant { return i.m_suggestedFollowUpPrompt; } },
     };
+    return s_fields.at(name)(*this);
 }
 
 InstalledModels::InstalledModels(QObject *parent, bool selectable)
@@ -398,7 +422,6 @@ InstalledModels::InstalledModels(QObject *parent, bool selectable)
     connect(this, &InstalledModels::rowsInserted, this, &InstalledModels::countChanged);
     connect(this, &InstalledModels::rowsRemoved, this, &InstalledModels::countChanged);
     connect(this, &InstalledModels::modelReset, this, &InstalledModels::countChanged);
-    connect(this, &InstalledModels::layoutChanged, this, &InstalledModels::countChanged);
 }
 
 bool InstalledModels::filterAcceptsRow(int sourceRow,
@@ -423,7 +446,6 @@ DownloadableModels::DownloadableModels(QObject *parent)
     connect(this, &DownloadableModels::rowsInserted, this, &DownloadableModels::countChanged);
     connect(this, &DownloadableModels::rowsRemoved, this, &DownloadableModels::countChanged);
     connect(this, &DownloadableModels::modelReset, this, &DownloadableModels::countChanged);
-    connect(this, &DownloadableModels::layoutChanged, this, &DownloadableModels::countChanged);
 }
 
 bool DownloadableModels::filterAcceptsRow(int sourceRow,
@@ -489,41 +511,58 @@ ModelList::ModelList()
     m_selectableModels->setSourceModel(this);
     m_downloadableModels->setSourceModel(this);
 
-    connect(MySettings::globalInstance(), &MySettings::modelPathChanged, this, &ModelList::updateModelsFromDirectory);
-    connect(MySettings::globalInstance(), &MySettings::modelPathChanged, this, &ModelList::updateModelsFromJson);
-    connect(MySettings::globalInstance(), &MySettings::modelPathChanged, this, &ModelList::updateModelsFromSettings);
-    connect(MySettings::globalInstance(), &MySettings::nameChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::temperatureChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::topPChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::minPChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::topKChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::maxLengthChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::promptBatchSizeChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::contextLengthChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::gpuLayersChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::repeatPenaltyChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::repeatPenaltyTokensChanged, this, &ModelList::updateDataForSettings);;
-    connect(MySettings::globalInstance(), &MySettings::promptTemplateChanged, this, &ModelList::updateDataForSettings);
-    connect(MySettings::globalInstance(), &MySettings::systemPromptChanged, this, &ModelList::updateDataForSettings);
+    auto *mySettings = MySettings::globalInstance();
+    connect(mySettings, &MySettings::nameChanged,                this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::temperatureChanged,         this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::topPChanged,                this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::minPChanged,                this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::topKChanged,                this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::maxLengthChanged,           this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::promptBatchSizeChanged,     this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::contextLengthChanged,       this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::gpuLayersChanged,           this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::repeatPenaltyChanged,       this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::repeatPenaltyTokensChanged, this, &ModelList::updateDataForSettings     );
+    connect(mySettings, &MySettings::chatTemplateChanged,        this, &ModelList::maybeUpdateDataForSettings);
+    connect(mySettings, &MySettings::systemMessageChanged,       this, &ModelList::maybeUpdateDataForSettings);
+
+    connect(this, &ModelList::dataChanged, this, &ModelList::onDataChanged);
+
     connect(&m_networkManager, &QNetworkAccessManager::sslErrors, this, &ModelList::handleSslErrors);
 
     updateModelsFromJson();
     updateModelsFromSettings();
     updateModelsFromDirectory();
 
+    connect(mySettings, &MySettings::modelPathChanged, this, &ModelList::updateModelsFromDirectory);
+    connect(mySettings, &MySettings::modelPathChanged, this, &ModelList::updateModelsFromJson     );
+    connect(mySettings, &MySettings::modelPathChanged, this, &ModelList::updateModelsFromSettings );
+
     QCoreApplication::instance()->installEventFilter(this);
 }
 
+// an easier way to listen for model info and setting changes
+void ModelList::onDataChanged(const QModelIndex &topLeft, const QModelIndex &bottomRight, const QList<int> &roles)
+{
+    Q_UNUSED(roles)
+    for (int row = topLeft.row(); row <= bottomRight.row(); row++) {
+        auto index = topLeft.siblingAtRow(row);
+        auto id = index.data(ModelList::IdRole).toString();
+        if (auto info = modelInfo(id); !info.id().isNull())
+            emit modelInfoChanged(info);
+    }
+}
+
 QString ModelList::compatibleModelNameHash(QUrl baseUrl, QString modelName) {
     QCryptographicHash sha256(QCryptographicHash::Sha256);
     sha256.addData((baseUrl.toString() + "_" + modelName).toUtf8());
     return sha256.result().toHex();
-};
+}
 
 QString ModelList::compatibleModelFilename(QUrl baseUrl, QString modelName) {
     QString hash(compatibleModelNameHash(baseUrl, modelName));
     return QString(u"gpt4all-%1-capi.rmodel"_s).arg(hash);
-};
+}
 
 bool ModelList::eventFilter(QObject *obj, QEvent *ev)
 {
@@ -774,10 +813,10 @@ QVariant ModelList::dataInternal(const ModelInfo *info, int role) const
             return info->repeatPenalty();
         case RepeatPenaltyTokensRole:
             return info->repeatPenaltyTokens();
-        case PromptTemplateRole:
-            return info->promptTemplate();
-        case SystemPromptRole:
-            return info->systemPrompt();
+        case ChatTemplateRole:
+            return QVariant::fromValue(info->chatTemplate());
+        case SystemMessageRole:
+            return QVariant::fromValue(info->systemMessage());
         case ChatNamePromptRole:
             return info->chatNamePrompt();
         case SuggestedFollowUpPromptRole:
@@ -821,7 +860,11 @@ QVariant ModelList::data(const QModelIndex &index, int role) const
 
 void ModelList::updateData(const QString &id, const QVector<QPair<int, QVariant>> &data)
 {
+    // We only sort when one of the fields used by the sorting algorithm actually changes that
+    // is implicated or used by the sorting algorithm
+    bool shouldSort = false;
     int index;
+
     {
         QMutexLocker locker(&m_mutex);
         if (!m_modelMap.contains(id)) {
@@ -836,10 +879,6 @@ void ModelList::updateData(const QString &id, const QVector<QPair<int, QVariant>
             return;
         }
 
-        // We only sort when one of the fields used by the sorting algorithm actually changes that
-        // is implicated or used by the sorting algorithm
-        bool shouldSort = false;
-
         for (const auto &d : data) {
             const int role = d.first;
             const QVariant value = d.second;
@@ -950,10 +989,10 @@ void ModelList::updateData(const QString &id, const QVector<QPair<int, QVariant>
                 info->setRepeatPenalty(value.toDouble()); break;
             case RepeatPenaltyTokensRole:
                 info->setRepeatPenaltyTokens(value.toInt()); break;
-            case PromptTemplateRole:
-                info->setPromptTemplate(value.toString()); break;
-            case SystemPromptRole:
-                info->setSystemPrompt(value.toString()); break;
+            case ChatTemplateRole:
+                info->m_chatTemplate = value.toString(); break;
+            case SystemMessageRole:
+                info->m_systemMessage = value.toString(); break;
             case ChatNamePromptRole:
                 info->setChatNamePrompt(value.toString()); break;
             case SuggestedFollowUpPromptRole:
@@ -1000,21 +1039,12 @@ void ModelList::updateData(const QString &id, const QVector<QPair<int, QVariant>
             info->isEmbeddingModel = LLModel::Implementation::isEmbeddingModel(modelPath.toStdString());
             info->checkedEmbeddingModel = true;
         }
-
-        if (shouldSort) {
-            auto s = m_discoverSort;
-            auto d = m_discoverSortDirection;
-            std::stable_sort(m_models.begin(), m_models.end(), [s, d](const ModelInfo* lhs, const ModelInfo* rhs) {
-                return ModelList::lessThan(lhs, rhs, s, d);
-            });
-        }
     }
+
     emit dataChanged(createIndex(index, 0), createIndex(index, 0));
 
-    // FIXME(jared): for some reason these don't update correctly when the source model changes, so we explicitly invalidate them
-    m_selectableModels->invalidate();
-    m_installedModels->invalidate();
-    m_downloadableModels->invalidate();
+    if (shouldSort)
+        resortModel();
 
     emit selectableModelListChanged();
 }
@@ -1063,11 +1093,11 @@ ModelInfo ModelList::modelInfo(const QString &id) const
     return *m_modelMap.value(id);
 }
 
-ModelInfo ModelList::modelInfoByFilename(const QString &filename) const
+ModelInfo ModelList::modelInfoByFilename(const QString &filename, bool allowClone) const
 {
     QMutexLocker locker(&m_mutex);
     for (ModelInfo *info : m_models)
-        if (info->filename() == filename)
+        if (info->filename() == filename && (allowClone || !info->isClone()))
             return *info;
     return ModelInfo();
 }
@@ -1087,6 +1117,20 @@ QString ModelList::clone(const ModelInfo &model)
     const QString id = Network::globalInstance()->generateUniqueId();
     addModel(id);
 
+    QString chatTemplate, systemMessage;
+    if (auto tmpl = model.chatTemplate().asModern()) {
+        chatTemplate = *tmpl;
+    } else {
+        qWarning("ModelList Warning: attempted to clone model with legacy chat template");
+        return {};
+    }
+    if (auto msg = model.systemMessage().asModern()) {
+        systemMessage = *msg;
+    } else {
+        qWarning("ModelList Warning: attempted to clone model with legacy system message");
+        return {};
+    }
+
     QVector<QPair<int, QVariant>> data {
         { ModelList::InstalledRole, model.installed },
         { ModelList::IsCloneRole, true },
@@ -1106,8 +1150,8 @@ QString ModelList::clone(const ModelInfo &model)
         { ModelList::GpuLayersRole, model.gpuLayers() },
         { ModelList::RepeatPenaltyRole, model.repeatPenalty() },
         { ModelList::RepeatPenaltyTokensRole, model.repeatPenaltyTokens() },
-        { ModelList::PromptTemplateRole, model.promptTemplate() },
-        { ModelList::SystemPromptRole, model.systemPrompt() },
+        { ModelList::ChatTemplateRole,  chatTemplate  },
+        { ModelList::SystemMessageRole, systemMessage },
         { ModelList::ChatNamePromptRole, model.chatNamePrompt() },
         { ModelList::SuggestedFollowUpPromptRole, model.suggestedFollowUpPrompt() },
     };
@@ -1122,7 +1166,6 @@ void ModelList::removeClone(const ModelInfo &model)
         return;
 
     removeInternal(model);
-    emit layoutChanged();
 }
 
 void ModelList::removeInstalled(const ModelInfo &model)
@@ -1131,24 +1174,25 @@ void ModelList::removeInstalled(const ModelInfo &model)
     Q_ASSERT(!model.isClone());
     Q_ASSERT(model.isDiscovered() || model.isCompatibleApi || model.description() == "" /*indicates sideloaded*/);
     removeInternal(model);
-    emit layoutChanged();
+}
+
+int ModelList::indexByModelId(const QString &id) const
+{
+    QMutexLocker locker(&m_mutex);
+    if (auto it = m_modelMap.find(id); it != m_modelMap.cend())
+        return m_models.indexOf(*it);
+    return -1;
 }
 
 void ModelList::removeInternal(const ModelInfo &model)
 {
-    const bool hasModel = contains(model.id());
-    Q_ASSERT(hasModel);
-    if (!hasModel) {
+    int indexOfModel = indexByModelId(model.id());
+    Q_ASSERT(indexOfModel != -1);
+    if (indexOfModel == -1) {
         qWarning() << "ERROR: model list does not contain" << model.id();
         return;
     }
 
-    int indexOfModel = 0;
-    {
-        QMutexLocker locker(&m_mutex);
-        ModelInfo *info = m_modelMap.value(model.id());
-        indexOfModel = m_models.indexOf(info);
-    }
     beginRemoveRows(QModelIndex(), indexOfModel, indexOfModel);
     {
         QMutexLocker locker(&m_mutex);
@@ -1323,8 +1367,6 @@ void ModelList::processModelDirectory(const QString &path)
                 // The description is hard-coded into "GPT4All.ini" due to performance issue.
                 // If the description goes to be dynamic from its .rmodel file, it will get high I/O usage while using the ModelList.
                 data.append({ DescriptionRole, description });
-                // Prompt template should be clear while using ChatML format which is using in most of OpenAI-Compatible API server.
-                data.append({ PromptTemplateRole, "%1" });
             }
             updateData(id, data);
         }
@@ -1344,15 +1386,32 @@ void ModelList::updateModelsFromDirectory()
     }
 }
 
-#define MODELS_VERSION 3
+static QString modelsJsonFilename()
+{
+    return QStringLiteral("models" MODELS_JSON_VERSION ".json");
+}
+
+static std::optional<QFile> modelsJsonCacheFile()
+{
+    constexpr auto loc = QStandardPaths::CacheLocation;
+    QString modelsJsonFname = modelsJsonFilename();
+    if (auto path = QStandardPaths::locate(loc, modelsJsonFname); !path.isEmpty())
+        return std::make_optional<QFile>(path);
+    if (auto path = QStandardPaths::writableLocation(loc); !path.isEmpty())
+        return std::make_optional<QFile>(u"%1/%2"_s.arg(path, modelsJsonFname));
+    return std::nullopt;
+}
 
 void ModelList::updateModelsFromJson()
 {
+    QString modelsJsonFname = modelsJsonFilename();
+
 #if defined(USE_LOCAL_MODELSJSON)
-    QUrl jsonUrl("file://" + QDir::homePath() + u"/dev/large_language_models/gpt4all/gpt4all-chat/metadata/models%1.json"_s.arg(MODELS_VERSION));
+    QUrl jsonUrl(u"file://%1/dev/large_language_models/gpt4all/gpt4all-chat/metadata/%2"_s.arg(QDir::homePath(), modelsJsonFname));
 #else
-    QUrl jsonUrl(u"http://gpt4all.io/models/models%1.json"_s.arg(MODELS_VERSION));
+    QUrl jsonUrl(u"http://gpt4all.io/models/%1"_s.arg(modelsJsonFname));
 #endif
+
     QNetworkRequest request(jsonUrl);
     QSslConfiguration conf = request.sslConfiguration();
     conf.setPeerVerifyMode(QSslSocket::VerifyNone);
@@ -1371,18 +1430,15 @@ void ModelList::updateModelsFromJson()
         qWarning() << "WARNING: Could not download models.json synchronously";
         updateModelsFromJsonAsync();
 
-        QSettings settings;
-        QFileInfo info(settings.fileName());
-        QString dirPath = info.canonicalPath();
-        const QString modelsConfig = dirPath + "/models.json";
-        QFile file(modelsConfig);
-        if (!file.open(QIODeviceBase::ReadOnly)) {
-            qWarning() << "ERROR: Couldn't read models config file: " << modelsConfig;
-        } else {
-            QByteArray jsonData = file.readAll();
-            file.close();
+        auto cacheFile = modelsJsonCacheFile();
+        if (!cacheFile) {
+            // no known location
+        } else if (cacheFile->open(QIODeviceBase::ReadOnly)) {
+            QByteArray jsonData = cacheFile->readAll();
+            cacheFile->close();
             parseModelsJsonFile(jsonData, false);
-        }
+        } else if (cacheFile->exists())
+            qWarning() << "ERROR: Couldn't read models.json cache file: " << cacheFile->fileName();
     }
     delete jsonReply;
 }
@@ -1391,12 +1447,14 @@ void ModelList::updateModelsFromJsonAsync()
 {
     m_asyncModelRequestOngoing = true;
     emit asyncModelRequestOngoingChanged();
+    QString modelsJsonFname = modelsJsonFilename();
 
 #if defined(USE_LOCAL_MODELSJSON)
-    QUrl jsonUrl("file://" + QDir::homePath() + u"/dev/large_language_models/gpt4all/gpt4all-chat/metadata/models%1.json"_s.arg(MODELS_VERSION));
+    QUrl jsonUrl(u"file://%1/dev/large_language_models/gpt4all/gpt4all-chat/metadata/%2"_s.arg(QDir::homePath(), modelsJsonFname));
 #else
-    QUrl jsonUrl(u"http://gpt4all.io/models/models%1.json"_s.arg(MODELS_VERSION));
+    QUrl jsonUrl(u"http://gpt4all.io/models/%1"_s.arg(modelsJsonFname));
 #endif
+
     QNetworkRequest request(jsonUrl);
     QSslConfiguration conf = request.sslConfiguration();
     conf.setPeerVerifyMode(QSslSocket::VerifyNone);
@@ -1444,9 +1502,20 @@ void ModelList::handleSslErrors(QNetworkReply *reply, const QList<QSslError> &er
         qWarning() << "ERROR: Received ssl error:" << e.errorString() << "for" << url;
 }
 
+void ModelList::maybeUpdateDataForSettings(const ModelInfo &info, bool fromInfo)
+{
+    // ignore updates that were *because* of a dataChanged - would cause a circular dependency
+    int idx;
+    if (!fromInfo && (idx = indexByModelId(info.id())) != -1) {
+        emit dataChanged(index(idx, 0), index(idx, 0));
+        emit selectableModelListChanged();
+    }
+}
+
 void ModelList::updateDataForSettings()
 {
     emit dataChanged(index(0, 0), index(m_models.size() - 1, 0));
+    emit selectableModelListChanged();
 }
 
 void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
@@ -1459,17 +1528,14 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
     }
 
     if (save) {
-        QSettings settings;
-        QFileInfo info(settings.fileName());
-        QString dirPath = info.canonicalPath();
-        const QString modelsConfig = dirPath + "/models.json";
-        QFile file(modelsConfig);
-        if (!file.open(QIODeviceBase::WriteOnly)) {
-            qWarning() << "ERROR: Couldn't write models config file: " << modelsConfig;
-        } else {
-            file.write(jsonData);
-            file.close();
-        }
+        auto cacheFile = modelsJsonCacheFile();
+        if (!cacheFile) {
+            // no known location
+        } else if (QFileInfo(*cacheFile).dir().mkpath(u"."_s) && cacheFile->open(QIODeviceBase::WriteOnly)) {
+            cacheFile->write(jsonData);
+            cacheFile->close();
+        } else
+            qWarning() << "ERROR: Couldn't write models config file: " << cacheFile->fileName();
     }
 
     QJsonArray jsonArray = document.array();
@@ -1556,10 +1622,10 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
             data.append({ ModelList::RepeatPenaltyRole, obj["repeatPenalty"].toDouble() });
         if (obj.contains("repeatPenaltyTokens"))
             data.append({ ModelList::RepeatPenaltyTokensRole, obj["repeatPenaltyTokens"].toInt() });
-        if (obj.contains("promptTemplate"))
-            data.append({ ModelList::PromptTemplateRole, obj["promptTemplate"].toString() });
-        if (obj.contains("systemPrompt"))
-            data.append({ ModelList::SystemPromptRole, obj["systemPrompt"].toString() });
+        if (auto it = obj.find("chatTemplate"_L1); it != obj.end())
+            data.append({ ModelList::ChatTemplateRole, it->toString() });
+        if (auto it = obj.find("systemMessage"_L1); it != obj.end())
+            data.append({ ModelList::SystemMessageRole, it->toString() });
         updateData(id, data);
     }
 
@@ -1751,6 +1817,9 @@ void ModelList::updateDiscoveredInstalled(const ModelInfo &info)
     updateData(info.id(), data);
 }
 
+// FIXME(jared): This should only contain fields without reasonable defaults such as name, description, and URL.
+//               For other settings, there is no authoritative value and we should load the setting lazily like we do
+//               for any other override.
 void ModelList::updateModelsFromSettings()
 {
     QSettings settings;
@@ -1765,12 +1834,27 @@ void ModelList::updateModelsFromSettings()
 
         // If we can't find the corresponding file, then ignore it as this reflects a stale model.
         // The file could have been deleted manually by the user for instance or temporarily renamed.
-        if (!settings.contains(g + "/filename") || !modelExists(settings.value(g + "/filename").toString()))
-            continue;
+        QString filename;
+        {
+            auto value = settings.value(u"%1/filename"_s.arg(g));
+            if (!value.isValid() || !modelExists(filename = value.toString()))
+                continue;
+        }
+
+        QVector<QPair<int, QVariant>> data;
+
+        // load data from base model
+        // FIXME(jared): how does "Restore Defaults" work for other settings of clones which we don't do this for?
+        if (auto base = modelInfoByFilename(filename, /*allowClone*/ false); !base.id().isNull()) {
+            if (auto tmpl = base.m_chatTemplate)
+                data.append({ ModelList::ChatTemplateRole,  *tmpl });
+            if (auto msg  = base.m_systemMessage; !msg.isNull())
+                data.append({ ModelList::SystemMessageRole,  msg  });
+        }
 
         addModel(id);
 
-        QVector<QPair<int, QVariant>> data;
+        // load data from settings
         if (settings.contains(g + "/name")) {
             const QString name = settings.value(g + "/name").toString();
             data.append({ ModelList::NameRole, name });
@@ -1855,14 +1939,6 @@ void ModelList::updateModelsFromSettings()
             const int repeatPenaltyTokens = settings.value(g + "/repeatPenaltyTokens").toInt();
             data.append({ ModelList::RepeatPenaltyTokensRole, repeatPenaltyTokens });
         }
-        if (settings.contains(g + "/promptTemplate")) {
-            const QString promptTemplate = settings.value(g + "/promptTemplate").toString();
-            data.append({ ModelList::PromptTemplateRole, promptTemplate });
-        }
-        if (settings.contains(g + "/systemPrompt")) {
-            const QString systemPrompt = settings.value(g + "/systemPrompt").toString();
-            data.append({ ModelList::SystemPromptRole, systemPrompt });
-        }
         if (settings.contains(g + "/chatNamePrompt")) {
             const QString chatNamePrompt = settings.value(g + "/chatNamePrompt").toString();
             data.append({ ModelList::ChatNamePromptRole, chatNamePrompt });
@@ -1928,7 +2004,6 @@ void ModelList::clearDiscoveredModels()
     }
     for (ModelInfo &info : infos)
         removeInternal(info);
-    emit layoutChanged();
 }
 
 float ModelList::discoverProgress() const
@@ -2100,7 +2175,7 @@ void ModelList::parseDiscoveryJsonFile(const QByteArray &jsonData)
     emit discoverProgressChanged();
     if (!m_discoverNumberOfResults) {
         m_discoverInProgress = false;
-        emit discoverInProgressChanged();;
+        emit discoverInProgressChanged();
     }
 }
 
@@ -2176,9 +2251,8 @@ void ModelList::handleDiscoveryItemFinished()
     emit discoverProgressChanged();
 
     if (discoverProgress() >= 1.0) {
-        emit layoutChanged();
         m_discoverInProgress = false;
-        emit discoverInProgressChanged();;
+        emit discoverInProgressChanged();
     }
 
     reply->deleteLater();

gpt4all-chat/src/modellist.h

@@ -5,12 +5,14 @@
 #include <QByteArray>
 #include <QDateTime>
 #include <QHash>
+#include <QLatin1StringView>
 #include <QList>
 #include <QMutex>
 #include <QNetworkAccessManager>
 #include <QNetworkReply>
 #include <QObject>
 #include <QPair>
+#include <QQmlEngine>
 #include <QSortFilterProxyModel>
 #include <QSslError>
 #include <QString>
@@ -19,11 +21,53 @@
 #include <Qt>
 #include <QtGlobal>
 
+#include <optional>
 #include <utility>
 
 using namespace Qt::Literals::StringLiterals;
 
 
+class UpgradeableSetting {
+    Q_GADGET
+    QML_ANONYMOUS
+
+    // NOTE: Unset implies there is neither a value nor a default
+    enum class State { Unset, Legacy, Modern };
+
+    Q_PROPERTY(bool     isSet     READ isSet   )
+    Q_PROPERTY(bool     isLegacy  READ isLegacy)
+    Q_PROPERTY(bool     isModern  READ isModern)
+    Q_PROPERTY(QVariant value READ value) // string or null
+
+public:
+    struct legacy_tag_t { explicit legacy_tag_t() = default; };
+    static inline constexpr legacy_tag_t legacy_tag = legacy_tag_t();
+
+    UpgradeableSetting()                           : m_state(State::Unset ) {}
+    UpgradeableSetting(legacy_tag_t, QString value): m_state(State::Legacy), m_value(std::move(value)) {}
+    UpgradeableSetting(              QString value): m_state(State::Modern), m_value(std::move(value)) {}
+
+    bool     isSet   () const { return m_state != State::Unset;  }
+    bool     isLegacy() const { return m_state == State::Legacy; }
+    bool     isModern() const { return m_state == State::Modern; }
+    QVariant value   () const { return m_state == State::Unset ? QVariant::fromValue(nullptr) : m_value; }
+
+    friend bool operator==(const UpgradeableSetting &a, const UpgradeableSetting &b)
+    { return a.m_state == b.m_state && (a.m_state == State::Unset || a.m_value == b.m_value); }
+
+    // returns std::nullopt if there is a legacy template or it is not set
+    std::optional<QString> asModern() const
+    {
+        if (m_state == State::Modern)
+            return m_value;
+        return std::nullopt;
+    }
+
+private:
+    State   m_state;
+    QString m_value;
+};
+
 struct ModelInfo {
     Q_GADGET
     Q_PROPERTY(QString id READ id WRITE setId)
@@ -69,8 +113,11 @@ struct ModelInfo {
     Q_PROPERTY(int maxGpuLayers READ maxGpuLayers)
     Q_PROPERTY(double repeatPenalty READ repeatPenalty WRITE setRepeatPenalty)
     Q_PROPERTY(int repeatPenaltyTokens READ repeatPenaltyTokens WRITE setRepeatPenaltyTokens)
-    Q_PROPERTY(QString promptTemplate READ promptTemplate WRITE setPromptTemplate)
-    Q_PROPERTY(QString systemPrompt READ systemPrompt WRITE setSystemPrompt)
+    // user-defined chat template and system message must be written through settings because of their legacy compat
+    Q_PROPERTY(QVariant           defaultChatTemplate  READ defaultChatTemplate )
+    Q_PROPERTY(UpgradeableSetting chatTemplate         READ chatTemplate        )
+    Q_PROPERTY(QString            defaultSystemMessage READ defaultSystemMessage)
+    Q_PROPERTY(UpgradeableSetting systemMessage        READ systemMessage       )
     Q_PROPERTY(QString chatNamePrompt READ chatNamePrompt WRITE setChatNamePrompt)
     Q_PROPERTY(QString suggestedFollowUpPrompt READ suggestedFollowUpPrompt WRITE setSuggestedFollowUpPrompt)
     Q_PROPERTY(int likes READ likes WRITE setLikes)
@@ -178,19 +225,22 @@ public:
     void setRepeatPenalty(double p);
     int repeatPenaltyTokens() const;
     void setRepeatPenaltyTokens(int t);
-    QString promptTemplate() const;
-    void setPromptTemplate(const QString &t);
-    QString systemPrompt() const;
-    void setSystemPrompt(const QString &p);
+    QVariant defaultChatTemplate() const;
+    UpgradeableSetting chatTemplate() const;
+    QString defaultSystemMessage() const;
+    UpgradeableSetting systemMessage() const;
     QString chatNamePrompt() const;
     void setChatNamePrompt(const QString &p);
     QString suggestedFollowUpPrompt() const;
     void setSuggestedFollowUpPrompt(const QString &p);
 
+    // Some metadata must be saved to settings because it does not have a meaningful default from some other source.
+    // This is useful for fields such as name, description, and URL.
+    // It is true for any models that have not been installed from models.json.
     bool shouldSaveMetadata() const;
 
 private:
-    QVariantMap getFields() const;
+    QVariant getField(QLatin1StringView name) const;
 
     QString m_id;
     QString m_name;
@@ -216,11 +266,13 @@ private:
     mutable int m_maxGpuLayers        = -1;
     double  m_repeatPenalty           = 1.18;
     int     m_repeatPenaltyTokens     = 64;
-    QString m_promptTemplate          = "### Human:\n%1\n\n### Assistant:\n";
-    QString m_systemPrompt            = "### System:\nYou are an AI assistant who gives a quality response to whatever humans ask of you.\n\n";
+            std::optional<QString> m_chatTemplate;
+    mutable std::optional<QString> m_modelChatTemplate;
+    QString m_systemMessage;
     QString m_chatNamePrompt          = "Describe the above conversation in seven words or less.";
     QString m_suggestedFollowUpPrompt = "Suggest three very short factual follow-up questions that have not been answered yet or cannot be found inspired by the previous conversation and excerpts.";
     friend class MySettings;
+    friend class ModelList;
 };
 Q_DECLARE_METATYPE(ModelInfo)
 
@@ -340,8 +392,8 @@ public:
         GpuLayersRole,
         RepeatPenaltyRole,
         RepeatPenaltyTokensRole,
-        PromptTemplateRole,
-        SystemPromptRole,
+        ChatTemplateRole,
+        SystemMessageRole,
         ChatNamePromptRole,
         SuggestedFollowUpPromptRole,
         MinPRole,
@@ -394,8 +446,8 @@ public:
         roles[GpuLayersRole] = "gpuLayers";
         roles[RepeatPenaltyRole] = "repeatPenalty";
         roles[RepeatPenaltyTokensRole] = "repeatPenaltyTokens";
-        roles[PromptTemplateRole] = "promptTemplate";
-        roles[SystemPromptRole] = "systemPrompt";
+        roles[ChatTemplateRole] = "chatTemplate";
+        roles[SystemMessageRole] = "systemMessage";
         roles[ChatNamePromptRole] = "chatNamePrompt";
         roles[SuggestedFollowUpPromptRole] = "suggestedFollowUpPrompt";
         roles[LikesRole] = "likes";
@@ -416,7 +468,7 @@ public:
     bool contains(const QString &id) const;
     bool containsByFilename(const QString &filename) const;
     Q_INVOKABLE ModelInfo modelInfo(const QString &id) const;
-    Q_INVOKABLE ModelInfo modelInfoByFilename(const QString &filename) const;
+    Q_INVOKABLE ModelInfo modelInfoByFilename(const QString &filename, bool allowClone = true) const;
     Q_INVOKABLE bool isUniqueName(const QString &name) const;
     Q_INVOKABLE QString clone(const ModelInfo &model);
     Q_INVOKABLE void removeClone(const ModelInfo &model);
@@ -476,15 +528,18 @@ Q_SIGNALS:
     void discoverSortChanged();
     void discoverProgressChanged();
     void discoverInProgressChanged();
+    void modelInfoChanged(const ModelInfo &info);
 
 protected:
     bool eventFilter(QObject *obj, QEvent *ev) override;
 
 private Q_SLOTS:
+    void onDataChanged(const QModelIndex &topLeft, const QModelIndex &bottomRight, const QList<int> &roles);
     void resortModel();
     void updateModelsFromJson();
     void updateModelsFromJsonAsync();
     void updateModelsFromSettings();
+    void maybeUpdateDataForSettings(const ModelInfo &info, bool fromInfo);
     void updateDataForSettings();
     void handleModelsJsonDownloadFinished();
     void handleModelsJsonDownloadErrorOccurred(QNetworkReply::NetworkError code);
@@ -495,6 +550,9 @@ private Q_SLOTS:
     void handleSslErrors(QNetworkReply *reply, const QList<QSslError> &errors);
 
 private:
+    // Return the index of the model with the given id, or -1 if not found.
+    int indexByModelId(const QString &id) const;
+
     void removeInternal(const ModelInfo &model);
     void clearDiscoveredModels();
     bool modelExists(const QString &fileName) const;

gpt4all-chat/src/mysettings.cpp

@@ -1,5 +1,8 @@
 #include "mysettings.h"
 
+#include "chatllm.h"
+#include "modellist.h"
+
 #include <gpt4all-backend/llmodel.h>
 
 #include <QDebug>
@@ -29,8 +32,13 @@ static const QStringList suggestionModeNames { "LocalDocsOnly", "On", "Off" };
 static const QStringList chatThemeNames      { "Light", "Dark", "LegacyDark" };
 static const QStringList fontSizeNames       { "Small", "Medium", "Large" };
 
-// FIXME: All of these default strings that are shown in the UI for settings need to be marked as
-// translatable
+// psuedo-enum
+namespace ModelSettingsKey { namespace {
+    auto ChatTemplate   = "chatTemplate"_L1;
+    auto PromptTemplate = "promptTemplate"_L1; // legacy
+    auto SystemMessage  = "systemMessage"_L1;
+    auto SystemPrompt   = "systemPrompt"_L1;   // legacy
+} } // namespace ModelSettingsKey::(anonymous)
 
 namespace defaults {
 
@@ -48,7 +56,7 @@ static const QVariantMap basicDefaults {
     { "fontSize",                 QVariant::fromValue(FontSize::Small) },
     { "lastVersionStarted",       "" },
     { "networkPort",              4891, },
-    { "saveChatsContext",         false },
+    { "systemTray",               false },
     { "serverChat",               false },
     { "userDefaultModel",         "Application default" },
     { "suggestionMode",           QVariant::fromValue(SuggestionMode::LocalDocsOnly) },
@@ -146,6 +154,11 @@ static QStringList getUiLanguages(const QString &modelPath)
     return languageList;
 }
 
+static QString modelSettingName(const ModelInfo &info, auto &&name)
+{
+    return u"model-%1/%2"_s.arg(info.id(), name);
+}
+
 class MyPrivateSettings: public MySettings { };
 Q_GLOBAL_STATIC(MyPrivateSettings, settingsInstance)
 MySettings *MySettings::globalInstance()
@@ -161,6 +174,34 @@ MySettings::MySettings()
 {
 }
 
+QVariant MySettings::checkJinjaTemplateError(const QString &tmpl)
+{
+    if (auto err = ChatLLM::checkJinjaTemplateError(tmpl.toStdString()))
+        return QString::fromStdString(*err);
+    return QVariant::fromValue(nullptr);
+}
+
+// Unset settings come from ModelInfo. Listen for changes so we can emit our own setting-specific signals.
+void MySettings::onModelInfoChanged(const QModelIndex &topLeft, const QModelIndex &bottomRight, const QList<int> &roles)
+{
+    auto settingChanged = [&](const auto &info, auto role, const auto &name) {
+        return (roles.isEmpty() || roles.contains(role)) && !m_settings.contains(modelSettingName(info, name));
+    };
+
+    auto &modelList = dynamic_cast<const ModelList &>(*QObject::sender());
+    for (int row = topLeft.row(); row <= bottomRight.row(); row++) {
+        using enum ModelList::Roles;
+        using namespace ModelSettingsKey;
+        auto index = topLeft.siblingAtRow(row);
+        if (auto info = modelList.modelInfo(index.data(IdRole).toString()); !info.id().isNull()) {
+            if (settingChanged(info, ChatTemplateRole, ChatTemplate))
+                emit chatTemplateChanged(info, /*fromInfo*/ true);
+            if (settingChanged(info, SystemMessageRole, SystemMessage))
+                emit systemMessageChanged(info, /*fromInfo*/ true);
+        }
+    }
+}
+
 QVariant MySettings::getBasicSetting(const QString &name) const
 {
     return m_settings.value(name, basicDefaults.value(name));
@@ -186,15 +227,15 @@ void MySettings::restoreModelDefaults(const ModelInfo &info)
     setModelTemperature(info, info.m_temperature);
     setModelTopP(info, info.m_topP);
     setModelMinP(info, info.m_minP);
-    setModelTopK(info, info.m_topK);;
+    setModelTopK(info, info.m_topK);
     setModelMaxLength(info, info.m_maxLength);
     setModelPromptBatchSize(info, info.m_promptBatchSize);
     setModelContextLength(info, info.m_contextLength);
     setModelGpuLayers(info, info.m_gpuLayers);
     setModelRepeatPenalty(info, info.m_repeatPenalty);
     setModelRepeatPenaltyTokens(info, info.m_repeatPenaltyTokens);
-    setModelPromptTemplate(info, info.m_promptTemplate);
-    setModelSystemPrompt(info, info.m_systemPrompt);
+    resetModelChatTemplate (info);
+    resetModelSystemMessage(info);
     setModelChatNamePrompt(info, info.m_chatNamePrompt);
     setModelSuggestedFollowUpPrompt(info, info.m_suggestedFollowUpPrompt);
 }
@@ -205,7 +246,7 @@ void MySettings::restoreApplicationDefaults()
     setFontSize(basicDefaults.value("fontSize").value<FontSize>());
     setDevice(defaults::device);
     setThreadCount(defaults::threadCount);
-    setSaveChatsContext(basicDefaults.value("saveChatsContext").toBool());
+    setSystemTray(basicDefaults.value("systemTray").toBool());
     setServerChat(basicDefaults.value("serverChat").toBool());
     setNetworkPort(basicDefaults.value("networkPort").toInt());
     setModelPath(defaultLocalModelsPath());
@@ -250,29 +291,37 @@ void MySettings::setModelName(const ModelInfo &info, const QString &value, bool
         emit nameChanged(info);
 }
 
-static QString modelSettingName(const ModelInfo &info, const QString &name)
+QVariant MySettings::getModelSetting(QLatin1StringView name, const ModelInfo &info) const
 {
-    return u"model-%1/%2"_s.arg(info.id(), name);
+    QLatin1StringView nameL1(name);
+    return m_settings.value(modelSettingName(info, nameL1), info.getField(nameL1));
 }
 
-QVariant MySettings::getModelSetting(const QString &name, const ModelInfo &info) const
+QVariant MySettings::getModelSetting(const char *name, const ModelInfo &info) const
 {
-    return m_settings.value(modelSettingName(info, name), info.getFields().value(name));
+    return getModelSetting(QLatin1StringView(name), info);
 }
 
-void MySettings::setModelSetting(const QString &name, const ModelInfo &info, const QVariant &value, bool force,
+void MySettings::setModelSetting(QLatin1StringView name, const ModelInfo &info, const QVariant &value, bool force,
                                  bool signal)
 {
     if (!force && (info.id().isEmpty() || getModelSetting(name, info) == value))
         return;
 
-    QString settingName = modelSettingName(info, name);
-    if (info.getFields().value(name) == value && !info.shouldSaveMetadata())
+    QLatin1StringView nameL1(name);
+    QString settingName = modelSettingName(info, nameL1);
+    if (info.getField(nameL1) == value && !info.shouldSaveMetadata())
         m_settings.remove(settingName);
     else
         m_settings.setValue(settingName, value);
     if (signal && !force)
-        QMetaObject::invokeMethod(this, u"%1Changed"_s.arg(name).toLatin1().constData(), Q_ARG(ModelInfo, info));
+        QMetaObject::invokeMethod(this, u"%1Changed"_s.arg(nameL1).toLatin1().constData(), Q_ARG(ModelInfo, info));
+}
+
+void MySettings::setModelSetting(const char *name, const ModelInfo &info, const QVariant &value, bool force,
+                                 bool signal)
+{
+    setModelSetting(QLatin1StringView(name), info, value, force, signal);
 }
 
 QString   MySettings::modelFilename               (const ModelInfo &info) const { return getModelSetting("filename",                info).toString(); }
@@ -295,11 +344,68 @@ int       MySettings::modelContextLength          (const ModelInfo &info) const
 int       MySettings::modelGpuLayers              (const ModelInfo &info) const { return getModelSetting("gpuLayers",               info).toInt(); }
 double    MySettings::modelRepeatPenalty          (const ModelInfo &info) const { return getModelSetting("repeatPenalty",           info).toDouble(); }
 int       MySettings::modelRepeatPenaltyTokens    (const ModelInfo &info) const { return getModelSetting("repeatPenaltyTokens",     info).toInt(); }
-QString   MySettings::modelPromptTemplate         (const ModelInfo &info) const { return getModelSetting("promptTemplate",          info).toString(); }
-QString   MySettings::modelSystemPrompt           (const ModelInfo &info) const { return getModelSetting("systemPrompt",            info).toString(); }
 QString   MySettings::modelChatNamePrompt         (const ModelInfo &info) const { return getModelSetting("chatNamePrompt",          info).toString(); }
 QString   MySettings::modelSuggestedFollowUpPrompt(const ModelInfo &info) const { return getModelSetting("suggestedFollowUpPrompt", info).toString(); }
 
+auto MySettings::getUpgradeableModelSetting(
+    const ModelInfo &info, QLatin1StringView legacyKey, QLatin1StringView newKey
+) const -> UpgradeableSetting
+{
+    if (info.id().isEmpty()) {
+        qWarning("%s: got null model", Q_FUNC_INFO);
+        return {};
+    }
+
+    auto value = m_settings.value(modelSettingName(info, legacyKey));
+    if (value.isValid())
+        return { UpgradeableSetting::legacy_tag, value.toString() };
+
+    value = getModelSetting(newKey, info);
+    if (!value.isNull())
+        return value.toString();
+    return {}; // neither a default nor an override
+}
+
+bool MySettings::isUpgradeableModelSettingSet(
+    const ModelInfo &info, QLatin1StringView legacyKey, QLatin1StringView newKey
+) const
+{
+    if (info.id().isEmpty()) {
+        qWarning("%s: got null model", Q_FUNC_INFO);
+        return false;
+    }
+
+    if (m_settings.contains(modelSettingName(info, legacyKey)))
+        return true;
+
+    // NOTE: unlike getUpgradeableSetting(), this ignores the default
+    return m_settings.contains(modelSettingName(info, newKey));
+}
+
+auto MySettings::modelChatTemplate(const ModelInfo &info) const -> UpgradeableSetting
+{
+    using namespace ModelSettingsKey;
+    return getUpgradeableModelSetting(info, PromptTemplate, ChatTemplate);
+}
+
+bool MySettings::isModelChatTemplateSet(const ModelInfo &info) const
+{
+    using namespace ModelSettingsKey;
+    return isUpgradeableModelSettingSet(info, PromptTemplate, ChatTemplate);
+}
+
+auto MySettings::modelSystemMessage(const ModelInfo &info) const -> UpgradeableSetting
+{
+    using namespace ModelSettingsKey;
+    return getUpgradeableModelSetting(info, SystemPrompt, SystemMessage);
+}
+
+bool MySettings::isModelSystemMessageSet(const ModelInfo &info) const
+{
+    using namespace ModelSettingsKey;
+    return isUpgradeableModelSettingSet(info, SystemPrompt, SystemMessage);
+}
+
 void MySettings::setModelFilename(const ModelInfo &info, const QString &value, bool force)
 {
     setModelSetting("filename", info, value, force, true);
@@ -400,14 +506,77 @@ void MySettings::setModelRepeatPenaltyTokens(const ModelInfo &info, int value, b
     setModelSetting("repeatPenaltyTokens", info, value, force, true);
 }
 
-void MySettings::setModelPromptTemplate(const ModelInfo &info, const QString &value, bool force)
-{
-    setModelSetting("promptTemplate", info, value, force, true);
+bool MySettings::setUpgradeableModelSetting(
+    const ModelInfo &info, const QString &value, QLatin1StringView legacyKey, QLatin1StringView newKey
+) {
+    if (info.id().isEmpty()) {
+        qWarning("%s: got null model", Q_FUNC_INFO);
+        return false;
+    }
+
+    auto legacyModelKey = modelSettingName(info, legacyKey);
+    auto newModelKey    = modelSettingName(info, newKey   );
+    bool changed = false;
+    if (m_settings.contains(legacyModelKey)) {
+        m_settings.remove(legacyModelKey);
+        changed = true;
+    }
+    auto oldValue = m_settings.value(newModelKey);
+    if (!oldValue.isValid() || oldValue.toString() != value) {
+        m_settings.setValue(newModelKey, value);
+        changed = true;
+    }
+    return changed;
 }
 
-void MySettings::setModelSystemPrompt(const ModelInfo &info, const QString &value, bool force)
+bool MySettings::resetUpgradeableModelSetting(
+    const ModelInfo &info, QLatin1StringView legacyKey, QLatin1StringView newKey
+) {
+    if (info.id().isEmpty()) {
+        qWarning("%s: got null model", Q_FUNC_INFO);
+        return false;
+    }
+
+    auto legacyModelKey = modelSettingName(info, legacyKey);
+    auto newModelKey    = modelSettingName(info, newKey   );
+    bool changed        = false;
+    if (m_settings.contains(legacyModelKey)) {
+        m_settings.remove(legacyModelKey);
+        changed = true;
+    }
+    if (m_settings.contains(newModelKey)) {
+        m_settings.remove(newModelKey);
+        changed = true;
+    }
+    return changed;
+}
+
+void MySettings::setModelChatTemplate(const ModelInfo &info, const QString &value)
 {
-    setModelSetting("systemPrompt", info, value, force, true);
+    using namespace ModelSettingsKey;
+    if (setUpgradeableModelSetting(info, value, PromptTemplate, ChatTemplate))
+        emit chatTemplateChanged(info);
+}
+
+void MySettings::resetModelChatTemplate(const ModelInfo &info)
+{
+    using namespace ModelSettingsKey;
+    if (resetUpgradeableModelSetting(info, PromptTemplate, ChatTemplate))
+        emit chatTemplateChanged(info);
+}
+
+void MySettings::setModelSystemMessage(const ModelInfo &info, const QString &value)
+{
+    using namespace ModelSettingsKey;
+    if (setUpgradeableModelSetting(info, value, SystemPrompt, SystemMessage))
+        emit systemMessageChanged(info);
+}
+
+void MySettings::resetModelSystemMessage(const ModelInfo &info)
+{
+    using namespace ModelSettingsKey;
+    if (resetUpgradeableModelSetting(info, SystemPrompt, SystemMessage))
+        emit systemMessageChanged(info);
 }
 
 void MySettings::setModelChatNamePrompt(const ModelInfo &info, const QString &value, bool force)
@@ -443,7 +612,7 @@ void MySettings::setThreadCount(int value)
     emit threadCountChanged();
 }
 
-bool        MySettings::saveChatsContext() const        { return getBasicSetting("saveChatsContext"        ).toBool(); }
+bool        MySettings::systemTray() const              { return getBasicSetting("systemTray"              ).toBool(); }
 bool        MySettings::serverChat() const              { return getBasicSetting("serverChat"              ).toBool(); }
 int         MySettings::networkPort() const             { return getBasicSetting("networkPort"             ).toInt(); }
 QString     MySettings::userDefaultModel() const        { return getBasicSetting("userDefaultModel"        ).toString(); }
@@ -461,7 +630,7 @@ ChatTheme      MySettings::chatTheme() const      { return ChatTheme     (getEnu
 FontSize       MySettings::fontSize() const       { return FontSize      (getEnumSetting("fontSize",  fontSizeNames)); }
 SuggestionMode MySettings::suggestionMode() const { return SuggestionMode(getEnumSetting("suggestionMode", suggestionModeNames)); }
 
-void MySettings::setSaveChatsContext(bool value)                      { setBasicSetting("saveChatsContext",         value); }
+void MySettings::setSystemTray(bool value)                            { setBasicSetting("systemTray",               value); }
 void MySettings::setServerChat(bool value)                            { setBasicSetting("serverChat",               value); }
 void MySettings::setNetworkPort(int value)                            { setBasicSetting("networkPort",              value); }
 void MySettings::setUserDefaultModel(const QString &value)            { setBasicSetting("userDefaultModel",         value); }

gpt4all-chat/src/mysettings.h

@@ -4,6 +4,9 @@
 #include "modellist.h" // IWYU pragma: keep
 
 #include <QDateTime>
+#include <QLatin1StringView>
+#include <QList>
+#include <QModelIndex>
 #include <QObject>
 #include <QSettings>
 #include <QString>
@@ -48,7 +51,7 @@ class MySettings : public QObject
 {
     Q_OBJECT
     Q_PROPERTY(int threadCount READ threadCount WRITE setThreadCount NOTIFY threadCountChanged)
-    Q_PROPERTY(bool saveChatsContext READ saveChatsContext WRITE setSaveChatsContext NOTIFY saveChatsContextChanged)
+    Q_PROPERTY(bool systemTray READ systemTray WRITE setSystemTray NOTIFY systemTrayChanged)
     Q_PROPERTY(bool serverChat READ serverChat WRITE setServerChat NOTIFY serverChatChanged)
     Q_PROPERTY(QString modelPath READ modelPath WRITE setModelPath NOTIFY modelPathChanged)
     Q_PROPERTY(QString userDefaultModel READ userDefaultModel WRITE setUserDefaultModel NOTIFY userDefaultModelChanged)
@@ -74,9 +77,18 @@ class MySettings : public QObject
     Q_PROPERTY(SuggestionMode suggestionMode READ suggestionMode WRITE setSuggestionMode NOTIFY suggestionModeChanged)
     Q_PROPERTY(QStringList uiLanguages MEMBER m_uiLanguages CONSTANT)
 
+private:
+    explicit MySettings();
+    ~MySettings() override = default;
+
+public Q_SLOTS:
+    void onModelInfoChanged(const QModelIndex &topLeft, const QModelIndex &bottomRight, const QList<int> &roles = {});
+
 public:
     static MySettings *globalInstance();
 
+    Q_INVOKABLE static QVariant checkJinjaTemplateError(const QString &tmpl);
+
     // Restore methods
     Q_INVOKABLE void restoreModelDefaults(const ModelInfo &info);
     Q_INVOKABLE void restoreApplicationDefaults();
@@ -124,10 +136,14 @@ public:
     Q_INVOKABLE void setModelRepeatPenalty(const ModelInfo &info, double value, bool force = false);
     int modelRepeatPenaltyTokens(const ModelInfo &info) const;
     Q_INVOKABLE void setModelRepeatPenaltyTokens(const ModelInfo &info, int value, bool force = false);
-    QString modelPromptTemplate(const ModelInfo &info) const;
-    Q_INVOKABLE void setModelPromptTemplate(const ModelInfo &info, const QString &value, bool force = false);
-    QString modelSystemPrompt(const ModelInfo &info) const;
-    Q_INVOKABLE void setModelSystemPrompt(const ModelInfo &info, const QString &value, bool force = false);
+    auto modelChatTemplate(const ModelInfo &info) const -> UpgradeableSetting;
+    Q_INVOKABLE bool isModelChatTemplateSet(const ModelInfo &info) const;
+    Q_INVOKABLE void setModelChatTemplate(const ModelInfo &info, const QString &value);
+    Q_INVOKABLE void resetModelChatTemplate(const ModelInfo &info);
+    auto modelSystemMessage(const ModelInfo &info) const -> UpgradeableSetting;
+    Q_INVOKABLE bool isModelSystemMessageSet(const ModelInfo &info) const;
+    Q_INVOKABLE void setModelSystemMessage(const ModelInfo &info, const QString &value);
+    Q_INVOKABLE void resetModelSystemMessage(const ModelInfo &info);
     int modelContextLength(const ModelInfo &info) const;
     Q_INVOKABLE void setModelContextLength(const ModelInfo &info, int value, bool force = false);
     int modelGpuLayers(const ModelInfo &info) const;
@@ -140,8 +156,8 @@ public:
     // Application settings
     int threadCount() const;
     void setThreadCount(int value);
-    bool saveChatsContext() const;
-    void setSaveChatsContext(bool value);
+    bool systemTray() const;
+    void setSystemTray(bool value);
     bool serverChat() const;
     void setServerChat(bool value);
     QString modelPath();
@@ -212,12 +228,12 @@ Q_SIGNALS:
     void gpuLayersChanged(const ModelInfo &info);
     void repeatPenaltyChanged(const ModelInfo &info);
     void repeatPenaltyTokensChanged(const ModelInfo &info);
-    void promptTemplateChanged(const ModelInfo &info);
-    void systemPromptChanged(const ModelInfo &info);
+    void chatTemplateChanged(const ModelInfo &info, bool fromInfo = false);
+    void systemMessageChanged(const ModelInfo &info, bool fromInfo = false);
     void chatNamePromptChanged(const ModelInfo &info);
     void suggestedFollowUpPromptChanged(const ModelInfo &info);
     void threadCountChanged();
-    void saveChatsContextChanged();
+    void systemTrayChanged();
     void serverChatChanged();
     void modelPathChanged();
     void userDefaultModelChanged();
@@ -241,6 +257,30 @@ Q_SIGNALS:
     void suggestionModeChanged();
     void languageAndLocaleChanged();
 
+private:
+    QVariant getBasicSetting(const QString &name) const;
+    void setBasicSetting(const QString &name, const QVariant &value, std::optional<QString> signal = std::nullopt);
+    int getEnumSetting(const QString &setting, const QStringList &valueNames) const;
+    QVariant getModelSetting(QLatin1StringView name, const ModelInfo &info) const;
+    QVariant getModelSetting(const char *name, const ModelInfo &info) const;
+    void setModelSetting(QLatin1StringView name, const ModelInfo &info, const QVariant &value, bool force,
+                         bool signal = false);
+    void setModelSetting(const char *name, const ModelInfo &info, const QVariant &value, bool force,
+                         bool signal = false);
+    auto getUpgradeableModelSetting(
+        const ModelInfo &info, QLatin1StringView legacyKey, QLatin1StringView newKey
+    ) const -> UpgradeableSetting;
+    bool isUpgradeableModelSettingSet(
+        const ModelInfo &info, QLatin1StringView legacyKey, QLatin1StringView newKey
+    ) const;
+    bool setUpgradeableModelSetting(
+        const ModelInfo &info, const QString &value, QLatin1StringView legacyKey, QLatin1StringView newKey
+    );
+    bool resetUpgradeableModelSetting(
+        const ModelInfo &info, QLatin1StringView legacyKey, QLatin1StringView newKey
+    );
+    QString filePathForLocale(const QLocale &locale);
+
 private:
     QSettings m_settings;
     bool m_forceMetal;
@@ -249,18 +289,7 @@ private:
     const QStringList m_uiLanguages;
     std::unique_ptr<QTranslator> m_translator;
 
-private:
-    explicit MySettings();
-    ~MySettings() {}
     friend class MyPrivateSettings;
-
-    QVariant getBasicSetting(const QString &name) const;
-    void setBasicSetting(const QString &name, const QVariant &value, std::optional<QString> signal = std::nullopt);
-    int getEnumSetting(const QString &setting, const QStringList &valueNames) const;
-    QVariant getModelSetting(const QString &name, const ModelInfo &info) const;
-    void setModelSetting(const QString &name, const ModelInfo &info, const QVariant &value, bool force,
-                         bool signal = false);
-    QString filePathForLocale(const QLocale &locale);
 };
 
 #endif // MYSETTINGS_H

gpt4all-chat/src/network.cpp

@@ -8,6 +8,7 @@
 #include "localdocsmodel.h"
 #include "modellist.h"
 #include "mysettings.h"
+#include "utils.h"
 
 #include <gpt4all-backend/llmodel.h>
 
@@ -192,11 +193,14 @@ bool Network::packageAndSendJson(const QString &ingestId, const QString &json)
         return false;
     }
 
+    auto *currentChat = ChatListModel::globalInstance()->currentChat();
+    Q_ASSERT(currentChat);
+    auto modelInfo = currentChat->modelInfo();
+
     Q_ASSERT(doc.isObject());
-    Q_ASSERT(ChatListModel::globalInstance()->currentChat());
     QJsonObject object = doc.object();
     object.insert("source", "gpt4all-chat");
-    object.insert("agent_id", ChatListModel::globalInstance()->currentChat()->modelInfo().filename());
+    object.insert("agent_id", modelInfo.filename());
     object.insert("submitter_id", m_uniqueId);
     object.insert("ingest_id", ingestId);
 
@@ -204,8 +208,9 @@ bool Network::packageAndSendJson(const QString &ingestId, const QString &json)
     if (!attribution.isEmpty())
         object.insert("network/attribution", attribution);
 
-    QString promptTemplate = ChatListModel::globalInstance()->currentChat()->modelInfo().promptTemplate();
-    object.insert("prompt_template", promptTemplate);
+    if (!modelInfo.id().isNull())
+        if (auto tmpl = modelInfo.chatTemplate().asModern())
+            object.insert("chat_template"_L1, *tmpl);
 
     QJsonDocument newDoc;
     newDoc.setObject(object);
@@ -358,7 +363,8 @@ void Network::sendStartup()
 
 void Network::trackChatEvent(const QString &ev, QVariantMap props)
 {
-    const auto &curChat = ChatListModel::globalInstance()->currentChat();
+    auto *curChat = ChatListModel::globalInstance()->currentChat();
+    Q_ASSERT(curChat);
     if (!props.contains("model"))
         props.insert("model", curChat->modelInfo().filename());
     props.insert("device_backend", curChat->deviceBackend());
@@ -366,7 +372,7 @@ void Network::trackChatEvent(const QString &ev, QVariantMap props)
     props.insert("doc_collections_enabled", curChat->collectionList().count());
     props.insert("doc_collections_total", LocalDocs::globalInstance()->localDocsModel()->rowCount());
     props.insert("datalake_active", MySettings::globalInstance()->networkIsActive());
-    props.insert("using_server", ChatListModel::globalInstance()->currentChat()->isServer());
+    props.insert("using_server", curChat->isServer());
     trackEvent(ev, props);
 }
 

gpt4all-chat/src/server.cpp

@@ -37,6 +37,10 @@
 #include <unordered_map>
 #include <utility>
 
+#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
+#   include <QTcpServer>
+#endif
+
 using namespace std::string_literals;
 using namespace Qt::Literals::StringLiterals;
 
@@ -309,11 +313,8 @@ const std::unordered_map<BaseCompletionRequest::Type, const char *> BaseCompleti
 class ChatRequest : public BaseCompletionRequest {
 public:
     struct Message {
-        enum class Role : uint8_t {
-            User,
-            Assistant,
-        };
-        Role role;
+        enum class Role { System, User, Assistant };
+        Role    role;
         QString content;
     };
 
@@ -345,7 +346,6 @@ protected:
         this->messages.clear();
         {
             QCborArray arr = value.toArray();
-            Message::Role nextRole = Message::Role::User;
             for (qsizetype i = 0; i < arr.size(); i++) {
                 const auto &elem = arr[i];
                 if (!elem.isMap())
@@ -356,9 +356,9 @@ protected:
                 QCborMap msg = elem.toMap();
                 Message res;
                 QString role = takeValue(msg, "role", String, /*required*/ true).toString();
-                if (role == u"system"_s)
-                    continue; // FIXME(jared): don't ignore these
-                if (role == u"user"_s) {
+                if (role == u"system"_s) {
+                    res.role = Message::Role::System;
+                } else if (role == u"user"_s) {
                     res.role = Message::Role::User;
                 } else if (role == u"assistant"_s) {
                     res.role = Message::Role::Assistant;
@@ -370,13 +370,7 @@ protected:
                     ));
                 }
                 res.content = takeValue(msg, "content", String, /*required*/ true).toString();
-                if (res.role != nextRole)
-                    throw InvalidRequestError(fmt::format(
-                        "Invalid 'messages[{}].role': did not expect '{}' here", i, role
-                    ));
                 this->messages.append(res);
-                nextRole = res.role == Message::Role::User ? Message::Role::Assistant
-                                                           : Message::Role::User;
 
                 if (!msg.isEmpty())
                     throw InvalidRequestError(fmt::format(
@@ -435,7 +429,6 @@ T &parseRequest(T &request, QJsonObject &&obj)
 Server::Server(Chat *chat)
     : ChatLLM(chat, true /*isServer*/)
     , m_chat(chat)
-    , m_server(nullptr)
 {
     connect(this, &Server::threadStarted, this, &Server::start);
     connect(this, &Server::databaseResultsChanged, this, &Server::handleDatabaseResultsChanged);
@@ -457,10 +450,23 @@ static QJsonObject requestFromJson(const QByteArray &request)
 void Server::start()
 {
     m_server = std::make_unique<QHttpServer>(this);
-    if (!m_server->listen(QHostAddress::LocalHost, MySettings::globalInstance()->networkPort())) {
-        qWarning() << "ERROR: Unable to start the server";
+#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
+    auto *tcpServer = new QTcpServer(m_server.get());
+#else
+    auto *tcpServer = m_server.get();
+#endif
+
+    auto port = MySettings::globalInstance()->networkPort();
+    if (!tcpServer->listen(QHostAddress::LocalHost, port)) {
+        qWarning() << "Server ERROR: Failed to listen on port" << port;
         return;
     }
+#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
+    if (!m_server->bind(tcpServer)) {
+        qWarning() << "Server ERROR: Failed to HTTP server to socket" << port;
+        return;
+    }
+#endif
 
     m_server->route("/v1/models", QHttpServerRequest::Method::Get,
         [](const QHttpServerRequest &) {
@@ -600,13 +606,21 @@ void Server::start()
         }
     );
 
-    m_server->afterRequest([] (QHttpServerResponse &&resp) {
+#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
+    m_server->addAfterRequestHandler(this, [](const QHttpServerRequest &req, QHttpServerResponse &resp) {
+        Q_UNUSED(req);
+        auto headers = resp.headers();
+        headers.append("Access-Control-Allow-Origin"_L1, "*"_L1);
+        resp.setHeaders(std::move(headers));
+    });
+#else
+    m_server->afterRequest([](QHttpServerResponse &&resp) {
         resp.addHeader("Access-Control-Allow-Origin", "*");
         return std::move(resp);
     });
+#endif
 
-    connect(this, &Server::requestServerNewPromptResponsePair, m_chat,
-        &Chat::serverNewPromptResponsePair, Qt::BlockingQueuedConnection);
+    connect(this, &Server::requestResetResponseState, m_chat, &Chat::resetResponseState, Qt::BlockingQueuedConnection);
 }
 
 static auto makeError(auto &&...args) -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>
@@ -617,6 +631,10 @@ static auto makeError(auto &&...args) -> std::pair<QHttpServerResponse, std::opt
 auto Server::handleCompletionRequest(const CompletionRequest &request)
     -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>
 {
+    Q_ASSERT(m_chatModel);
+
+    auto *mySettings = MySettings::globalInstance();
+
     ModelInfo modelInfo = ModelList::globalInstance()->defaultModelInfo();
     const QList<ModelInfo> modelList = ModelList::globalInstance()->selectableModelList();
     for (const ModelInfo &info : modelList) {
@@ -629,10 +647,6 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
         }
     }
 
-    // adds prompt/response items to GUI
-    emit requestServerNewPromptResponsePair(request.prompt); // blocks
-    resetResponse();
-
     // load the new model if necessary
     setShouldBeLoaded(true);
 
@@ -641,47 +655,55 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
         return makeError(QHttpServerResponder::StatusCode::InternalServerError);
     }
 
+    emit requestResetResponseState(); // blocks
+    qsizetype prevMsgIndex = m_chatModel->count() - 1;
+    if (prevMsgIndex >= 0)
+        m_chatModel->updateCurrentResponse(prevMsgIndex, false);
+
     // NB: this resets the context, regardless of whether this model is already loaded
     if (!loadModel(modelInfo)) {
         std::cerr << "ERROR: couldn't load model " << modelInfo.name().toStdString() << std::endl;
         return makeError(QHttpServerResponder::StatusCode::InternalServerError);
     }
 
-    // FIXME(jared): taking parameters from the UI inhibits reproducibility of results
-    const int  top_k          = modelInfo.topK();
-    const int  n_batch        = modelInfo.promptBatchSize();
-    const auto repeat_penalty = float(modelInfo.repeatPenalty());
-    const int  repeat_last_n  = modelInfo.repeatPenaltyTokens();
+    // add prompt/response items to GUI
+    m_chatModel->appendPrompt(request.prompt);
+    m_chatModel->appendResponse();
 
+    // FIXME(jared): taking parameters from the UI inhibits reproducibility of results
+    LLModel::PromptContext promptCtx {
+        .n_predict      = request.max_tokens,
+        .top_k          = mySettings->modelTopK(modelInfo),
+        .top_p          = request.top_p,
+        .min_p          = request.min_p,
+        .temp           = request.temperature,
+        .n_batch        = mySettings->modelPromptBatchSize(modelInfo),
+        .repeat_penalty = float(mySettings->modelRepeatPenalty(modelInfo)),
+        .repeat_last_n  = mySettings->modelRepeatPenaltyTokens(modelInfo),
+    };
+
+    auto promptUtf8 = request.prompt.toUtf8();
     int promptTokens = 0;
     int responseTokens = 0;
-    QList<QPair<QString, QList<ResultInfo>>> responses;
+    QStringList responses;
     for (int i = 0; i < request.n; ++i) {
-        if (!promptInternal(
-            m_collections,
-            request.prompt,
-            /*promptTemplate*/ u"%1"_s,
-            request.max_tokens,
-            top_k,
-            request.top_p,
-            request.min_p,
-            request.temperature,
-            n_batch,
-            repeat_penalty,
-            repeat_last_n)) {
-
-            std::cerr << "ERROR: couldn't prompt model " << modelInfo.name().toStdString() << std::endl;
+        PromptResult result;
+        try {
+            result = promptInternal(std::string_view(promptUtf8.cbegin(), promptUtf8.cend()),
+                                    promptCtx,
+                                    /*usedLocalDocs*/ false);
+        } catch (const std::exception &e) {
+            emit responseChanged(e.what());
+            emit responseStopped(0);
             return makeError(QHttpServerResponder::StatusCode::InternalServerError);
         }
-        QString resp = response(/*trim*/ false);
+        QString resp = QString::fromUtf8(result.response);
         if (request.echo)
             resp = request.prompt + resp;
-        responses.append({resp, m_databaseResults});
-        if (!promptTokens)
-            promptTokens = m_promptTokens;
-        responseTokens += m_promptResponseTokens - m_promptTokens;
-        if (i < request.n - 1)
-            resetResponse();
+        responses << resp;
+        if (i == 0)
+            promptTokens = result.promptTokens;
+        responseTokens += result.responseTokens;
     }
 
     QJsonObject responseObject {
@@ -692,25 +714,13 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
     };
 
     QJsonArray choices;
-    {
-        int index = 0;
-        for (const auto &r : responses) {
-            QString result = r.first;
-            QList<ResultInfo> infos = r.second;
-            QJsonObject choice {
-                { "text",          result                                                   },
-                { "index",         index++                                                  },
-                { "logprobs",      QJsonValue::Null                                         },
-                { "finish_reason", responseTokens == request.max_tokens ? "length" : "stop" },
-            };
-            if (MySettings::globalInstance()->localDocsShowReferences()) {
-                QJsonArray references;
-                for (const auto &ref : infos)
-                    references.append(resultToJson(ref));
-                choice.insert("references", references.isEmpty() ? QJsonValue::Null : QJsonValue(references));
-            }
-            choices.append(choice);
-        }
+    for (qsizetype i = 0; auto &resp : std::as_const(responses)) {
+        choices << QJsonObject {
+            { "text",          resp                                                     },
+            { "index",         i++                                                      },
+            { "logprobs",      QJsonValue::Null                                         },
+            { "finish_reason", responseTokens == request.max_tokens ? "length" : "stop" },
+        };
     }
 
     responseObject.insert("choices", choices);
@@ -726,6 +736,8 @@ auto Server::handleCompletionRequest(const CompletionRequest &request)
 auto Server::handleChatRequest(const ChatRequest &request)
     -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>
 {
+    auto *mySettings = MySettings::globalInstance();
+
     ModelInfo modelInfo = ModelList::globalInstance()->defaultModelInfo();
     const QList<ModelInfo> modelList = ModelList::globalInstance()->selectableModelList();
     for (const ModelInfo &info : modelList) {
@@ -746,83 +758,58 @@ auto Server::handleChatRequest(const ChatRequest &request)
         return makeError(QHttpServerResponder::StatusCode::InternalServerError);
     }
 
+    emit requestResetResponseState(); // blocks
+
     // NB: this resets the context, regardless of whether this model is already loaded
     if (!loadModel(modelInfo)) {
         std::cerr << "ERROR: couldn't load model " << modelInfo.name().toStdString() << std::endl;
         return makeError(QHttpServerResponder::StatusCode::InternalServerError);
     }
 
-    const QString promptTemplate = modelInfo.promptTemplate();
-    const int     top_k          = modelInfo.topK();
-    const int     n_batch        = modelInfo.promptBatchSize();
-    const auto    repeat_penalty = float(modelInfo.repeatPenalty());
-    const int     repeat_last_n  = modelInfo.repeatPenaltyTokens();
+    m_chatModel->updateCurrentResponse(m_chatModel->count() - 1, false);
 
-    int promptTokens = 0;
+    Q_ASSERT(!request.messages.isEmpty());
+
+    // adds prompt/response items to GUI
+    std::vector<ChatItem> chatItems;
+    for (auto &message : request.messages) {
+        using enum ChatRequest::Message::Role;
+        switch (message.role) {
+            case System:    chatItems.emplace_back(ChatItem::system_tag, message.content); break;
+            case User:      chatItems.emplace_back(ChatItem::prompt_tag, message.content); break;
+            case Assistant: chatItems.emplace_back(ChatItem::response_tag, /*currentResponse*/ false); break;
+        }
+    }
+    m_chatModel->appendResponseWithHistory(chatItems);
+
+    // FIXME(jared): taking parameters from the UI inhibits reproducibility of results
+    LLModel::PromptContext promptCtx {
+        .n_predict      = request.max_tokens,
+        .top_k          = mySettings->modelTopK(modelInfo),
+        .top_p          = request.top_p,
+        .min_p          = request.min_p,
+        .temp           = request.temperature,
+        .n_batch        = mySettings->modelPromptBatchSize(modelInfo),
+        .repeat_penalty = float(mySettings->modelRepeatPenalty(modelInfo)),
+        .repeat_last_n  = mySettings->modelRepeatPenaltyTokens(modelInfo),
+    };
+
+    int promptTokens   = 0;
     int responseTokens = 0;
     QList<QPair<QString, QList<ResultInfo>>> responses;
-    Q_ASSERT(!request.messages.isEmpty());
-    Q_ASSERT(request.messages.size() % 2 == 1);
-    for (int i = 0; i < request.messages.size() - 2; i += 2) {
-        using enum ChatRequest::Message::Role;
-        auto &user      = request.messages[i];
-        auto &assistant = request.messages[i + 1];
-        Q_ASSERT(user.role      == User);
-        Q_ASSERT(assistant.role == Assistant);
-
-        // adds prompt/response items to GUI
-        emit requestServerNewPromptResponsePair(user.content); // blocks
-        resetResponse();
-
-        if (!promptInternal(
-            {},
-            user.content,
-            promptTemplate,
-            request.max_tokens,
-            top_k,
-            request.top_p,
-            request.min_p,
-            request.temperature,
-            n_batch,
-            repeat_penalty,
-            repeat_last_n,
-            assistant.content)
-        ) {
-            std::cerr << "ERROR: couldn't prompt model " << modelInfo.name().toStdString() << std::endl;
-            return makeError(QHttpServerResponder::StatusCode::InternalServerError);
-        }
-        promptTokens += m_promptResponseTokens; // previous responses are part of current prompt
-    }
-
-    QString lastMessage = request.messages.last().content;
-    // adds prompt/response items to GUI
-    emit requestServerNewPromptResponsePair(lastMessage); // blocks
-    resetResponse();
-
     for (int i = 0; i < request.n; ++i) {
-        if (!promptInternal(
-            m_collections,
-            lastMessage,
-            promptTemplate,
-            request.max_tokens,
-            top_k,
-            request.top_p,
-            request.min_p,
-            request.temperature,
-            n_batch,
-            repeat_penalty,
-            repeat_last_n)
-        ) {
-            std::cerr << "ERROR: couldn't prompt model " << modelInfo.name().toStdString() << std::endl;
+        ChatPromptResult result;
+        try {
+            result = promptInternalChat(m_collections, promptCtx);
+        } catch (const std::exception &e) {
+            emit responseChanged(e.what());
+            emit responseStopped(0);
             return makeError(QHttpServerResponder::StatusCode::InternalServerError);
         }
-        responses.append({response(), m_databaseResults});
-        // FIXME(jared): these are UI counts and do not include framing tokens, which they should
+        responses.emplace_back(result.response, result.databaseResults);
         if (i == 0)
-            promptTokens += m_promptTokens;
-        responseTokens += m_promptResponseTokens - m_promptTokens;
-        if (i != request.n - 1)
-            resetResponse();
+            promptTokens = result.promptTokens;
+        responseTokens += result.responseTokens;
     }
 
     QJsonObject responseObject {

gpt4all-chat/src/server.h

@@ -33,7 +33,7 @@ public Q_SLOTS:
     void start();
 
 Q_SIGNALS:
-    void requestServerNewPromptResponsePair(const QString &prompt, const QList<PromptAttachment> &attachments = {});
+    void requestResetResponseState();
 
 private:
     auto handleCompletionRequest(const CompletionRequest &request) -> std::pair<QHttpServerResponse, std::optional<QJsonObject>>;

gpt4all-chat/src/utils.h

@@ -3,23 +3,41 @@
 #include <fmt/base.h>
 #include <fmt/format.h>
 
+#include <QByteArray>
+#include <QJsonValue>
+#include <QLatin1StringView>
 #include <QString>
+#include <QStringView>
+#include <QUtf8StringView>
 #include <QVariant>
 
-#include <string>
+#include <initializer_list>
+#include <string_view>
+#include <utility>
+
+class QJsonObject;
 
 
 // fmtlib formatters for QString and QVariant
 
-#define MAKE_FORMATTER(type, conversion)                                      \
-    template <>                                                               \
-    struct fmt::formatter<type, char>: fmt::formatter<std::string, char> {    \
-        template <typename FmtContext>                                        \
-        FmtContext::iterator format(const type &value, FmtContext &ctx) const \
-        {                                                                     \
-            return formatter<std::string, char>::format(conversion, ctx);     \
-        }                                                                     \
+#define MAKE_FORMATTER(type, conversion)                                        \
+    template <>                                                                 \
+    struct fmt::formatter<type, char>: fmt::formatter<std::string_view, char> { \
+        template <typename FmtContext>                                          \
+        FmtContext::iterator format(const type &value, FmtContext &ctx) const   \
+        {                                                                       \
+            auto valueUtf8 = (conversion);                                      \
+            std::string_view view(valueUtf8.cbegin(), valueUtf8.cend());        \
+            return formatter<std::string_view, char>::format(view, ctx);        \
+        }                                                                       \
     }
 
-MAKE_FORMATTER(QString,  value.toStdString()           );
-MAKE_FORMATTER(QVariant, value.toString().toStdString());
+MAKE_FORMATTER(QUtf8StringView, value                    );
+MAKE_FORMATTER(QStringView,     value.toUtf8()           );
+MAKE_FORMATTER(QString,         value.toUtf8()           );
+MAKE_FORMATTER(QVariant,        value.toString().toUtf8());
+
+// alternative to QJsonObject's initializer_list constructor that accepts Latin-1 strings
+QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringView, QJsonValue>> args);
+
+#include "utils.inl"

gpt4all-chat/src/utils.inl

@@ -0,0 +1,9 @@
+#include <QJsonObject>
+
+inline QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringView, QJsonValue>> args)
+{
+    QJsonObject obj;
+    for (auto &arg : args)
+        obj.insert(arg.first, arg.second);
+    return obj;
+}

gpt4all-chat/src/xlsxtomd.cpp

@@ -10,8 +10,10 @@
 #include <QDateTime>
 #include <QDebug>
 #include <QList>
+#include <QRegularExpression>
 #include <QString>
 #include <QStringList>
+#include <QStringView>
 #include <QVariant>
 #include <QtGlobal>
 #include <QtLogging>
@@ -30,10 +32,10 @@ static QString formatCellText(const QXlsx::Cell *cell)
     QString cellText;
 
     // Determine the cell type based on format
-    if (format.isDateTimeFormat()) {
+    if (cell->isDateTime()) {
         // Handle DateTime
-        QDateTime dateTime = value.toDateTime();
-        cellText = dateTime.isValid() ? dateTime.toString("yyyy-MM-dd") : value.toString();
+        QDateTime dateTime = cell->dateTime().toDateTime();
+        cellText = dateTime.isValid() ? dateTime.toString(QStringView(u"yyyy-MM-dd")) : value.toString();
     } else {
         cellText = value.toString();
     }
@@ -41,23 +43,32 @@ static QString formatCellText(const QXlsx::Cell *cell)
     if (cellText.isEmpty())
         return QString();
 
-    // Apply Markdown and HTML formatting based on font styles
-    QString formattedText = cellText;
-
-    if (format.fontBold() && format.fontItalic())
-        formattedText = "***" + formattedText + "***";
-    else if (format.fontBold())
-        formattedText = "**" + formattedText + "**";
-    else if (format.fontItalic())
-        formattedText = "*" + formattedText + "*";
+    // Escape special characters
+    static QRegularExpression special(
+        QStringLiteral(
+            R"(()([\\`*_[\]<>()!|])|)"    // special characters
+            R"(^(\s*)(#+(?:\s|$))|)"      // headings
+            R"(^(\s*[0-9])(\.(?:\s|$))|)" // ordered lists ("1. a")
+            R"(^(\s*)([+-](?:\s|$)))"     // unordered lists ("- a")
+        ),
+        QRegularExpression::MultilineOption
+    );
+    cellText.replace(special, uR"(\1\\2)"_s);
+    cellText.replace(u'&', "&amp;"_L1);
+    cellText.replace(u'<', "&lt;"_L1);
+    cellText.replace(u'>', "&gt;"_L1);
 
+    // Apply Markdown formatting based on font styles
+    if (format.fontUnderline())
+        cellText = u"_%1_"_s.arg(cellText);
+    if (format.fontBold())
+        cellText = u"**%1**"_s.arg(cellText);
+    if (format.fontItalic())
+        cellText = u"*%1*"_s.arg(cellText);
     if (format.fontStrikeOut())
-        formattedText = "~~" + formattedText + "~~";
+        cellText = u"~~%1~~"_s.arg(cellText);
 
-    // Escape pipe characters to prevent Markdown table issues
-    formattedText.replace("|", "\\|");
-
-    return formattedText;
+    return cellText;
 }
 
 static QString getCellValue(QXlsx::Worksheet *sheet, int row, int col)
@@ -113,7 +124,7 @@ QString XLSXToMD::toMarkdown(QIODevice *xlsxDevice)
             continue;
         }
 
-        markdown += u"## %1\n\n"_s.arg(sheetName);
+        markdown += u"### %1\n\n"_s.arg(sheetName);
 
         // Determine the used range
         QXlsx::CellRange range = sheet->dimension();
@@ -124,44 +135,35 @@ QString XLSXToMD::toMarkdown(QIODevice *xlsxDevice)
 
         if (firstRow > lastRow || firstCol > lastCol) {
             qWarning() << "Sheet" << sheetName << "is empty.";
-            markdown += "*No data available.*\n\n";
+            markdown += QStringView(u"*No data available.*\n\n");
             continue;
         }
 
-        // Assume the first row is the header
-        int headerRow = firstRow;
+        auto appendRow = [&markdown](auto &list) { markdown += u"|%1|\n"_s.arg(list.join(u'|')); };
 
-        // Collect headers
+        // Empty header
+        static QString header(u' ');
+        static QString separator(u'-');
         QStringList headers;
-        for (int col = firstCol; col <= lastCol; ++col) {
-            QString header = getCellValue(sheet, headerRow, col);
-            headers << header;
-        }
-
-        // Create Markdown header row
-        QString headerRowMarkdown = "|" + headers.join("|") + "|";
-        markdown += headerRowMarkdown + "\n";
-
-        // Create Markdown separator row
         QStringList separators;
-        for (int i = 0; i < headers.size(); ++i)
-            separators << "---";
-        QString separatorRow = "|" + separators.join("|") + "|";
-        markdown += separatorRow + "\n";
+        for (int col = firstCol; col <= lastCol; ++col) {
+            headers << header;
+            separators << separator;
+        }
+        appendRow(headers);
+        appendRow(separators);
 
-        // Iterate through data rows (starting from the row after header)
-        for (int row = headerRow + 1; row <= lastRow; ++row) {
+        // Iterate through data rows
+        for (int row = firstRow; row <= lastRow; ++row) {
             QStringList rowData;
             for (int col = firstCol; col <= lastCol; ++col) {
                 QString cellText = getCellValue(sheet, row, col);
-                rowData << cellText;
+                rowData << (cellText.isEmpty() ? u" "_s : cellText);
             }
-
-            QString dataRow = "|" + rowData.join("|") + "|";
-            markdown += dataRow + "\n";
+            appendRow(rowData);
         }
 
-        markdown += "\n"; // Add an empty line between sheets
+        markdown += u'\n'; // Add an empty line between sheets
     }
     return markdown;
 }

gpt4all-chat/test-requirements.txt

@@ -0,0 +1,2 @@
+pytest~=8.3
+requests~=2.32

gpt4all-chat/tests/CMakeLists.txt

@@ -0,0 +1,30 @@
+include(FetchContent)
+
+find_package(Python3 3.12 REQUIRED COMPONENTS Interpreter)
+
+# Google test download and setup
+FetchContent_Declare(
+    googletest
+    URL https://github.com/google/googletest/archive/refs/tags/v1.15.2.zip
+)
+FetchContent_MakeAvailable(googletest)
+
+configure_file(python/config.py.in "${CMAKE_CURRENT_SOURCE_DIR}/python/config.py")
+
+add_test(NAME ChatPythonTests
+    COMMAND ${Python3_EXECUTABLE} -m pytest --color=yes "${CMAKE_CURRENT_SOURCE_DIR}/python"
+)
+set_tests_properties(ChatPythonTests PROPERTIES
+    ENVIRONMENT "CHAT_EXECUTABLE=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/chat;TEST_MODEL_PATH=${TEST_MODEL_PATH}"
+    TIMEOUT 60
+)
+
+add_executable(gpt4all_tests
+    cpp/test_main.cpp
+    cpp/basic_test.cpp
+)
+
+target_link_libraries(gpt4all_tests PRIVATE gtest gtest_main)
+
+include(GoogleTest)
+gtest_discover_tests(gpt4all_tests)

gpt4all-chat/tests/cpp/basic_test.cpp

@@ -0,0 +1,5 @@
+#include <gtest/gtest.h>
+
+TEST(BasicTest, TestInitialization) {
+    EXPECT_TRUE(true);
+}

gpt4all-chat/tests/cpp/test_main.cpp

@@ -0,0 +1,6 @@
+#include <gtest/gtest.h>
+
+int main(int argc, char **argv) {
+    ::testing::InitGoogleTest(&argc, argv);
+    return RUN_ALL_TESTS();
+}