chat: cut v3.9.0 release (#3461)

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

.circleci/config.yml

@@ -8,6 +8,10 @@ workflows:
   generate-config:
     jobs:
       - path-filtering/filter:
+          filters:
+            tags:
+              only:
+                - /.*/
           base-revision: main
           config-path: .circleci/continue_config.yml
           mapping: |

.circleci/continue_config.yml

@@ -22,10 +22,17 @@ jobs:
   # work around CircleCI-Public/path-filtering-orb#20
   noop:
     docker:
-      - image: cimg/base:current
+      - image: cimg/base:stable
     steps:
       - run: "true"
-
+  validate-commit-on-main:
+    docker:
+      - image: cimg/base:stable
+    steps:
+      - checkout
+      - run:
+          name: Verify that commit is on the main branch
+          command: git merge-base --is-ancestor HEAD main
   build-offline-chat-installer-macos:
     macos:
       xcode: 15.4.0
@@ -48,10 +55,15 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
-            hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
+            wget "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-macOS-x64-4.8.1.dmg"
+            hdiutil attach qt-online-installer-macOS-x64-4.8.1.dmg
+            /Volumes/qt-online-installer-macOS-x64-4.8.1/qt-online-installer-macOS-x64-4.8.1.app/Contents/MacOS/qt-online-installer-macOS-x64-4.8.1 \
+              --no-force-installations --no-default-installations --no-size-checking --default-answer \
+              --accept-licenses --confirm-command --accept-obligations --email "$QT_EMAIL" --password "$QT_PASSWORD" \
+              install \
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.clang_64 qt.qt6.682.addons.qt5compat \
+              extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
+            hdiutil detach /Volumes/qt-online-installer-macOS-x64-4.8.1
       - run:
           name: Setup Keychain
           command: |
@@ -72,19 +84,29 @@ jobs:
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
               -S ../gpt4all-chat -B . -G Ninja \
               -DCMAKE_BUILD_TYPE=Release \
-              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.5.1/macos/lib/cmake \
+              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.8.2/macos/lib/cmake \
               -DCMAKE_MAKE_PROGRAM:FILEPATH=~/Qt/Tools/Ninja/ninja \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
               -DBUILD_UNIVERSAL=ON \
               -DCMAKE_OSX_DEPLOYMENT_TARGET=12.6 \
               -DGGML_METAL_MACOSX_VERSION_MIN=12.6 \
-              -DMACDEPLOYQT=~/Qt/6.5.1/macos/bin/macdeployqt \
+              -DMACDEPLOYQT=~/Qt/6.8.2/macos/bin/macdeployqt \
               -DGPT4ALL_OFFLINE_INSTALLER=ON \
               -DGPT4ALL_SIGN_INSTALL=ON \
               -DGPT4ALL_GEN_CPACK_CONFIG=ON
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target package
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake . -DGPT4ALL_GEN_CPACK_CONFIG=OFF
+            # The 'install' step here *should* be completely unnecessary. There is absolutely no reason we should have
+            # to copy all of the build artifacts to an output directory that we do not use (because we package GPT4All
+            # as an installer instead).
+            # However, because of the way signing is implemented in the cmake script, the *source* files are signed at
+            # install time instead of the *installed* files. This side effect is the *only* way libraries that are not
+            # processed by macdeployqt, such as libllmodel.so, get signed (at least, with -DBUILD_UNIVERSAL=ON).
+            # Also, we have to run this as a *separate* step. Telling cmake to run both targets in one command causes it
+            # to execute them in parallel, since it is not aware of the dependency of the package target on the install
+            # target.
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target install
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target package
             ccache -s
             mkdir upload
@@ -164,6 +186,21 @@ jobs:
             xcrun stapler staple build/upload/gpt4all-installer-darwin-signed.dmg
       - store_artifacts:
           path: build/upload
+      - run:
+          name: Install Rosetta
+          command: softwareupdate --install-rosetta --agree-to-license  # needed for QtIFW
+      - run:
+          name: Test installation and verify that it is signed
+          command: |
+            set -e
+            hdiutil attach build/upload/gpt4all-installer-darwin-signed.dmg
+            codesign --verify --deep --verbose /Volumes/gpt4all-installer-darwin/gpt4all-installer-darwin.app
+            /Volumes/gpt4all-installer-darwin/gpt4all-installer-darwin.app/Contents/MacOS/gpt4all-installer-darwin \
+              --no-size-checking --default-answer --accept-licenses --confirm-command \
+              install gpt4all
+            codesign --verify --deep --verbose /Applications/gpt4all/bin/gpt4all.app
+            codesign --verify --deep --verbose /Applications/gpt4all/maintenancetool.app
+            hdiutil detach /Volumes/gpt4all-installer-darwin
 
   build-online-chat-installer-macos:
     macos:
@@ -187,10 +224,15 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
-            hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
+            wget "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-macOS-x64-4.8.1.dmg"
+            hdiutil attach qt-online-installer-macOS-x64-4.8.1.dmg
+            /Volumes/qt-online-installer-macOS-x64-4.8.1/qt-online-installer-macOS-x64-4.8.1.app/Contents/MacOS/qt-online-installer-macOS-x64-4.8.1 \
+              --no-force-installations --no-default-installations --no-size-checking --default-answer \
+              --accept-licenses --confirm-command --accept-obligations --email "$QT_EMAIL" --password "$QT_PASSWORD" \
+              install \
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.clang_64 qt.qt6.682.addons.qt5compat \
+              extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
+            hdiutil detach /Volumes/qt-online-installer-macOS-x64-4.8.1
       - run:
           name: Setup Keychain
           command: |
@@ -211,19 +253,21 @@ jobs:
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
               -S ../gpt4all-chat -B . -G Ninja \
               -DCMAKE_BUILD_TYPE=Release \
-              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.5.1/macos/lib/cmake \
+              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.8.2/macos/lib/cmake \
               -DCMAKE_MAKE_PROGRAM:FILEPATH=~/Qt/Tools/Ninja/ninja \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
               -DBUILD_UNIVERSAL=ON \
               -DCMAKE_OSX_DEPLOYMENT_TARGET=12.6 \
               -DGGML_METAL_MACOSX_VERSION_MIN=12.6 \
-              -DMACDEPLOYQT=~/Qt/6.5.1/macos/bin/macdeployqt \
+              -DMACDEPLOYQT=~/Qt/6.8.2/macos/bin/macdeployqt \
               -DGPT4ALL_OFFLINE_INSTALLER=OFF \
               -DGPT4ALL_SIGN_INSTALL=ON \
               -DGPT4ALL_GEN_CPACK_CONFIG=ON
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target package
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake . -DGPT4ALL_GEN_CPACK_CONFIG=OFF
+            # See comment above related to the 'install' target.
+            ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target install
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake --build . --target package
             ccache -s
             mkdir upload
@@ -304,6 +348,22 @@ jobs:
             xcrun stapler staple build/upload/gpt4all-installer-darwin-signed.dmg
       - store_artifacts:
           path: build/upload
+      - run:
+          name: Install Rosetta
+          command: softwareupdate --install-rosetta --agree-to-license  # needed for QtIFW
+      - run:
+          name: Test installation and verify that it is signed
+          command: |
+            set -e
+            hdiutil attach build/upload/gpt4all-installer-darwin-signed.dmg
+            codesign --verify --deep --verbose /Volumes/gpt4all-installer-darwin/gpt4all-installer-darwin.app
+            tar -xf build/upload/repository.tar.gz
+            /Volumes/gpt4all-installer-darwin/gpt4all-installer-darwin.app/Contents/MacOS/gpt4all-installer-darwin \
+              --no-size-checking --default-answer --accept-licenses --confirm-command --set-temp-repository repository \
+              install gpt4all
+            codesign --verify --deep --verbose /Applications/gpt4all/bin/gpt4all.app
+            codesign --verify --deep --verbose /Applications/gpt4all/maintenancetool.app
+            hdiutil detach /Volumes/gpt4all-installer-darwin
 
   build-offline-chat-installer-linux:
     machine:
@@ -340,9 +400,13 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            wget "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-linux-x64-4.8.1.run"
+            chmod +x qt-online-installer-linux-x64-4.8.1.run
+            ./qt-online-installer-linux-x64-4.8.1.run --no-force-installations --no-default-installations \
+              --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations \
+              --email "$QT_EMAIL" --password "$QT_PASSWORD" install \
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.linux_gcc_64 qt.qt6.682.addons.qt5compat \
+              qt.qt6.682.debug_info extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
       - run:
           name: Build linuxdeployqt
           command: |
@@ -353,7 +417,7 @@ jobs:
           no_output_timeout: 30m
           command: |
             set -eo pipefail
-            export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
+            export CMAKE_PREFIX_PATH=~/Qt/6.8.2/gcc_64/lib/cmake
             export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.8/bin
             export PATH=$PATH:/usr/local/cuda/bin
             ccache -o "cache_dir=${PWD}/../.ccache" -o max_size=500M -p -z
@@ -382,6 +446,13 @@ jobs:
           when: always
           paths:
             - ../.ccache
+      - run:
+          name: Test installation
+          command: |
+            mkdir ~/Desktop
+            build/upload/gpt4all-installer-linux.run --no-size-checking --default-answer --accept-licenses \
+              --confirm-command \
+              install gpt4all
 
   build-online-chat-installer-linux:
     machine:
@@ -418,9 +489,13 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            wget "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-linux-x64-4.8.1.run"
+            chmod +x qt-online-installer-linux-x64-4.8.1.run
+            ./qt-online-installer-linux-x64-4.8.1.run --no-force-installations --no-default-installations \
+              --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations \
+              --email "$QT_EMAIL" --password "$QT_PASSWORD" install \
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.linux_gcc_64 qt.qt6.682.addons.qt5compat \
+              qt.qt6.682.debug_info extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
       - run:
           name: Build linuxdeployqt
           command: |
@@ -431,7 +506,7 @@ jobs:
           no_output_timeout: 30m
           command: |
             set -eo pipefail
-            export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
+            export CMAKE_PREFIX_PATH=~/Qt/6.8.2/gcc_64/lib/cmake
             export PATH=$PATH:$HOME/Qt/Tools/QtInstallerFramework/4.8/bin
             export PATH=$PATH:/usr/local/cuda/bin
             ccache -o "cache_dir=${PWD}/../.ccache" -o max_size=500M -p -z
@@ -461,6 +536,14 @@ jobs:
           when: always
           paths:
             - ../.ccache
+      - run:
+          name: Test installation
+          command: |
+            mkdir ~/Desktop
+            build/upload/gpt4all-installer-linux.run --no-size-checking --default-answer --accept-licenses \
+              --confirm-command \
+              --set-temp-repository build/_CPack_Packages/Linux/IFW/gpt4all-installer-linux/repository \
+              install gpt4all
 
   build-offline-chat-installer-windows:
     machine:
@@ -484,8 +567,12 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            wget.exe "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-windows-x64-4.8.1.exe"
+            & .\qt-online-installer-windows-x64-4.8.1.exe --no-force-installations --no-default-installations `
+              --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations `
+              --email "${Env:QT_EMAIL}" --password "${Env:QT_PASSWORD}" install `
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.win64_msvc2022_64 qt.qt6.682.addons.qt5compat `
+              qt.qt6.682.debug_info extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
       - run:
           name: Install VulkanSDK
           command: |
@@ -526,7 +613,7 @@ jobs:
             & "C:\Qt\Tools\CMake_64\bin\cmake.exe" `
               -S ..\gpt4all-chat -B . -G Ninja `
               -DCMAKE_BUILD_TYPE=Release `
-              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_64" `
+              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.8.2\msvc2022_64" `
               "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
               -DCMAKE_C_COMPILER_LAUNCHER=ccache `
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache `
@@ -586,6 +673,12 @@ jobs:
             AzureSignTool.exe sign -du "https://gpt4all.io/index.html" -kvu https://gpt4all.vault.azure.net -kvi "$Env:AZSignGUID" -kvs "$Env:AZSignPWD" -kvc "$Env:AZSignCertName" -kvt "$Env:AZSignTID" -tr http://timestamp.digicert.com -v "$($(Get-Location).Path)\build\upload\gpt4all-installer-win64.exe"
       - store_artifacts:
           path: build/upload
+      - run:
+          name: Test installation
+          command: |
+            build\upload\gpt4all-installer-win64.exe --no-size-checking --default-answer --accept-licenses `
+              --confirm-command `
+              install gpt4all
 
   build-online-chat-installer-windows:
     machine:
@@ -608,8 +701,12 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            wget.exe "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-windows-x64-4.8.1.exe"
+            & .\qt-online-installer-windows-x64-4.8.1.exe --no-force-installations --no-default-installations `
+              --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations `
+              --email "${Env:QT_EMAIL}" --password "${Env:QT_PASSWORD}" install `
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.win64_msvc2022_64 qt.qt6.682.addons.qt5compat `
+              qt.qt6.682.debug_info extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
       - run:
           name: Install VulkanSDK
           command: |
@@ -655,7 +752,7 @@ jobs:
             & "C:\Qt\Tools\CMake_64\bin\cmake.exe" `
               -S ..\gpt4all-chat -B . -G Ninja `
               -DCMAKE_BUILD_TYPE=Release `
-              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_64" `
+              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.8.2\msvc2022_64" `
               "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
               -DCMAKE_C_COMPILER_LAUNCHER=ccache `
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache `
@@ -722,6 +819,13 @@ jobs:
             AzureSignTool.exe sign -du "https://gpt4all.io/index.html" -kvu https://gpt4all.vault.azure.net -kvi "$Env:AZSignGUID" -kvs "$Env:AZSignPWD" -kvc "$Env:AZSignCertName" -kvt "$Env:AZSignTID" -tr http://timestamp.digicert.com -v "$($(Get-Location).Path)/build/upload/gpt4all-installer-win64.exe"
       - store_artifacts:
           path: build/upload
+      - run:
+          name: Test installation
+          command: |
+            Expand-Archive -LiteralPath build\upload\repository.zip -DestinationPath .
+            build\upload\gpt4all-installer-win64.exe --no-size-checking --default-answer --accept-licenses `
+              --confirm-command --set-temp-repository repository `
+              install gpt4all
 
   build-offline-chat-installer-windows-arm:
     machine:
@@ -745,13 +849,13 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            wget.exe "https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe"
-            # note: need x86_64 toolset for windeployqt.exe, which isn't provided with the ARM64 cross toolset
-            & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations `
+            wget.exe "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-windows-x64-4.8.1.exe"
+            & .\qt-online-installer-windows-x64-4.8.1.exe --no-force-installations --no-default-installations `
               --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations `
-              --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install `
-              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 `
-              qt.qt6.651.win64_msvc2019_arm64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qthttpserver
+              --email "${Env:QT_EMAIL}" --password "${Env:QT_PASSWORD}" install `
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.win64_msvc2022_64 `
+              qt.qt6.682.win64_msvc2022_arm64_cross_compiled qt.qt6.682.addons.qt5compat qt.qt6.682.debug_info `
+              qt.qt6.682.addons.qthttpserver
       - run:
           name: "Install Dotnet 8"
           command: |
@@ -781,14 +885,14 @@ jobs:
             & "C:\Qt\Tools\CMake_64\bin\cmake.exe" `
               -S ..\gpt4all-chat -B . -G Ninja `
               -DCMAKE_BUILD_TYPE=Release `
-              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_arm64" `
+              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.8.2\msvc2022_arm64" `
               "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
-              "-DCMAKE_TOOLCHAIN_FILE=C:\Qt\6.5.1\msvc2019_arm64\lib\cmake\Qt6\qt.toolchain.cmake" `
+              "-DCMAKE_TOOLCHAIN_FILE=C:\Qt\6.8.2\msvc2022_arm64\lib\cmake\Qt6\qt.toolchain.cmake" `
               -DCMAKE_C_COMPILER_LAUNCHER=ccache `
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache `
               -DLLMODEL_CUDA=OFF `
               -DLLMODEL_KOMPUTE=OFF `
-              "-DWINDEPLOYQT=C:\Qt\6.5.1\msvc2019_64\bin\windeployqt.exe;--qtpaths;C:\Qt\6.5.1\msvc2019_arm64\bin\qtpaths.bat" `
+              "-DWINDEPLOYQT=C:\Qt\6.8.2\msvc2022_64\bin\windeployqt.exe;--qtpaths;C:\Qt\6.8.2\msvc2022_arm64\bin\qtpaths.bat" `
               -DGPT4ALL_TEST=OFF `
               -DGPT4ALL_OFFLINE_INSTALLER=ON
             & "C:\Qt\Tools\Ninja\ninja.exe"
@@ -844,6 +948,12 @@ jobs:
             AzureSignTool.exe sign -du "https://gpt4all.io/index.html" -kvu https://gpt4all.vault.azure.net -kvi "$Env:AZSignGUID" -kvs "$Env:AZSignPWD" -kvc "$Env:AZSignCertName" -kvt "$Env:AZSignTID" -tr http://timestamp.digicert.com -v "$($(Get-Location).Path)\build\upload\gpt4all-installer-win64-arm.exe"
       - store_artifacts:
           path: build/upload
+      - run:
+          name: Test installation
+          command: |
+            build\upload\gpt4all-installer-win64-arm.exe --no-size-checking --default-answer --accept-licenses `
+              --confirm-command `
+              install gpt4all
 
   build-online-chat-installer-windows-arm:
     machine:
@@ -866,13 +976,13 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            wget.exe "https://gpt4all.io/ci/qt-unified-windows-x64-4.6.0-online.exe"
-            # note: need x86_64 toolset for windeployqt.exe, which isn't provided with the ARM64 cross toolset
-            & .\qt-unified-windows-x64-4.6.0-online.exe --no-force-installations --no-default-installations `
+            wget.exe "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-windows-x64-4.8.1.exe"
+            & .\qt-online-installer-windows-x64-4.8.1.exe --no-force-installations --no-default-installations `
               --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations `
-              --email ${Env:QT_EMAIL} --password ${Env:QT_PASSWORD} install `
-              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 `
-              qt.qt6.651.win64_msvc2019_arm64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qthttpserver
+              --email "${Env:QT_EMAIL}" --password "${Env:QT_PASSWORD}" install `
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.win64_msvc2022_64 `
+              qt.qt6.682.win64_msvc2022_arm64_cross_compiled qt.qt6.682.addons.qt5compat qt.qt6.682.debug_info `
+              qt.qt6.682.addons.qthttpserver
       - run:
           name: "Install Dotnet 8"
           command: |
@@ -907,14 +1017,14 @@ jobs:
             & "C:\Qt\Tools\CMake_64\bin\cmake.exe" `
               -S ..\gpt4all-chat -B . -G Ninja `
               -DCMAKE_BUILD_TYPE=Release `
-              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_arm64" `
+              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.8.2\msvc2022_arm64" `
               "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
-              "-DCMAKE_TOOLCHAIN_FILE=C:\Qt\6.5.1\msvc2019_arm64\lib\cmake\Qt6\qt.toolchain.cmake" `
+              "-DCMAKE_TOOLCHAIN_FILE=C:\Qt\6.8.2\msvc2022_arm64\lib\cmake\Qt6\qt.toolchain.cmake" `
               -DCMAKE_C_COMPILER_LAUNCHER=ccache `
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache `
               -DLLMODEL_CUDA=OFF `
               -DLLMODEL_KOMPUTE=OFF `
-              "-DWINDEPLOYQT=C:\Qt\6.5.1\msvc2019_64\bin\windeployqt.exe;--qtpaths;C:\Qt\6.5.1\msvc2019_arm64\bin\qtpaths.bat" `
+              "-DWINDEPLOYQT=C:\Qt\6.8.2\msvc2022_64\bin\windeployqt.exe;--qtpaths;C:\Qt\6.8.2\msvc2022_arm64\bin\qtpaths.bat" `
               -DGPT4ALL_TEST=OFF `
               -DGPT4ALL_OFFLINE_INSTALLER=OFF
             & "C:\Qt\Tools\Ninja\ninja.exe"
@@ -977,6 +1087,13 @@ jobs:
             AzureSignTool.exe sign -du "https://gpt4all.io/index.html" -kvu https://gpt4all.vault.azure.net -kvi "$Env:AZSignGUID" -kvs "$Env:AZSignPWD" -kvc "$Env:AZSignCertName" -kvt "$Env:AZSignTID" -tr http://timestamp.digicert.com -v "$($(Get-Location).Path)/build/upload/gpt4all-installer-win64-arm.exe"
       - store_artifacts:
           path: build/upload
+      - run:
+          name: Test installation
+          command: |
+            Expand-Archive -LiteralPath build\upload\repository.zip -DestinationPath .
+            build\upload\gpt4all-installer-win64-arm.exe --no-size-checking --default-answer --accept-licenses `
+              --confirm-command --set-temp-repository repository `
+              install gpt4all
 
   build-gpt4all-chat-linux:
     machine:
@@ -1013,14 +1130,18 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.gcc_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            wget "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-linux-x64-4.8.1.run"
+            chmod +x qt-online-installer-linux-x64-4.8.1.run
+            ./qt-online-installer-linux-x64-4.8.1.run --no-force-installations --no-default-installations \
+              --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations \
+              --email "$QT_EMAIL" --password "$QT_PASSWORD" install \
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.linux_gcc_64 qt.qt6.682.addons.qt5compat \
+              qt.qt6.682.debug_info extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
       - run:
           name: Build
           no_output_timeout: 30m
           command: |
-            export CMAKE_PREFIX_PATH=~/Qt/6.5.1/gcc_64/lib/cmake
+            export CMAKE_PREFIX_PATH=~/Qt/6.8.2/gcc_64/lib/cmake
             export PATH=$PATH:/usr/local/cuda/bin
             ccache -o "cache_dir=${PWD}/../.ccache" -o max_size=500M -p -z
             ~/Qt/Tools/CMake/bin/cmake \
@@ -1061,8 +1182,12 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 ${Env:QT_PASSWORD} install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.win64_msvc2019_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
+            wget.exe "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-windows-x64-4.8.1.exe"
+            & .\qt-online-installer-windows-x64-4.8.1.exe --no-force-installations --no-default-installations `
+              --no-size-checking --default-answer --accept-licenses --confirm-command --accept-obligations `
+              --email "${Env:QT_EMAIL}" --password "${Env:QT_PASSWORD}" install `
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.win64_msvc2022_64 qt.qt6.682.addons.qt5compat `
+              qt.qt6.682.debug_info extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
       - run:
           name: Install VulkanSDK
           command: |
@@ -1087,7 +1212,7 @@ jobs:
             & "C:\Qt\Tools\CMake_64\bin\cmake.exe" `
               -S gpt4all-chat -B build -G Ninja `
               -DCMAKE_BUILD_TYPE=Release `
-              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.5.1\msvc2019_64" `
+              "-DCMAKE_PREFIX_PATH:PATH=C:\Qt\6.8.2\msvc2022_64" `
               "-DCMAKE_MAKE_PROGRAM:FILEPATH=C:\Qt\Tools\Ninja\ninja.exe" `
               -DCMAKE_C_COMPILER_LAUNCHER=ccache `
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache `
@@ -1123,10 +1248,15 @@ jobs:
       - run:
           name: Installing Qt
           command: |
-            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 $QT_PASSWORD install qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.651.clang_64 qt.qt6.651.qt5compat qt.qt6.651.debug_info qt.qt6.651.addons.qtpdf qt.qt6.651.addons.qthttpserver
-            hdiutil detach /Volumes/qt-unified-macOS-x64-4.6.0-online
+            wget "https://qt.mirror.constant.com/archive/online_installers/4.8/qt-online-installer-macOS-x64-4.8.1.dmg"
+            hdiutil attach qt-online-installer-macOS-x64-4.8.1.dmg
+            /Volumes/qt-online-installer-macOS-x64-4.8.1/qt-online-installer-macOS-x64-4.8.1.app/Contents/MacOS/qt-online-installer-macOS-x64-4.8.1 \
+              --no-force-installations --no-default-installations --no-size-checking --default-answer \
+              --accept-licenses --confirm-command --accept-obligations --email "$QT_EMAIL" --password "$QT_PASSWORD" \
+              install \
+              qt.tools.cmake qt.tools.ifw.48 qt.tools.ninja qt.qt6.682.clang_64 qt.qt6.682.addons.qt5compat \
+              extensions.qtpdf.682 qt.qt6.682.addons.qthttpserver
+            hdiutil detach /Volumes/qt-online-installer-macOS-x64-4.8.1
       - run:
           name: Build
           no_output_timeout: 30m
@@ -1135,7 +1265,7 @@ jobs:
             ~/Qt/Tools/CMake/CMake.app/Contents/bin/cmake \
               -S gpt4all-chat -B build -G Ninja \
               -DCMAKE_BUILD_TYPE=Release \
-              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.5.1/macos/lib/cmake \
+              -DCMAKE_PREFIX_PATH:PATH=~/Qt/6.8.2/macos/lib/cmake \
               -DCMAKE_MAKE_PROGRAM:FILEPATH=~/Qt/Tools/Ninja/ninja \
               -DCMAKE_C_COMPILER_LAUNCHER=ccache \
               -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
@@ -1169,7 +1299,7 @@ jobs:
             cd gpt4all-bindings/typescript
             npm run docs:build
 
-  build-py-docs:
+  deploy-docs:
     docker:
       - image: circleci/python:3.8
     steps:
@@ -1373,7 +1503,7 @@ jobs:
           paths:
             - "*.whl"
 
-  publish-wheels:
+  deploy-wheels:
     docker:
       - image: circleci/python:3.8
     steps:
@@ -1670,7 +1800,7 @@ jobs:
             - prebuilds/win32-x64/*.node
             - runtimes/win32-x64/*-*.dll
 
-  prepare-npm-pkg:
+  deploy-npm-pkg:
     docker:
       - image: cimg/base:stable
     steps:
@@ -1726,11 +1856,27 @@ jobs:
             npm set //registry.npmjs.org/:_authToken=$NPM_TOKEN
             npm publish
 
-main_only: &main_only
+# only run a job on the main branch
+job_only_main: &job_only_main
   filters:
     branches:
       only: main
 
+# allow a job to run on tags as well as commits
+job_allow_tags: &job_allow_tags
+  filters:
+    tags:
+      only:
+        - /.*/
+
+# standard chat workflow filter
+workflow-when-chat-requested: &workflow-when-chat-requested
+  when:
+    and:
+      - or: [ << pipeline.parameters.run-all-workflows >>, << pipeline.parameters.run-chat-workflow >> ]
+      - not:
+          equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
+
 workflows:
   version: 2
   noop:
@@ -1741,23 +1887,143 @@ workflows:
           - << pipeline.parameters.run-python-workflow >>
           - << pipeline.parameters.run-ts-workflow >>
           - << pipeline.parameters.run-chat-workflow >>
+          - equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
     jobs:
       - noop
-  build-chat-offline-installers:
+  schedule:
+    # only run when scheduled by CircleCI
     when:
-      or:
-        - << pipeline.parameters.run-all-workflows >>
-        - << pipeline.parameters.run-chat-workflow >>
+      equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
     jobs:
-      - hold:
+      - build-offline-chat-installer-macos:
+          context: gpt4all
+      - build-offline-chat-installer-windows:
+          context: gpt4all
+      - build-offline-chat-installer-windows-arm:
+          context: gpt4all
+      - build-offline-chat-installer-linux:
+          context: gpt4all
+      - sign-online-chat-installer-macos:
+          context: gpt4all
+          requires:
+            - build-online-chat-installer-macos
+      - notarize-online-chat-installer-macos:
+          context: gpt4all
+          requires:
+            - sign-online-chat-installer-macos
+      - sign-online-chat-installer-windows:
+          context: gpt4all
+          requires:
+            - build-online-chat-installer-windows
+      - sign-online-chat-installer-windows-arm:
+          context: gpt4all
+          requires:
+            - build-online-chat-installer-windows-arm
+  build-chat-installers-release:
+    # only run on main branch tags that start with 'v' and a digit
+    when:
+      and:
+        - matches: { pattern: '^v\d.*', value: << pipeline.git.tag >> }
+        - not:
+            equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
+    jobs:
+      - validate-commit-on-main:
+          <<: *job_allow_tags
+      - build-offline-chat-installer-macos:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - build-offline-chat-installer-windows:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - build-offline-chat-installer-windows-arm:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - build-offline-chat-installer-linux:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - sign-offline-chat-installer-macos:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - build-offline-chat-installer-macos
+      - notarize-offline-chat-installer-macos:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - sign-offline-chat-installer-macos
+      - sign-offline-chat-installer-windows:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - build-offline-chat-installer-windows
+      - sign-offline-chat-installer-windows-arm:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - build-offline-chat-installer-windows-arm
+      - build-online-chat-installer-macos:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - build-online-chat-installer-windows:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - build-online-chat-installer-windows-arm:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - build-online-chat-installer-linux:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - validate-commit-on-main
+      - sign-online-chat-installer-macos:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - build-online-chat-installer-macos
+      - notarize-online-chat-installer-macos:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - sign-online-chat-installer-macos
+      - sign-online-chat-installer-windows:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - build-online-chat-installer-windows
+      - sign-online-chat-installer-windows-arm:
+          <<: *job_allow_tags
+          context: gpt4all
+          requires:
+            - build-online-chat-installer-windows-arm
+  build-chat-offline-installers:
+    <<: *workflow-when-chat-requested
+    jobs:
+      - build-hold:
+          type: approval
+      - sign-hold:
           type: approval
       - build-offline-chat-installer-macos:
           context: gpt4all
           requires:
-            - hold
+            - build-hold
       - sign-offline-chat-installer-macos:
           context: gpt4all
           requires:
+            - sign-hold
             - build-offline-chat-installer-macos
       - notarize-offline-chat-installer-macos:
           context: gpt4all
@@ -1766,77 +2032,69 @@ workflows:
       - build-offline-chat-installer-windows:
           context: gpt4all
           requires:
-            - hold
+            - build-hold
       - sign-offline-chat-installer-windows:
           context: gpt4all
           requires:
+            - sign-hold
             - build-offline-chat-installer-windows
       - build-offline-chat-installer-windows-arm:
           context: gpt4all
           requires:
-            - hold
+            - build-hold
       - sign-offline-chat-installer-windows-arm:
           context: gpt4all
           requires:
+            - sign-hold
             - build-offline-chat-installer-windows-arm
       - build-offline-chat-installer-linux:
           context: gpt4all
           requires:
-            - hold
+            - build-hold
   build-chat-online-installers:
-    when:
-      or:
-        - << pipeline.parameters.run-all-workflows >>
-        - << pipeline.parameters.run-chat-workflow >>
+    <<: *workflow-when-chat-requested
     jobs:
-      - hold:
-          <<: *main_only
+      - build-hold:
+          type: approval
+      - sign-hold:
           type: approval
       - build-online-chat-installer-macos:
-          <<: *main_only
           context: gpt4all
           requires:
-            - hold
+            - build-hold
       - sign-online-chat-installer-macos:
-          <<: *main_only
           context: gpt4all
           requires:
+            - sign-hold
             - build-online-chat-installer-macos
       - notarize-online-chat-installer-macos:
-          <<: *main_only
           context: gpt4all
           requires:
             - sign-online-chat-installer-macos
       - build-online-chat-installer-windows:
-          <<: *main_only
           context: gpt4all
           requires:
-            - hold
+            - build-hold
       - sign-online-chat-installer-windows:
-          <<: *main_only
           context: gpt4all
           requires:
+            - sign-hold
             - build-online-chat-installer-windows
       - build-online-chat-installer-windows-arm:
-          <<: *main_only
           context: gpt4all
           requires:
-            - hold
+            - build-hold
       - sign-online-chat-installer-windows-arm:
-          <<: *main_only
           context: gpt4all
           requires:
+            - sign-hold
             - build-online-chat-installer-windows-arm
       - build-online-chat-installer-linux:
-          <<: *main_only
           context: gpt4all
           requires:
-            - hold
+            - build-hold
   build-and-test-gpt4all-chat:
-    when:
-      or:
-        - << pipeline.parameters.run-all-workflows >>
-        - << pipeline.parameters.run-chat-workflow >>
+    <<: *workflow-when-chat-requested
     jobs:
       - hold:
           type: approval
@@ -1854,23 +2112,25 @@ workflows:
             - hold
   deploy-docs:
     when:
-      or:
-        - << pipeline.parameters.run-all-workflows >>
-        - << pipeline.parameters.run-python-workflow >>
+      and:
+        - equal: [ << pipeline.git.branch >>, main ]
+        - or:
+            - << pipeline.parameters.run-all-workflows >>
+            - << pipeline.parameters.run-python-workflow >>
+        - not:
+            equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
     jobs:
-      - build-ts-docs:
-          <<: *main_only
-      - build-py-docs:
-          <<: *main_only
+      - deploy-docs:
           context: gpt4all
   build-python:
     when:
-      or:
-        - << pipeline.parameters.run-all-workflows >>
-        - << pipeline.parameters.run-python-workflow >>
+      and:
+        - or: [ << pipeline.parameters.run-all-workflows >>, << pipeline.parameters.run-python-workflow >> ]
+        - not:
+            equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
     jobs:
       - pypi-hold:
-          <<: *main_only
+          <<: *job_only_main
           type: approval
       - hold:
           type: approval
@@ -1883,8 +2143,8 @@ workflows:
       - build-py-windows:
           requires:
             - hold
-      - publish-wheels:
-          <<: *main_only
+      - deploy-wheels:
+          <<: *job_only_main
           context: gpt4all
           requires:
             - pypi-hold
@@ -1893,36 +2153,29 @@ workflows:
             - build-py-macos
   build-bindings:
     when:
-      or:
-       - << pipeline.parameters.run-all-workflows >>
-       - << pipeline.parameters.run-python-workflow >>
-       - << pipeline.parameters.run-ts-workflow >>
+      and:
+        - or: [ << pipeline.parameters.run-all-workflows >>, << pipeline.parameters.run-ts-workflow >> ]
+        - not:
+            equal: [ << pipeline.trigger_source >>, scheduled_pipeline ]
     jobs:
-      - hold:
+      - backend-hold:
           type: approval
       - nodejs-hold:
           type: approval
       - npm-hold:
-          <<: *main_only
+          <<: *job_only_main
+          type: approval
+      - docs-hold:
           type: approval
       - build-bindings-backend-linux:
           requires:
-            - hold
+            - backend-hold
       - build-bindings-backend-macos:
           requires:
-            - hold
+            - backend-hold
       - build-bindings-backend-windows:
           requires:
-            - hold
-
-      # NodeJs Jobs
-      - prepare-npm-pkg:
-          <<: *main_only
-          requires:
-            - npm-hold
-            - build-nodejs-linux
-            - build-nodejs-windows
-            - build-nodejs-macos
+            - backend-hold
       - build-nodejs-linux:
           requires:
             - nodejs-hold
@@ -1935,3 +2188,13 @@ workflows:
           requires:
             - nodejs-hold
             - build-bindings-backend-macos
+      - build-ts-docs:
+          requires:
+            - docs-hold
+      - deploy-npm-pkg:
+          <<: *job_only_main
+          requires:
+            - npm-hold
+            - build-nodejs-linux
+            - build-nodejs-windows
+            - build-nodejs-macos

.gitmodules

@@ -17,9 +17,9 @@
 [submodule "gpt4all-chat/deps/QXlsx"]
 	path = gpt4all-chat/deps/QXlsx
 	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
+[submodule "gpt4all-chat/deps/minja"]
+	path = gpt4all-chat/deps/minja
+	url = https://github.com/nomic-ai/minja.git
+[submodule "gpt4all-chat/deps/json"]
+	path = gpt4all-chat/deps/json
+	url = https://github.com/nlohmann/json.git

README.md

@@ -1,5 +1,9 @@
 <h1 align="center">GPT4All</h1>
 
+<p align="center">
+  Now with support for DeepSeek R1 Distillations
+</p>
+
 <p align="center">
   <a href="https://www.nomic.ai/gpt4all">Website</a> &bull; <a href="https://docs.gpt4all.io">Documentation</a> &bull; <a href="https://discord.gg/mGZE39AS3e">Discord</a> &bull; <a href="https://www.youtube.com/watch?v=gQcZDXRVJok">YouTube Tutorial</a>
 </p>

gpt4all-backend/src/llamamodel.cpp

@@ -53,6 +53,7 @@ static const std::vector<const char *> KNOWN_ARCHES {
     "gpt2",
     // "gptj", -- no inference code
     "gptneox",
+    "granitemoe",
     "mpt",
     "baichuan",
     "starcoder",
@@ -80,6 +81,7 @@ static const std::vector<const char *> KNOWN_ARCHES {
     "command-r",
     // "dbrx", -- 16x12B parameters
     "olmo",
+    "olmoe",
     "openelm",
     // "arctic", -- 10B+128x3.66B parameters
     "deepseek2",

gpt4all-chat/CHANGELOG.md

@@ -4,11 +4,45 @@ 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/).
 
+## [3.9.0] - 2025-02-04
+
+### Added
+- Whitelist OLMoE and Granite MoE model architectures (no Vulkan) (by [@ThiloteE](https://github.com/ThiloteE) in [#3449](https://github.com/nomic-ai/gpt4all/pull/3449))
+
+### Fixed
+- Fix "index N is not a prompt" when using LocalDocs with reasoning ([#3451](https://github.com/nomic-ai/gpt4all/pull/3451)
+- Work around rendering artifacts on Snapdragon SoCs with Windows ([#3450](https://github.com/nomic-ai/gpt4all/pull/3450))
+- Prevent DeepSeek-R1 reasoning from appearing in chat names and follow-up questions ([#3458](https://github.com/nomic-ai/gpt4all/pull/3458))
+- Fix LocalDocs crash on Windows ARM when reading PDFs ([#3460](https://github.com/nomic-ai/gpt4all/pull/3460))
+- Fix UI freeze when chat template is `{#` ([#3446](https://github.com/nomic-ai/gpt4all/pull/3446))
+
+## [3.8.0] - 2025-01-30
+
+### Added
+- Support DeepSeek-R1 Qwen models ([#3431](https://github.com/nomic-ai/gpt4all/pull/3431))
+- Support for think tags in the GUI ([#3440](https://github.com/nomic-ai/gpt4all/pull/3440))
+- Support specifying SHA256 hash in models3.json instead of MD5 ([#3437](https://github.com/nomic-ai/gpt4all/pull/3437))
+
+### Changed
+- Use minja instead of Jinja2Cpp for significantly improved template compatibility ([#3433](https://github.com/nomic-ai/gpt4all/pull/3433))
+
+### Fixed
+- Fix regression while using localdocs with server API ([#3410](https://github.com/nomic-ai/gpt4all/pull/3410))
+- Don't show system messages in server chat view ([#3411](https://github.com/nomic-ai/gpt4all/pull/3411))
+- Fix `codesign --verify` failure on macOS ([#3413](https://github.com/nomic-ai/gpt4all/pull/3413))
+- Code Interpreter: Fix console.log not accepting a single string after v3.7.0 ([#3426](https://github.com/nomic-ai/gpt4all/pull/3426))
+- Fix Phi 3.1 Mini 128K Instruct template (by [@ThiloteE](https://github.com/ThiloteE) in [#3412](https://github.com/nomic-ai/gpt4all/pull/3412))
+- Don't block the gui thread for reasoning ([#3435](https://github.com/nomic-ai/gpt4all/pull/3435))
+- Fix corruption of unicode in output of reasoning models ([#3443](https://github.com/nomic-ai/gpt4all/pull/3443))
+
 ## [3.7.0] - 2025-01-21
 
 ### Added
 - Add support for the Windows ARM64 target platform (CPU-only) ([#3385](https://github.com/nomic-ai/gpt4all/pull/3385))
 
+### Changed
+- Update from Qt 6.5.1 to 6.8.1 ([#3386](https://github.com/nomic-ai/gpt4all/pull/3386))
+
 ### Fixed
 - Fix the timeout error in code interpreter ([#3369](https://github.com/nomic-ai/gpt4all/pull/3369))
 - Fix code interpreter console.log not accepting multiple arguments ([#3371](https://github.com/nomic-ai/gpt4all/pull/3371))
@@ -17,7 +51,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - Save chats on quit, even if the window isn't closed first ([#3387](https://github.com/nomic-ai/gpt4all/pull/3387))
 - Add chat template replacements for five new models and fix EM German Mistral ([#3393](https://github.com/nomic-ai/gpt4all/pull/3393))
 - Fix crash when entering `{{ a["foo"(` as chat template ([#3394](https://github.com/nomic-ai/gpt4all/pull/3394))
-- Sign the maintenance tool on macOS to prevent crash on Sonoma ([#3391](https://github.com/nomic-ai/gpt4all/pull/3391))
+- Sign the maintenance tool on macOS to prevent crash on Sequoia ([#3391](https://github.com/nomic-ai/gpt4all/pull/3391))
 - Jinja2Cpp: Fix operator precedence in 'not X is defined' ([#3402](https://github.com/nomic-ai/gpt4all/pull/3402))
 
 ## [3.6.1] - 2024-12-20
@@ -261,6 +295,9 @@ 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))
 
+[3.9.0]: https://github.com/nomic-ai/gpt4all/compare/v3.8.0...v3.9.0
+[3.8.0]: https://github.com/nomic-ai/gpt4all/compare/v3.7.0...v3.8.0
+[3.7.0]: https://github.com/nomic-ai/gpt4all/compare/v3.6.1...v3.7.0
 [3.6.1]: https://github.com/nomic-ai/gpt4all/compare/v3.6.0...v3.6.1
 [3.6.0]: https://github.com/nomic-ai/gpt4all/compare/v3.5.3...v3.6.0
 [3.5.3]: https://github.com/nomic-ai/gpt4all/compare/v3.5.2...v3.5.3

gpt4all-chat/CMakeLists.txt

@@ -3,13 +3,17 @@ 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 7)
+set(APP_VERSION_MINOR 9)
 set(APP_VERSION_PATCH 0)
 set(APP_VERSION_BASE "${APP_VERSION_MAJOR}.${APP_VERSION_MINOR}.${APP_VERSION_PATCH}")
 set(APP_VERSION "${APP_VERSION_BASE}")
 
 project(gpt4all VERSION ${APP_VERSION_BASE} LANGUAGES CXX C)
 
+if (CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
+  set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install CACHE PATH "..." FORCE)
+endif()
+
 if(APPLE)
   option(BUILD_UNIVERSAL "Build a Universal binary on macOS" OFF)
   if(BUILD_UNIVERSAL)
@@ -31,6 +35,8 @@ option(GPT4ALL_SIGN_INSTALL "Sign installed binaries and installers (requires si
 option(GPT4ALL_GEN_CPACK_CONFIG "Generate the CPack config.xml in the package step and nothing else." OFF)
 set(GPT4ALL_USE_QTPDF "AUTO" CACHE STRING "Whether to Use QtPDF for LocalDocs. If OFF or not available on this platform, PDFium is used.")
 set_property(CACHE GPT4ALL_USE_QTPDF PROPERTY STRINGS AUTO ON OFF)
+set(GPT4ALL_FORCE_D3D12 "AUTO" CACHE STRING "Whether to use Direct3D 12 as the Qt scene graph backend. Defaults to ON on Windows ARM.")
+set_property(CACHE GPT4ALL_FORCE_D3D12 PROPERTY STRINGS AUTO ON OFF)
 
 include(cmake/cpack_config.cmake)
 
@@ -47,7 +53,7 @@ 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
+    # Enable accurate __cplusplus macro
     add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/Zc:__cplusplus>)
 endif()
 
@@ -86,12 +92,6 @@ include_directories("${CMAKE_CURRENT_BINARY_DIR}")
 set(CMAKE_AUTOMOC ON)
 set(CMAKE_AUTORCC ON)
 
-# Generate a header file with the version number
-configure_file(
-  "${CMAKE_CURRENT_SOURCE_DIR}/cmake/config.h.in"
-  "${CMAKE_CURRENT_BINARY_DIR}/config.h"
-)
-
 set(CMAKE_FIND_PACKAGE_TARGETS_GLOBAL ON)
 set(GPT4ALL_QT_COMPONENTS Core HttpServer LinguistTools Quick QuickDialogs2 Sql Svg)
 set(GPT4ALL_USING_QTPDF OFF)
@@ -126,10 +126,24 @@ 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 ${CMAKE_BINARY_DIR}/install CACHE PATH "..." FORCE)
+set(GPT4ALL_CONFIG_FORCE_D3D12 -1)
+if (NOT CMAKE_SYSTEM_NAME MATCHES Windows OR Qt6_VERSION VERSION_LESS "6.6")
+    # Direct3D 12 is not available.
+    if (GPT4ALL_FORCE_D3D12 STREQUAL "ON")
+        message(FATAL_ERROR "Cannot use Direct3D 12 on this platform.")
+    endif()
+elseif (GPT4ALL_FORCE_D3D12 MATCHES "^(ON|AUTO)$")
+    if (GPT4ALL_FORCE_D3D12 STREQUAL "ON" OR CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|AARCH64|arm64|ARM64)$")
+        set(GPT4ALL_CONFIG_FORCE_D3D12 1)
+    endif()
 endif()
 
+# Generate a header file for configuration
+configure_file(
+    "${CMAKE_CURRENT_SOURCE_DIR}/src/config.h.in"
+    "${CMAKE_CURRENT_BINARY_DIR}/config.h"
+)
+
 add_subdirectory(deps)
 add_subdirectory(../gpt4all-backend llmodel)
 
@@ -437,7 +451,10 @@ else()
     target_link_libraries(chat PRIVATE pdfium)
 endif()
 target_link_libraries(chat
-    PRIVATE llmodel SingleApplication fmt::fmt duckx::duckx QXlsx jinja2cpp)
+    PRIVATE llmodel SingleApplication fmt::fmt duckx::duckx QXlsx)
+target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/json/include)
+target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/json/include/nlohmann)
+target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/minja/include)
 
 if (APPLE)
     target_link_libraries(chat PRIVATE ${COCOA_LIBRARY})
@@ -445,12 +462,18 @@ endif()
 
 # -- install --
 
+if (APPLE)
+    set(GPT4ALL_LIB_DEST bin/gpt4all.app/Contents/Frameworks)
+else()
+    set(GPT4ALL_LIB_DEST lib)
+endif()
+
 install(TARGETS chat DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN})
 
 install(
     TARGETS llmodel
-    LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
-    RUNTIME DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
+    LIBRARY DESTINATION ${GPT4ALL_LIB_DEST} COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
+    RUNTIME DESTINATION bin                 COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
 )
 
 # We should probably iterate through the list of the cmake for backend, but these need to be installed
@@ -473,8 +496,8 @@ endif()
 
 install(
     TARGETS ${MODEL_IMPL_TARGETS}
-    LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
-    RUNTIME DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
+    LIBRARY DESTINATION ${GPT4ALL_LIB_DEST} COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
+    RUNTIME DESTINATION lib                 COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
 )
 
 if(APPLE AND GPT4ALL_SIGN_INSTALL)
@@ -503,7 +526,7 @@ if (LLMODEL_CUDA)
         TARGETS llamamodel-mainline-cuda
                 llamamodel-mainline-cuda-avxonly
         RUNTIME_DEPENDENCY_SET llama-cuda-deps
-        LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
+        LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so
         RUNTIME DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
     )
     if (WIN32)
@@ -520,9 +543,9 @@ endif()
 if (NOT GPT4ALL_USING_QTPDF)
     # Install PDFium
     if (WIN32)
-        install(FILES "${PDFium_LIBRARY}" DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN})
+        install(FILES ${PDFium_LIBRARY} DESTINATION bin                 COMPONENT ${COMPONENT_NAME_MAIN})  # .dll
     else()
-        install(FILES "${PDFium_LIBRARY}" DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN})
+        install(FILES ${PDFium_LIBRARY} DESTINATION ${GPT4ALL_LIB_DEST} COMPONENT ${COMPONENT_NAME_MAIN})  # .so/.dylib
     endif()
 endif()
 

gpt4all-chat/README.md

@@ -1,45 +0,0 @@
-# gpt4all-chat
-
-Cross platform Qt based GUI for GPT4All versions with GPT-J as the base
-model. NOTE: The model seen in the screenshot is actually a preview of a
-new training run for GPT4All based on GPT-J. The GPT4All project is busy
-at work getting ready to release this model including installers for all
-three major OS's. In the meantime, you can try this UI out with the original
-GPT-J model by following build instructions below.
-
-![image](https://user-images.githubusercontent.com/50458173/231464085-da9edff6-a593-410e-8f38-7513f75c8aab.png)
-
-## Install
-
-One click installers for macOS, Linux, and Windows at https://www.nomic.ai/gpt4all
-
-## Features
-
-* Cross-platform (Linux, Windows, MacOSX)
-* The UI is made to look and feel like you've come to expect from a chatty gpt
-* Check for updates so you can always stay fresh with latest models
-* Easy to install with precompiled binaries available for all three major desktop platforms
-* Multi-modal - Ability to load more than one model and switch between them
-* Multi-chat - a list of current and past chats and the ability to save/delete/export and switch between
-* Supports models that are supported by llama.cpp
-* Model downloader in GUI featuring many popular open source models
-* Settings dialog to change temp, top_p, min_p, top_k, threads, etc
-* Copy your conversation to clipboard
-* RAG via LocalDocs feature
-* Check for updates to get the very latest GUI
-
-## Building and running
-
-* Follow the visual instructions on the [build_and_run](build_and_run.md) page
-
-## Getting the latest
-
-If you've already checked out the source code and/or built the program make sure when you do a git fetch to get the latest changes and that you also do `git submodule update --init --recursive` to update the submodules. (If you ever run into trouble, deinitializing via `git submodule deinit -f .` and then initializing again via `git submodule update --init --recursive` fixes most issues)
-
-## Contributing
-
-* Pull requests welcome. See the feature wish list for ideas :)
-
-
-## License
-The source code of this chat interface is currently under a MIT license.

gpt4all-chat/cmake/config.h.in

@@ -1,6 +0,0 @@
-#ifndef CONFIG_H
-#define CONFIG_H
-
-#define APP_VERSION "@APP_VERSION@"
-
-#endif // CONFIG_H

gpt4all-chat/cmake/cpack_config.cmake

@@ -37,7 +37,6 @@ set(CPACK_PACKAGE_VERSION_PATCH ${PROJECT_VERSION_PATCH})
 set(CPACK_PACKAGE_HOMEPAGE_URL "https://www.nomic.ai/gpt4all")
 set(CPACK_PACKAGE_ICON "${CMAKE_CURRENT_SOURCE_DIR}/icons/gpt4all-48.png")
 set(CPACK_RESOURCE_FILE_LICENSE ${CMAKE_CURRENT_SOURCE_DIR}/LICENSE)
-set(CPACK_RESOURCE_FILE_README ${CMAKE_CURRENT_SOURCE_DIR}/README.md)
 set(CPACK_PACKAGE_EXECUTABLES "GPT4All")
 set(CPACK_CREATE_DESKTOP_LINKS "GPT4All")
 set(CPACK_IFW_PACKAGE_NAME "GPT4All")

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

@@ -8,12 +8,6 @@ if (GPT4ALL_SIGN_INSTALL)
     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}
-     DESTINATION ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app/Contents/Frameworks)
-file(COPY ${MYLLMODELLIBS}
-     DESTINATION ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app/Contents/Frameworks)
 file(COPY "${CMAKE_CURRENT_SOURCE_DIR}/icons/gpt4all-32.png"
      DESTINATION ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data)
 file(COPY "${CMAKE_CURRENT_SOURCE_DIR}/icons/gpt4all-48.png"

gpt4all-chat/deps/CMakeLists.txt

@@ -15,43 +15,34 @@ 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)
-
 if (NOT GPT4ALL_USING_QTPDF)
     # If we do not use QtPDF, we need to get PDFium.
-    set(GPT4ALL_PDFIUM_TAG "chromium/6954")
+    set(GPT4ALL_PDFIUM_TAG "chromium/6996")
     if (CMAKE_SYSTEM_NAME MATCHES Linux)
         FetchContent_Declare(
             pdfium
             URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-linux-x64.tgz"
-            URL_HASH "SHA256=69917fd9543befc6c806254aff6c8a604d9e7cd3999a3e70fc32b8690d372da2"
+            URL_HASH "SHA256=68b381b87efed539f2e33ae1e280304c9a42643a878cc296c1d66a93b0cb4335"
         )
     elseif (CMAKE_SYSTEM_NAME MATCHES Windows)
         if (CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64|amd64)$")
             FetchContent_Declare(
                 pdfium
                 URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-win-x64.tgz"
-                URL_HASH "SHA256=62ecac78fbaf658457beaffcc05eb147f493d435a2e1309e6a731808b4e80d38"
+                URL_HASH "SHA256=83e714c302ceacccf403826d5cb57ea39b77f393d83b8d5781283012774a9378"
             )
         elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|AARCH64|arm64|ARM64)$")
             FetchContent_Declare(
                 pdfium
                 URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-win-arm64.tgz"
-                URL_HASH "SHA256=a0b69014467f2b9824776c064920bc95359c9ba0d88793bdda1894a0f22206f8"
+                URL_HASH "SHA256=78e77e871453a4915cbf66fb381b951c9932f88a747c6b2b33c9f27ec2371445"
             )
         endif()
     elseif (CMAKE_SYSTEM_NAME MATCHES Darwin)
         FetchContent_Declare(
             pdfium
             URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-mac-univ.tgz"
-            URL_HASH "SHA256=7442f1dc6bef90898b2b7bd38dbec369ddd81bbf66c1c5aac3a1b60e107098f9"
+            URL_HASH "SHA256=e7577f3242ff9c1df50025f9615673a43601a201bc51ee4792975f98920793a2"
         )
     endif()
 

gpt4all-chat/metadata/latestnews.md

@@ -1,16 +1,16 @@
 ## Latest News
 
-GPT4All v3.6.1 was released on December 20th and fixes issues with the stop generation and copy conversation buttons which were broken in v3.6.0.
+GPT4All v3.8.0 was released on January 30th. Changes include:
 
----
-
-GPT4All v3.6.0 was released on December 19th. Changes include:
-
-* **Reasoner v1:**
-  * Built-in javascript code interpreter tool.
-  * Custom curated model that utilizes the code interpreter to break down, analyze, perform, and verify complex reasoning tasks.
-* **Templates:** Automatically substitute chat templates that are not compatible with Jinja2Cpp in GGUFs.
-* **Fixes:**
-  * Remote model template to allow for XML in messages.
-  * Jinja2Cpp bug that broke system message detection in chat templates.
-  * LocalDocs sources displaying in unconsolidated form after v3.5.0.
+* **Native DeepSeek-R1-Distill Support:** GPT4All now has robust support for the DeepSeek-R1 family of distillations.
+  * Several model variants are now available on the downloads page.
+  * Reasoning (wrapped in "think" tags) is displayed similarly to the Reasoner model.
+  * The DeepSeek-R1 Qwen pretokenizer is now supported, resolving the loading failure in previous versions.
+  * The model is now configured with a GPT4All-compatible prompt template by default.
+* **Chat Templating Overhaul:** The template parser has been *completely* replaced with one that has much better compatibility with common models.
+* **Code Interpreter Fixes:**
+  * An issue preventing the code interpreter from logging a single string in v3.7.0 has been fixed.
+  * The UI no longer freezes while the code interpreter is running a computation.
+* **Local Server Fixes:**
+  * An issue preventing the server from using LocalDocs after the first request since v3.5.0 has been fixed.
+  * System messages are now correctly hidden from the message history.

gpt4all-chat/metadata/models3.json

@@ -32,6 +32,66 @@
     "systemPrompt": "",
     "chatTemplate": "{%- set loop_messages = messages %}\n{%- for message in loop_messages %}\n    {%- set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' %}\n    {{- content }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}"
   },
+  {
+    "order": "aa1",
+    "sha256sum": "5cd4ee65211770f1d99b4f6f4951780b9ef40e29314bd6542bb5bd0ad0bc29d1",
+    "name": "DeepSeek-R1-Distill-Qwen-7B",
+    "filename": "DeepSeek-R1-Distill-Qwen-7B-Q4_0.gguf",
+    "filesize": "4444121056",
+    "requires": "3.8.0",
+    "ramrequired": "8",
+    "parameters": "7 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Qwen2.5-Math-7B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-7B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
+  {
+    "order": "aa2",
+    "sha256sum": "906b3382f2680f4ce845459b4a122e904002b075238080307586bcffcde49eef",
+    "name": "DeepSeek-R1-Distill-Qwen-14B",
+    "filename": "DeepSeek-R1-Distill-Qwen-14B-Q4_0.gguf",
+    "filesize": "8544267680",
+    "requires": "3.8.0",
+    "ramrequired": "16",
+    "parameters": "14 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Qwen2.5-14B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-14B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
+  {
+    "order": "aa3",
+    "sha256sum": "0eb93e436ac8beec18aceb958c120d282cb2cf5451b23185e7be268fe9d375cc",
+    "name": "DeepSeek-R1-Distill-Llama-8B",
+    "filename": "DeepSeek-R1-Distill-Llama-8B-Q4_0.gguf",
+    "filesize": "4675894112",
+    "requires": "3.8.0",
+    "ramrequired": "8",
+    "parameters": "8 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Llama-3.1-8B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF/resolve/main/DeepSeek-R1-Distill-Llama-8B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
+  {
+    "order": "aa4",
+    "sha256sum": "b3af887d0a015b39fab2395e4faf682c1a81a6a3fd09a43f0d4292f7d94bf4d0",
+    "name": "DeepSeek-R1-Distill-Qwen-1.5B",
+    "filename": "DeepSeek-R1-Distill-Qwen-1.5B-Q4_0.gguf",
+    "filesize": "1068807776",
+    "requires": "3.8.0",
+    "ramrequired": "3",
+    "parameters": "1.5 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Qwen2.5-Math-1.5B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-1.5B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-1.5B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
   {
     "order": "b",
     "md5sum": "27b44e8ae1817525164ddf4f8dae8af4",
@@ -472,7 +532,7 @@
     "filename": "qwen2-1_5b-instruct-q4_0.gguf",
     "filesize": "937532800",
     "requires": "3.0",
-    "ramrequired": "4",
+    "ramrequired": "3",
     "parameters": "1.5 billion",
     "quant": "q4_0",
     "type": "qwen2",

gpt4all-chat/metadata/release.json

@@ -251,12 +251,22 @@
   },
   {
     "version": "3.6.0",
-    "notes": "* **Reasoner v1:**\n  * Built-in javascript code interpreter tool.\n  * Custom curated model that utilizes the code interpreter to break down, analyze, perform, and verify complex reasoning tasks.\n* **Templates:** Automatically substitute chat templates that are not compatible with Jinja2Cpp in GGUFs.\n* **Fixes:**\n  * Remote model template to allow for XML in messages.\n  * Jinja2Cpp bug that broke system message detection in chat templates.\n  * LocalDocs sources displaying in unconsolidated form after v3.5.0.",
+    "notes": "* **Reasoner v1:**\n  * Built-in javascript code interpreter tool.\n  * Custom curated model that utilizes the code interpreter to break down, analyze, perform, and verify complex reasoning tasks.\n* **Templates:** Automatically substitute chat templates that are not compatible with Jinja2Cpp in GGUFs.\n* **Fixes:**\n  * Remote model template to allow for XML in messages.\n  * Jinja2Cpp bug that broke system message detection in chat templates.\n  * LocalDocs sources displaying in unconsolidated form after v3.5.0.\n",
     "contributors": "* Adam Treat (Nomic AI)\n* Jared Van Bortel (Nomic AI)"
   },
   {
     "version": "3.6.1",
-    "notes": "* **Fixes:**\n  * The stop generation button no longer working in v3.6.0.\n  * The copy entire conversation button no longer working in v3.6.0.",
+    "notes": "* **Fixes:**\n  * The stop generation button no longer working in v3.6.0.\n  * The copy entire conversation button no longer working in v3.6.0.\n",
     "contributors": "* Adam Treat (Nomic AI)"
+  },
+  {
+    "version": "3.7.0",
+    "notes": "* **Windows ARM Support:** GPT4All now supports the Windows ARM platform, ensuring compatibility with devices powered by Qualcomm Snapdragon and Microsoft SQ-series processors.\n  * **NOTE:** Support for GPU and/or NPU acceleration is not available at this time. Only the CPU will be used to run LLMs.\n  * **NOTE:** You must install the new *Windows ARM* version of GPT4All from the website. The standard *Windows* version will not work due to emulation limitations.\n* **Fixed Updating on macOS:** The maintenance tool no longer crashes when attempting to update or uninstall GPT4All on Sequoia.\n  * **NOTE:** If you have installed the version from the GitHub releases as a workaround for this issue, you can safely uninstall it and switch back to the version from the website.\n* **Fixed Chat Saving on macOS:** Chats now save as expected when the application is quit with Command-Q.\n* **Code Interpreter Improvements:**\n  * The behavior when the code takes too long to execute and times out has been improved.\n  * console.log now accepts multiple arguments for better compatibility with native JavaScript.\n* **Chat Templating Improvements:**\n  * Two crashes and one compatibility issue have been fixed in the chat template parser.\n  * The default chat template for EM German Mistral has been fixed.\n  * Automatic replacements have been added for five new models as we continue to improve compatibility with common chat templates.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* Riccardo Giovanetti (`@Harvester62`)"
+  },
+  {
+    "version": "3.8.0",
+    "notes": "* **Native DeepSeek-R1-Distill Support:** GPT4All now has robust support for the DeepSeek-R1 family of distillations.\n  * Several model variants are now available on the downloads page.\n  * Reasoning (wrapped in \"think\" tags) is displayed similarly to the Reasoner model.\n  * The DeepSeek-R1 Qwen pretokenizer is now supported, resolving the loading failure in previous versions.\n  * The model is now configured with a GPT4All-compatible prompt template by default.\n* **Chat Templating Overhaul:** The template parser has been *completely* replaced with one that has much better compatibility with common models.\n* **Code Interpreter Fixes:**\n  * An issue preventing the code interpreter from logging a single string in v3.7.0 has been fixed.\n  * The UI no longer freezes while the code interpreter is running a computation.\n* **Local Server Fixes:**\n  * An issue preventing the server from using LocalDocs after the first request since v3.5.0 has been fixed.\n  * System messages are now correctly hidden from the message history.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* ThiloteE (`@ThiloteE`)"
   }
 ]

gpt4all-chat/qml/ChatCollapsibleItem.qml

@@ -13,6 +13,8 @@ ColumnLayout {
     property alias textContent: innerTextItem.textContent
     property bool isCurrent: false
     property bool isError: false
+    property bool isThinking: false
+    property int  thinkingTime: 0
 
     Layout.topMargin: 10
     Layout.bottomMargin: 10
@@ -26,16 +28,20 @@ ColumnLayout {
             anchors.bottom: parent.bottom
 
             Item {
-                width: myTextArea.width
-                height: myTextArea.height
+                Layout.preferredWidth: myTextArea.implicitWidth
+                Layout.preferredHeight: myTextArea.implicitHeight
                 TextArea {
                     id: myTextArea
                     text: {
                         if (isError)
                             return qsTr("Analysis encountered error");
                         if (isCurrent)
-                            return qsTr("Analyzing");
-                        return qsTr("Analyzed");
+                            return isThinking ? qsTr("Thinking") : qsTr("Analyzing");
+                        return isThinking
+                            ? qsTr("Thought for %1 %2")
+                                  .arg(Math.ceil(thinkingTime / 1000.0))
+                                  .arg(Math.ceil(thinkingTime / 1000.0) === 1 ? qsTr("second") : qsTr("seconds"))
+                            : qsTr("Analyzed");
                     }
                     padding: 0
                     font.pixelSize: theme.fontSizeLarger

gpt4all-chat/qml/ChatItemView.qml

@@ -110,6 +110,7 @@ GridLayout {
                         case Chat.PromptProcessing: return qsTr("processing ...")
                         case Chat.ResponseGeneration: return qsTr("generating response ...");
                         case Chat.GeneratingQuestions: return qsTr("generating questions ...");
+                        case Chat.ToolCallGeneration: return qsTr("generating toolcall ...");
                         default: return ""; // handle unexpected values
                         }
                     }
@@ -188,6 +189,17 @@ GridLayout {
                         isError: modelData.isToolCallError
                     }
                 }
+                DelegateChoice {
+                    roleValue: "Think: ";
+                    ChatCollapsibleItem {
+                        Layout.fillWidth: true
+                        textContent: modelData.content
+                        isCurrent: modelData.isCurrentResponse
+                        isError: false
+                        isThinking: true
+                        thinkingTime: modelData.thinkingTime
+                    }
+                }
             }
 
             delegate: chooser

gpt4all-chat/qml/ChatView.qml

@@ -828,7 +828,7 @@ Rectangle {
                                     textInput.cursorPosition = text.length;
                                 }
                                 height: visible ? implicitHeight : 0
-                                visible: name !== "ToolResponse: "
+                                visible: name !== "ToolResponse: " && name !== "System: "
                             }
 
                             remove: Transition {

gpt4all-chat/src/chat.cpp

@@ -181,6 +181,11 @@ QVariant Chat::popPrompt(int index)
 
 void Chat::stopGenerating()
 {
+    // In future if we have more than one tool we'll have to keep track of which tools are possibly
+    // running, but for now we only have one
+    Tool *toolInstance = ToolModel::globalInstance()->get(ToolCallConstants::CodeInterpreterFunction);
+    Q_ASSERT(toolInstance);
+    toolInstance->interrupt();
     m_llmodel->stopGenerating();
 }
 
@@ -242,56 +247,71 @@ void Chat::responseStopped(qint64 promptResponseMs)
 
     const QString possibleToolcall = m_chatModel->possibleToolcall();
 
-    ToolCallParser parser;
-    parser.update(possibleToolcall);
-
-    if (parser.state() == ToolEnums::ParseState::Complete) {
-        const QString toolCall = parser.toolCall();
-
-        // Regex to remove the formatting around the code
-        static const QRegularExpression regex("^\\s*```javascript\\s*|\\s*```\\s*$");
-        QString code = toolCall;
-        code.remove(regex);
-        code = code.trimmed();
-
-        // Right now the code interpreter is the only available tool
-        Tool *toolInstance = ToolModel::globalInstance()->get(ToolCallConstants::CodeInterpreterFunction);
-        Q_ASSERT(toolInstance);
-
-        // The param is the code
-        const ToolParam param = { "code", ToolEnums::ParamType::String, code };
-        const QString result = toolInstance->run({param}, 10000 /*msecs to timeout*/);
-        const ToolEnums::Error error = toolInstance->error();
-        const QString errorString = toolInstance->errorString();
-
-        // Update the current response with meta information about toolcall and re-parent
-        m_chatModel->updateToolCall({
-            ToolCallConstants::CodeInterpreterFunction,
-            { param },
-            result,
-            error,
-            errorString
-        });
-
-        ++m_consecutiveToolCalls;
-
-        // We limit the number of consecutive toolcalls otherwise we get into a potentially endless loop
-        if (m_consecutiveToolCalls < 3 || error == ToolEnums::Error::NoError) {
-            resetResponseState();
-            emit promptRequested(m_collections); // triggers a new response
-            return;
-        }
-    }
-
-    if (m_generatedName.isEmpty())
-        emit generateNameRequested();
-
-    m_consecutiveToolCalls = 0;
-    Network::globalInstance()->trackChatEvent("response_complete", {
+    Network::globalInstance()->trackChatEvent("response_stopped", {
         {"first", m_firstResponse},
         {"message_count", chatModel()->count()},
         {"$duration", promptResponseMs / 1000.},
     });
+
+    ToolCallParser parser;
+    parser.update(possibleToolcall.toUtf8());
+    if (parser.state() == ToolEnums::ParseState::Complete && parser.startTag() != ToolCallConstants::ThinkStartTag)
+        processToolCall(parser.toolCall());
+    else
+        responseComplete();
+}
+
+void Chat::processToolCall(const QString &toolCall)
+{
+    m_responseState = Chat::ToolCallGeneration;
+    emit responseStateChanged();
+    // Regex to remove the formatting around the code
+    static const QRegularExpression regex("^\\s*```javascript\\s*|\\s*```\\s*$");
+    QString code = toolCall;
+    code.remove(regex);
+    code = code.trimmed();
+
+    // Right now the code interpreter is the only available tool
+    Tool *toolInstance = ToolModel::globalInstance()->get(ToolCallConstants::CodeInterpreterFunction);
+    Q_ASSERT(toolInstance);
+    connect(toolInstance, &Tool::runComplete, this, &Chat::toolCallComplete, Qt::SingleShotConnection);
+
+    // The param is the code
+    const ToolParam param = { "code", ToolEnums::ParamType::String, code };
+    m_responseInProgress = true;
+    emit responseInProgressChanged();
+    toolInstance->run({param});
+}
+
+void Chat::toolCallComplete(const ToolCallInfo &info)
+{
+    // Update the current response with meta information about toolcall and re-parent
+    m_chatModel->updateToolCall(info);
+
+    ++m_consecutiveToolCalls;
+
+    m_responseInProgress = false;
+    emit responseInProgressChanged();
+
+    // We limit the number of consecutive toolcalls otherwise we get into a potentially endless loop
+    if (m_consecutiveToolCalls < 3 || info.error == ToolEnums::Error::NoError) {
+        resetResponseState();
+        emit promptRequested(m_collections); // triggers a new response
+        return;
+    }
+
+    responseComplete();
+}
+
+void Chat::responseComplete()
+{
+    if (m_generatedName.isEmpty())
+        emit generateNameRequested();
+
+    m_responseState = Chat::ResponseStopped;
+    emit responseStateChanged();
+
+    m_consecutiveToolCalls = 0;
     m_firstResponse = false;
 }
 
@@ -361,11 +381,8 @@ void Chat::trySwitchContextOfLoadedModel()
 
 void Chat::generatedNameChanged(const QString &name)
 {
-    // Only use the first three words maximum and remove newlines and extra spaces
-    m_generatedName = name.simplified();
-    QStringList words = m_generatedName.split(' ', Qt::SkipEmptyParts);
-    int wordCount = qMin(7, words.size());
-    m_name = words.mid(0, wordCount).join(' ');
+    m_generatedName = name;
+    m_name = name;
     emit nameChanged();
     m_needsSave = true;
 }

gpt4all-chat/src/chat.h

@@ -55,7 +55,8 @@ public:
         LocalDocsProcessing,
         PromptProcessing,
         GeneratingQuestions,
-        ResponseGeneration
+        ResponseGeneration,
+        ToolCallGeneration
     };
     Q_ENUM(ResponseState)
 
@@ -166,6 +167,9 @@ private Q_SLOTS:
     void promptProcessing();
     void generatingQuestions();
     void responseStopped(qint64 promptResponseMs);
+    void processToolCall(const QString &toolCall);
+    void toolCallComplete(const ToolCallInfo &info);
+    void responseComplete();
     void generatedNameChanged(const QString &name);
     void generatedQuestionFinished(const QString &question);
     void handleModelLoadingError(const QString &error);

gpt4all-chat/src/chatllm.cpp

@@ -12,12 +12,8 @@
 #include "toolcallparser.h"
 
 #include <fmt/format.h>
-
-#include <jinja2cpp/error_info.h>
-#include <jinja2cpp/template.h>
-#include <jinja2cpp/template_env.h>
-#include <jinja2cpp/user_callable.h>
-#include <jinja2cpp/value.h>
+#include <minja/minja.hpp>
+#include <nlohmann/json.hpp>
 
 #include <QDataStream>
 #include <QDebug>
@@ -33,6 +29,7 @@
 #include <QRegularExpression>
 #include <QRegularExpressionMatch>
 #include <QSet>
+#include <QTextStream>
 #include <QUrl>
 #include <QWaitCondition>
 #include <Qt>
@@ -60,44 +57,104 @@
 using namespace Qt::Literals::StringLiterals;
 using namespace ToolEnums;
 namespace ranges = std::ranges;
+using json = nlohmann::ordered_json;
 
 //#define DEBUG
 //#define DEBUG_MODEL_LOADING
 
 // NOTE: not threadsafe
-static jinja2::TemplateEnv *jinjaEnv()
+static const std::shared_ptr<minja::Context> &jinjaEnv()
 {
-    static std::optional<jinja2::TemplateEnv> environment;
+    static std::shared_ptr<minja::Context> environment;
     if (!environment) {
-        auto &env = environment.emplace();
-        auto &settings = env.GetSettings();
-        settings.trimBlocks = true;
-        settings.lstripBlocks = true;
-        env.AddGlobal("raise_exception", jinja2::UserCallable(
-            /*callable*/ [](auto &params) -> jinja2::Value {
-                auto messageArg = params.args.find("message");
-                if (messageArg == params.args.end() || !messageArg->second.isString())
-                    throw std::runtime_error("'message' argument to raise_exception() must be a string");
-                throw std::runtime_error(fmt::format("Jinja template error: {}", messageArg->second.asString()));
-            },
-            /*argsInfo*/ { jinja2::ArgInfo("message", /*isMandatory*/ true) }
-        ));
-        env.AddGlobal("strftime_now", jinja2::UserCallable(
-            /*callable*/ [](auto &params) -> jinja2::Value {
+        environment = minja::Context::builtins();
+        environment->set("strftime_now", minja::simple_function(
+            "strftime_now", { "format" },
+            [](const std::shared_ptr<minja::Context> &, minja::Value &args) -> minja::Value {
+                auto format = args.at("format").get<std::string>();
                 using Clock = std::chrono::system_clock;
-                auto formatArg = params.args.find("format");
-                if (formatArg == params.args.end() || !formatArg->second.isString())
-                    throw std::runtime_error("'format' argument to strftime_now() must be a string");
                 time_t nowUnix = Clock::to_time_t(Clock::now());
                 auto localDate = *std::localtime(&nowUnix);
                 std::ostringstream ss;
-                ss << std::put_time(&localDate, formatArg->second.asString().c_str());
+                ss << std::put_time(&localDate, format.c_str());
                 return ss.str();
-            },
-            /*argsInfo*/ { jinja2::ArgInfo("format", /*isMandatory*/ true) }
+            }
+        ));
+        environment->set("regex_replace", minja::simple_function(
+            "regex_replace", { "str", "pattern", "repl" },
+            [](const std::shared_ptr<minja::Context> &, minja::Value &args) -> minja::Value {
+                auto str     = args.at("str"    ).get<std::string>();
+                auto pattern = args.at("pattern").get<std::string>();
+                auto repl    = args.at("repl"   ).get<std::string>();
+                return std::regex_replace(str, std::regex(pattern), repl);
+            }
         ));
     }
-    return &*environment;
+    return environment;
+}
+
+class BaseResponseHandler {
+public:
+    virtual void onSplitIntoTwo    (const QString &startTag, const QString &firstBuffer, const QString &secondBuffer) = 0;
+    virtual void onSplitIntoThree  (const QString &secondBuffer, const QString &thirdBuffer)                          = 0;
+    // "old-style" responses, with all of the implementation details left in
+    virtual void onOldResponseChunk(const QByteArray &chunk)                                                          = 0;
+    // notify of a "new-style" response that has been cleaned of tool calling
+    virtual bool onBufferResponse  (const QString &response, int bufferIdx)                                           = 0;
+    // notify of a "new-style" response, no tool calling applicable
+    virtual bool onRegularResponse ()                                                                                 = 0;
+    virtual bool getStopGenerating () const                                                                           = 0;
+};
+
+static auto promptModelWithTools(
+    LLModel *model, const LLModel::PromptCallback &promptCallback, BaseResponseHandler &respHandler,
+    const LLModel::PromptContext &ctx, const QByteArray &prompt, const QStringList &toolNames
+) -> std::pair<QStringList, bool>
+{
+    ToolCallParser toolCallParser(toolNames);
+    auto handleResponse = [&toolCallParser, &respHandler](LLModel::Token token, std::string_view piece) -> bool {
+        Q_UNUSED(token)
+
+        toolCallParser.update(piece.data());
+
+        // Split the response into two if needed
+        if (toolCallParser.numberOfBuffers() < 2 && toolCallParser.splitIfPossible()) {
+            const auto parseBuffers = toolCallParser.buffers();
+            Q_ASSERT(parseBuffers.size() == 2);
+            respHandler.onSplitIntoTwo(toolCallParser.startTag(), parseBuffers.at(0), parseBuffers.at(1));
+        }
+
+        // Split the response into three if needed
+        if (toolCallParser.numberOfBuffers() < 3 && toolCallParser.startTag() == ToolCallConstants::ThinkStartTag
+            && toolCallParser.splitIfPossible()) {
+            const auto parseBuffers = toolCallParser.buffers();
+            Q_ASSERT(parseBuffers.size() == 3);
+            respHandler.onSplitIntoThree(parseBuffers.at(1), parseBuffers.at(2));
+        }
+
+        respHandler.onOldResponseChunk(QByteArray::fromRawData(piece.data(), piece.size()));
+
+        bool ok;
+        const auto parseBuffers = toolCallParser.buffers();
+        if (parseBuffers.size() > 1) {
+            ok = respHandler.onBufferResponse(parseBuffers.last(), parseBuffers.size() - 1);
+        } else {
+            ok = respHandler.onRegularResponse();
+        }
+        if (!ok)
+            return false;
+
+        const bool shouldExecuteToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete
+            && toolCallParser.startTag() != ToolCallConstants::ThinkStartTag;
+
+        return !shouldExecuteToolCall && !respHandler.getStopGenerating();
+    };
+    model->prompt(std::string_view(prompt), promptCallback, handleResponse, ctx);
+
+    const bool shouldExecuteToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete
+        && toolCallParser.startTag() != ToolCallConstants::ThinkStartTag;
+
+    return { toolCallParser.buffers(), shouldExecuteToolCall };
 }
 
 class LLModelStore {
@@ -738,7 +795,8 @@ std::vector<MessageItem> ChatLLM::forkConversation(const QString &prompt) const
         conversation.reserve(items.size() + 1);
         conversation.assign(items.begin(), items.end());
     }
-    conversation.emplace_back(MessageItem::Type::Prompt, prompt.toUtf8());
+    qsizetype nextIndex = conversation.empty() ? 0 : conversation.back().index().value() + 1;
+    conversation.emplace_back(nextIndex, MessageItem::Type::Prompt, prompt.toUtf8());
     return conversation;
 }
 
@@ -757,19 +815,18 @@ static uint parseJinjaTemplateVersion(QStringView tmpl)
     return 0;
 }
 
-static auto loadJinjaTemplate(
-    std::optional<jinja2::Template> &tmpl /*out*/, const std::string &source
-) -> jinja2::Result<void>
+static std::shared_ptr<minja::TemplateNode> loadJinjaTemplate(const std::string &source)
 {
-    tmpl.emplace(jinjaEnv());
-    return tmpl->Load(source);
+    return minja::Parser::parse(source, { .trim_blocks = true, .lstrip_blocks = true, .keep_trailing_newline = false });
 }
 
 std::optional<std::string> ChatLLM::checkJinjaTemplateError(const std::string &source)
 {
-    std::optional<jinja2::Template> tmpl;
-    if (auto res = loadJinjaTemplate(tmpl, source); !res)
-        return res.error().ToString();
+    try {
+        loadJinjaTemplate(source);
+    } catch (const std::runtime_error &e) {
+        return e.what();
+    }
     return std::nullopt;
 }
 
@@ -801,29 +858,29 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
     uint version = parseJinjaTemplateVersion(chatTemplate);
 
     auto makeMap = [version](const MessageItem &item) {
-        return jinja2::GenericMap([msg = std::make_shared<JinjaMessage>(version, item)] { return msg.get(); });
+        return JinjaMessage(version, item).AsJson();
     };
 
     std::unique_ptr<MessageItem> systemItem;
     bool useSystem = !isAllSpace(systemMessage);
 
-    jinja2::ValuesList messages;
+    json::array_t messages;
     messages.reserve(useSystem + items.size());
     if (useSystem) {
-        systemItem = std::make_unique<MessageItem>(MessageItem::Type::System, systemMessage.toUtf8());
+        systemItem = std::make_unique<MessageItem>(MessageItem::system_tag, systemMessage.toUtf8());
         messages.emplace_back(makeMap(*systemItem));
     }
     for (auto &item : items)
         messages.emplace_back(makeMap(item));
 
-    jinja2::ValuesList toolList;
+    json::array_t toolList;
     const int toolCount = ToolModel::globalInstance()->count();
     for (int i = 0; i < toolCount; ++i) {
         Tool *t = ToolModel::globalInstance()->get(i);
         toolList.push_back(t->jinjaValue());
     }
 
-    jinja2::ValuesMap params {
+    json::object_t params {
         { "messages",              std::move(messages) },
         { "add_generation_prompt", true                },
         { "toolList",              toolList            },
@@ -831,12 +888,14 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
     for (auto &[name, token] : model->specialTokens())
         params.emplace(std::move(name), std::move(token));
 
-    std::optional<jinja2::Template> tmpl;
-    auto maybeRendered = loadJinjaTemplate(tmpl, chatTemplate.toStdString())
-        .and_then([&] { return tmpl->RenderAsString(params); });
-    if (!maybeRendered)
-        throw std::runtime_error(fmt::format("Failed to parse chat template: {}", maybeRendered.error().ToString()));
-    return *maybeRendered;
+    try {
+        auto tmpl = loadJinjaTemplate(chatTemplate.toStdString());
+        auto context = minja::Context::make(minja::Value(std::move(params)), jinjaEnv());
+        return tmpl->render(context);
+    } catch (const std::runtime_error &e) {
+        throw std::runtime_error(fmt::format("Failed to parse chat template: {}", e.what()));
+    }
+    Q_UNREACHABLE();
 }
 
 auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const LLModel::PromptContext &ctx,
@@ -862,7 +921,7 @@ auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const LL
             // Find the prompt that represents the query. Server chats are flexible and may not have one.
             auto items = getChat();
             if (auto peer = m_chatModel->getPeer(items, items.end() - 1)) // peer of response
-                query = { *peer - items.begin(), (*peer)->content() };
+                query = { (*peer)->index().value(), (*peer)->content() };
         }
 
         if (query) {
@@ -888,6 +947,62 @@ auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const LL
     };
 }
 
+class ChatViewResponseHandler : public BaseResponseHandler {
+public:
+    ChatViewResponseHandler(ChatLLM *cllm, QElapsedTimer *totalTime, ChatLLM::PromptResult *result)
+        : m_cllm(cllm), m_totalTime(totalTime), m_result(result) {}
+
+    void onSplitIntoTwo(const QString &startTag, const QString &firstBuffer, const QString &secondBuffer) override
+    {
+        if (startTag == ToolCallConstants::ThinkStartTag)
+            m_cllm->m_chatModel->splitThinking({ firstBuffer, secondBuffer });
+        else
+            m_cllm->m_chatModel->splitToolCall({ firstBuffer, secondBuffer });
+    }
+
+    void onSplitIntoThree(const QString &secondBuffer, const QString &thirdBuffer) override
+    {
+        m_cllm->m_chatModel->endThinking({ secondBuffer, thirdBuffer }, m_totalTime->elapsed());
+    }
+
+    void onOldResponseChunk(const QByteArray &chunk) override
+    {
+        m_result->responseTokens++;
+        m_cllm->m_timer->inc();
+        m_result->response.append(chunk);
+    }
+
+    bool onBufferResponse(const QString &response, int bufferIdx) override
+    {
+        Q_UNUSED(bufferIdx)
+        try {
+            m_cllm->m_chatModel->setResponseValue(response);
+        } catch (const std::exception &e) {
+            // We have a try/catch here because the main thread might have removed the response from
+            // the chatmodel by erasing the conversation during the response... the main thread sets
+            // m_stopGenerating before doing so, but it doesn't wait after that to reset the chatmodel
+            Q_ASSERT(m_cllm->m_stopGenerating);
+            return false;
+        }
+        emit m_cllm->responseChanged();
+        return true;
+    }
+
+    bool onRegularResponse() override
+    {
+        auto respStr = QString::fromUtf8(m_result->response);
+        return onBufferResponse(removeLeadingWhitespace(respStr), 0);
+    }
+
+    bool getStopGenerating() const override
+    { return m_cllm->m_stopGenerating; }
+
+private:
+    ChatLLM               *m_cllm;
+    QElapsedTimer         *m_totalTime;
+    ChatLLM::PromptResult *m_result;
+};
+
 auto ChatLLM::promptInternal(
     const std::variant<std::span<const MessageItem>, std::string_view> &prompt,
     const LLModel::PromptContext &ctx,
@@ -935,53 +1050,22 @@ auto ChatLLM::promptInternal(
         return !m_stopGenerating;
     };
 
-    ToolCallParser toolCallParser;
-    auto handleResponse = [this, &result, &toolCallParser](LLModel::Token token, std::string_view piece) -> bool {
-        Q_UNUSED(token)
-        result.responseTokens++;
-        m_timer->inc();
-
-        // FIXME: This is *not* necessarily fully formed utf data because it can be partial at this point
-        // handle this like below where we have a QByteArray
-        toolCallParser.update(QString::fromStdString(piece.data()));
-
-        // Create a toolcall and split the response if needed
-        if (!toolCallParser.hasSplit() && toolCallParser.state() == ToolEnums::ParseState::Partial) {
-            const QPair<QString, QString> pair = toolCallParser.split();
-            m_chatModel->splitToolCall(pair);
-        }
-
-        result.response.append(piece.data(), piece.size());
-        auto respStr = QString::fromUtf8(result.response);
-
-        try {
-            if (toolCallParser.hasSplit())
-                m_chatModel->setResponseValue(toolCallParser.buffer());
-            else
-                m_chatModel->setResponseValue(removeLeadingWhitespace(respStr));
-        } catch (const std::exception &e) {
-            // We have a try/catch here because the main thread might have removed the response from
-            // the chatmodel by erasing the conversation during the response... the main thread sets
-            // m_stopGenerating before doing so, but it doesn't wait after that to reset the chatmodel
-            Q_ASSERT(m_stopGenerating);
-            return false;
-        }
-
-        emit responseChanged();
-
-        const bool foundToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete;
-        return !foundToolCall && !m_stopGenerating;
-    };
-
     QElapsedTimer totalTime;
     totalTime.start();
-    m_timer->start();
+    ChatViewResponseHandler respHandler(this, &totalTime, &result);
 
+    m_timer->start();
+    QStringList finalBuffers;
+    bool        shouldExecuteTool;
     try {
         emit promptProcessing();
         m_llModelInfo.model->setThreadCount(mySettings->threadCount());
         m_stopGenerating = false;
-        m_llModelInfo.model->prompt(conversation, handlePrompt, handleResponse, ctx);
+        std::tie(finalBuffers, shouldExecuteTool) = promptModelWithTools(
+            m_llModelInfo.model.get(), handlePrompt, respHandler, ctx,
+            QByteArray::fromRawData(conversation.data(), conversation.size()),
+            ToolCallConstants::AllTagNames
+        );
     } catch (...) {
         m_timer->stop();
         throw;
@@ -990,20 +1074,18 @@ auto ChatLLM::promptInternal(
     m_timer->stop();
     qint64 elapsed = totalTime.elapsed();
 
-    const bool foundToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete;
-
     // trim trailing whitespace
     auto respStr = QString::fromUtf8(result.response);
-    if (!respStr.isEmpty() && (std::as_const(respStr).back().isSpace() || foundToolCall)) {
-        if (toolCallParser.hasSplit())
-            m_chatModel->setResponseValue(toolCallParser.buffer());
+    if (!respStr.isEmpty() && (std::as_const(respStr).back().isSpace() || finalBuffers.size() > 1)) {
+        if (finalBuffers.size() > 1)
+            m_chatModel->setResponseValue(finalBuffers.last());
         else
             m_chatModel->setResponseValue(respStr.trimmed());
         emit responseChanged();
     }
 
     bool doQuestions = false;
-    if (!m_isServer && messageItems && !foundToolCall) {
+    if (!m_isServer && messageItems && !shouldExecuteTool) {
         switch (mySettings->suggestionMode()) {
             case SuggestionMode::On:            doQuestions = true;          break;
             case SuggestionMode::LocalDocsOnly: doQuestions = usedLocalDocs; break;
@@ -1081,6 +1163,66 @@ void ChatLLM::reloadModel()
         loadModel(m);
 }
 
+// This class throws discards the text within thinking tags, for use with chat names and follow-up questions.
+class SimpleResponseHandler : public BaseResponseHandler {
+public:
+    SimpleResponseHandler(ChatLLM *cllm)
+        : m_cllm(cllm) {}
+
+    void onSplitIntoTwo(const QString &startTag, const QString &firstBuffer, const QString &secondBuffer) override
+    { /* no-op */ }
+
+    void onSplitIntoThree(const QString &secondBuffer, const QString &thirdBuffer) override
+    { /* no-op */ }
+
+    void onOldResponseChunk(const QByteArray &chunk) override
+    { m_response.append(chunk); }
+
+    bool onBufferResponse(const QString &response, int bufferIdx) override
+    {
+        if (bufferIdx == 1)
+            return true; // ignore "think" content
+        return onSimpleResponse(response);
+    }
+
+    bool onRegularResponse() override
+    { return onBufferResponse(QString::fromUtf8(m_response), 0); }
+
+    bool getStopGenerating() const override
+    { return m_cllm->m_stopGenerating; }
+
+protected:
+    virtual bool onSimpleResponse(const QString &response) = 0;
+
+protected:
+    ChatLLM    *m_cllm;
+    QByteArray  m_response;
+};
+
+class NameResponseHandler : public SimpleResponseHandler {
+private:
+    // max length of chat names, in words
+    static constexpr qsizetype MAX_WORDS = 3;
+
+public:
+    using SimpleResponseHandler::SimpleResponseHandler;
+
+protected:
+    bool onSimpleResponse(const QString &response) override
+    {
+        QTextStream stream(const_cast<QString *>(&response), QIODeviceBase::ReadOnly);
+        QStringList words;
+        while (!stream.atEnd() && words.size() < MAX_WORDS) {
+            QString word;
+            stream >> word;
+            words << word;
+        }
+
+        emit m_cllm->generatedNameChanged(words.join(u' '));
+        return words.size() < MAX_WORDS || stream.atEnd();
+    }
+};
+
 void ChatLLM::generateName()
 {
     Q_ASSERT(isModelLoaded());
@@ -1097,23 +1239,15 @@ void ChatLLM::generateName()
         return;
     }
 
-    QByteArray response; // raw UTF-8
-
-    auto handleResponse = [this, &response](LLModel::Token token, std::string_view piece) -> bool {
-        Q_UNUSED(token)
-
-        response.append(piece.data(), piece.size());
-        QStringList words = QString::fromUtf8(response).simplified().split(u' ', Qt::SkipEmptyParts);
-        emit generatedNameChanged(words.join(u' '));
-        return words.size() <= 3;
-    };
+    NameResponseHandler respHandler(this);
 
     try {
-        m_llModelInfo.model->prompt(
-            applyJinjaTemplate(forkConversation(chatNamePrompt)),
-            [this](auto &&...) { return !m_stopGenerating; },
-            handleResponse,
-            promptContextFromSettings(m_modelInfo)
+        promptModelWithTools(
+            m_llModelInfo.model.get(),
+            /*promptCallback*/ [this](auto &&...) { return !m_stopGenerating; },
+            respHandler, promptContextFromSettings(m_modelInfo),
+            applyJinjaTemplate(forkConversation(chatNamePrompt)).c_str(),
+            { ToolCallConstants::ThinkTagName }
         );
     } catch (const std::exception &e) {
         qWarning() << "ChatLLM failed to generate name:" << e.what();
@@ -1125,13 +1259,43 @@ void ChatLLM::handleChatIdChanged(const QString &id)
     m_llmThread.setObjectName(id);
 }
 
-void ChatLLM::generateQuestions(qint64 elapsed)
-{
+class QuestionResponseHandler : public SimpleResponseHandler {
+public:
+    using SimpleResponseHandler::SimpleResponseHandler;
+
+protected:
+    bool onSimpleResponse(const QString &response) override
+    {
+        auto responseUtf8Bytes = response.toUtf8().slice(m_offset);
+        auto responseUtf8 = std::string(responseUtf8Bytes.begin(), responseUtf8Bytes.end());
+        // extract all questions from response
+        ptrdiff_t lastMatchEnd = -1;
+        auto it = std::sregex_iterator(responseUtf8.begin(), responseUtf8.end(), s_reQuestion);
+        auto end = std::sregex_iterator();
+        for (; it != end; ++it) {
+            auto pos = it->position();
+            auto len = it->length();
+            lastMatchEnd = pos + len;
+            emit m_cllm->generatedQuestionFinished(QString::fromUtf8(&responseUtf8[pos], len));
+        }
+
+        // remove processed input from buffer
+        if (lastMatchEnd != -1)
+            m_offset += lastMatchEnd;
+        return true;
+    }
+
+private:
     // FIXME: This only works with response by the model in english which is not ideal for a multi-language
     // model.
     // match whole question sentences
-    static const std::regex reQuestion(R"(\b(?:What|Where|How|Why|When|Who|Which|Whose|Whom)\b[^?]*\?)");
+    static inline const std::regex s_reQuestion { R"(\b(?:What|Where|How|Why|When|Who|Which|Whose|Whom)\b[^?]*\?)" };
 
+    qsizetype m_offset = 0;
+};
+
+void ChatLLM::generateQuestions(qint64 elapsed)
+{
     Q_ASSERT(isModelLoaded());
     if (!isModelLoaded()) {
         emit responseStopped(elapsed);
@@ -1149,39 +1313,17 @@ void ChatLLM::generateQuestions(qint64 elapsed)
 
     emit generatingQuestions();
 
-    std::string response; // raw UTF-8
-
-    auto handleResponse = [this, &response](LLModel::Token token, std::string_view piece) -> bool {
-        Q_UNUSED(token)
-
-        // add token to buffer
-        response.append(piece);
-
-        // extract all questions from response
-        ptrdiff_t lastMatchEnd = -1;
-        auto it = std::sregex_iterator(response.begin(), response.end(), reQuestion);
-        auto end = std::sregex_iterator();
-        for (; it != end; ++it) {
-            auto pos = it->position();
-            auto len = it->length();
-            lastMatchEnd = pos + len;
-            emit generatedQuestionFinished(QString::fromUtf8(&response[pos], len));
-        }
-
-        // remove processed input from buffer
-        if (lastMatchEnd != -1)
-            response.erase(0, lastMatchEnd);
-        return true;
-    };
+    QuestionResponseHandler respHandler(this);
 
     QElapsedTimer totalTime;
     totalTime.start();
     try {
-        m_llModelInfo.model->prompt(
-            applyJinjaTemplate(forkConversation(suggestedFollowUpPrompt)),
-            [this](auto &&...) { return !m_stopGenerating; },
-            handleResponse,
-            promptContextFromSettings(m_modelInfo)
+        promptModelWithTools(
+            m_llModelInfo.model.get(),
+            /*promptCallback*/ [this](auto &&...) { return !m_stopGenerating; },
+            respHandler, promptContextFromSettings(m_modelInfo),
+            applyJinjaTemplate(forkConversation(suggestedFollowUpPrompt)).c_str(),
+            { ToolCallConstants::ThinkTagName }
         );
     } catch (const std::exception &e) {
         qWarning() << "ChatLLM failed to generate follow-up questions:" << e.what();

gpt4all-chat/src/chatllm.h

@@ -30,6 +30,7 @@
 
 using namespace Qt::Literals::StringLiterals;
 
+class ChatViewResponseHandler;
 class QDataStream;
 
 // NOTE: values serialized to disk, do not change or reuse
@@ -285,6 +286,8 @@ private:
     bool m_isServer;
     bool m_forceMetal;
     bool m_reloadingToChangeVariant;
+    friend class ChatViewResponseHandler;
+    friend class SimpleResponseHandler;
 };
 
 #endif // CHATLLM_H

gpt4all-chat/src/chatmodel.cpp

@@ -41,6 +41,12 @@ void ChatItem::serializeText(QDataStream &stream, int version)
     stream << value;
 }
 
+void ChatItem::serializeThink(QDataStream &stream, int version)
+{
+    stream << value;
+    stream << thinkingTime;
+}
+
 void ChatItem::serializeSubItems(QDataStream &stream, int version)
 {
     stream << name;
@@ -50,6 +56,7 @@ void ChatItem::serializeSubItems(QDataStream &stream, int version)
         case ToolCall:      { serializeToolCall(stream, version);       break; }
         case ToolResponse:  { serializeToolResponse(stream, version);   break; }
         case Text:          { serializeText(stream, version);           break; }
+        case Think:         { serializeThink(stream, version);          break; }
         case System:
         case Prompt:
             throw std::invalid_argument(fmt::format("cannot serialize subitem type {}", int(typ)));
@@ -162,6 +169,13 @@ bool ChatItem::deserializeResponse(QDataStream &stream, int version)
     return true;
 }
 
+bool ChatItem::deserializeThink(QDataStream &stream, int version)
+{
+    stream >> value;
+    stream >> thinkingTime;
+    return true;
+}
+
 bool ChatItem::deserializeSubItems(QDataStream &stream, int version)
 {
     stream >> name;
@@ -177,6 +191,7 @@ bool ChatItem::deserializeSubItems(QDataStream &stream, int version)
         case ToolCall:      { deserializeToolCall(stream, version); break; }
         case ToolResponse:  { deserializeToolResponse(stream, version); break; }
         case Text:          { deserializeText(stream, version); break; }
+        case Think:         { deserializeThink(stream, version); break; }
         case System:
         case Prompt:
             throw std::invalid_argument(fmt::format("cannot serialize subitem type {}", int(typ)));

gpt4all-chat/src/chatmodel.h

@@ -94,11 +94,28 @@ class MessageItem
 public:
     enum class Type { System, Prompt, Response, ToolResponse };
 
-    MessageItem(Type type, QString content)
-        : m_type(type), m_content(std::move(content)) {}
+    struct system_tag_t { explicit system_tag_t() = default; };
+    static inline constexpr system_tag_t system_tag = system_tag_t{};
 
-    MessageItem(Type type, QString content, const QList<ResultInfo> &sources, const QList<PromptAttachment> &promptAttachments)
-        : m_type(type), m_content(std::move(content)), m_sources(sources), m_promptAttachments(promptAttachments) {}
+    MessageItem(qsizetype index, Type type, QString content)
+        : m_index(index), m_type(type), m_content(std::move(content))
+    {
+        Q_ASSERT(type != Type::System); // use system_tag constructor
+    }
+
+    // Construct a system message with no index, since they are never stored in the chat
+    MessageItem(system_tag_t, QString content)
+        : m_type(Type::System), m_content(std::move(content)) {}
+
+    MessageItem(qsizetype index, Type type, QString content, const QList<ResultInfo> &sources, const QList<PromptAttachment> &promptAttachments)
+        : m_index(index)
+        , m_type(type)
+        , m_content(std::move(content))
+        , m_sources(sources)
+        , m_promptAttachments(promptAttachments) {}
+
+    // index of the parent ChatItem (system, prompt, response) in its container
+    std::optional<qsizetype> index()   const { return m_index; }
 
     Type           type()    const { return m_type;    }
     const QString &content() const { return m_content; }
@@ -126,10 +143,11 @@ public:
     }
 
 private:
-    Type                    m_type;
-    QString                 m_content;
-    QList<ResultInfo>       m_sources;
-    QList<PromptAttachment> m_promptAttachments;
+    std::optional<qsizetype> m_index;
+    Type                     m_type;
+    QString                  m_content;
+    QList<ResultInfo>        m_sources;
+    QList<PromptAttachment>  m_promptAttachments;
 };
 Q_DECLARE_METATYPE(MessageItem)
 
@@ -159,8 +177,11 @@ class ChatItem : public QObject
     Q_PROPERTY(bool    thumbsUpState   MEMBER thumbsUpState  )
     Q_PROPERTY(bool    thumbsDownState MEMBER thumbsDownState)
 
+    // thinking
+    Q_PROPERTY(int     thinkingTime    MEMBER thinkingTime NOTIFY thinkingTimeChanged)
+
 public:
-    enum class Type { System, Prompt, Response, Text, ToolCall, ToolResponse };
+    enum class Type { System, Prompt, Response, Text, ToolCall, ToolResponse, Think };
 
     // tags for constructing ChatItems
     struct prompt_tag_t        { explicit prompt_tag_t       () = default; };
@@ -169,12 +190,14 @@ public:
     struct text_tag_t          { explicit text_tag_t         () = default; };
     struct tool_call_tag_t     { explicit tool_call_tag_t    () = default; };
     struct tool_response_tag_t { explicit tool_response_tag_t() = default; };
+    struct think_tag_t         { explicit think_tag_t        () = default; };
     static inline constexpr prompt_tag_t        prompt_tag        = prompt_tag_t        {};
     static inline constexpr response_tag_t      response_tag      = response_tag_t      {};
     static inline constexpr system_tag_t        system_tag        = system_tag_t        {};
     static inline constexpr text_tag_t          text_tag          = text_tag_t          {};
     static inline constexpr tool_call_tag_t     tool_call_tag     = tool_call_tag_t     {};
     static inline constexpr tool_response_tag_t tool_response_tag = tool_response_tag_t {};
+    static inline constexpr think_tag_t         think_tag         = think_tag_t         {};
 
 public:
     ChatItem(QObject *parent)
@@ -220,6 +243,10 @@ public:
         : ChatItem(parent)
     { this->name = u"ToolResponse: "_s; this->value = value; }
 
+    ChatItem(QObject *parent, think_tag_t, const QString &value)
+        : ChatItem(parent)
+    { this->name = u"Think: "_s; this->value = value; }
+
     Type type() const
     {
         if (name == u"System: "_s)
@@ -234,6 +261,8 @@ public:
             return Type::ToolCall;
         if (name == u"ToolResponse: "_s)
             return Type::ToolResponse;
+        if (name == u"Think: "_s)
+            return Type::Think;
         throw std::invalid_argument(fmt::format("Chat item has unknown label: {:?}", name));
     }
 
@@ -254,7 +283,7 @@ public:
         if (type() == Type::Response) {
             // We parse if this contains any part of a partial toolcall
             ToolCallParser parser;
-            parser.update(value);
+            parser.update(value.toUtf8());
 
             // If no tool call is detected, return the original value
             if (parser.startIndex() < 0)
@@ -265,9 +294,11 @@ public:
             return beforeToolCall;
         }
 
-        // For tool calls we only return content if it is the code interpreter
+        if (type() == Type::Think)
+            return thinkContent(value);
+
         if (type() == Type::ToolCall)
-            return codeInterpreterContent(value);
+            return toolCallContent(value);
 
         // We don't show any of content from the tool response in the GUI
         if (type() == Type::ToolResponse)
@@ -276,10 +307,21 @@ public:
         return value;
     }
 
-    QString codeInterpreterContent(const QString &value) const
+    QString thinkContent(const QString &value) const
     {
         ToolCallParser parser;
-        parser.update(value);
+        parser.update(value.toUtf8());
+
+        // Extract the content
+        QString content = parser.toolCall();
+        content = content.trimmed();
+        return content;
+    }
+
+    QString toolCallContent(const QString &value) const
+    {
+        ToolCallParser parser;
+        parser.update(value.toUtf8());
 
         // Extract the code
         QString code = parser.toolCall();
@@ -357,6 +399,12 @@ public:
         return toolCallInfo.error != ToolEnums::Error::NoError;
     }
 
+    void setThinkingTime(int t)
+    {
+        thinkingTime = t;
+        emit thinkingTimeChanged();
+    }
+
     // NB: Assumes response is not current.
     static ChatItem *fromMessageInput(QObject *parent, const MessageInput &message)
     {
@@ -369,7 +417,7 @@ public:
         Q_UNREACHABLE();
     }
 
-    MessageItem asMessageItem() const
+    MessageItem asMessageItem(qsizetype index) const
     {
         MessageItem::Type msgType;
         switch (auto typ = type()) {
@@ -380,9 +428,10 @@ public:
             case ToolResponse: msgType = MessageItem::Type::ToolResponse; break;
             case Text:
             case ToolCall:
+            case Think:
                 throw std::invalid_argument(fmt::format("cannot convert ChatItem type {} to message item", int(typ)));
         }
-        return { msgType, flattenedContent(), sources, promptAttachments };
+        return { index, msgType, flattenedContent(), sources, promptAttachments };
     }
 
     static QList<ResultInfo> consolidateSources(const QList<ResultInfo> &sources);
@@ -391,6 +440,7 @@ public:
     void serializeToolCall(QDataStream &stream, int version);
     void serializeToolResponse(QDataStream &stream, int version);
     void serializeText(QDataStream &stream, int version);
+    void serializeThink(QDataStream &stream, int version);
     void serializeSubItems(QDataStream &stream, int version); // recursive
     void serialize(QDataStream &stream, int version);
 
@@ -399,6 +449,7 @@ public:
     bool deserializeToolCall(QDataStream &stream, int version);
     bool deserializeToolResponse(QDataStream &stream, int version);
     bool deserializeText(QDataStream &stream, int version);
+    bool deserializeThink(QDataStream &stream, int version);
     bool deserializeSubItems(QDataStream &stream, int version); // recursive
     bool deserialize(QDataStream &stream, int version);
 
@@ -406,6 +457,7 @@ Q_SIGNALS:
     void contentChanged();
     void isTooCallErrorChanged();
     void isCurrentResponseChanged();
+    void thinkingTimeChanged();
 
 public:
 
@@ -429,6 +481,9 @@ public:
     bool    stopped         = false;
     bool    thumbsUpState   = false;
     bool    thumbsDownState = false;
+
+    // thinking time in ms
+    int     thinkingTime  = 0;
 };
 
 class ChatModel : public QAbstractListModel
@@ -500,6 +555,7 @@ private:
         return std::nullopt;
     }
 
+    // FIXME(jared): this should really be done at the parent level, not the sub-item level
     static std::optional<qsizetype> getPeerInternal(const MessageItem *arr, qsizetype size, qsizetype index)
     {
         qsizetype peer;
@@ -879,6 +935,70 @@ public:
         if (changed) emit dataChanged(createIndex(index, 0), createIndex(index, 0), {NewResponseRole});
     }
 
+    Q_INVOKABLE void splitThinking(const QPair<QString, QString> &split)
+    {
+        qsizetype index;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (m_chatItems.isEmpty() || m_chatItems.cend()[-1]->type() != ChatItem::Type::Response)
+                throw std::logic_error("can only set thinking on a chat that ends with a response");
+
+            index = m_chatItems.count() - 1;
+            ChatItem *currentResponse = m_chatItems.back();
+            Q_ASSERT(currentResponse->isCurrentResponse);
+
+            // Create a new response container for any text and the thinking
+            ChatItem *newResponse = new ChatItem(this, ChatItem::response_tag);
+
+            // Add preceding text if any
+            if (!split.first.isEmpty()) {
+                ChatItem *textItem = new ChatItem(this, ChatItem::text_tag, split.first);
+                newResponse->subItems.push_back(textItem);
+            }
+
+            // Add the thinking item
+            Q_ASSERT(!split.second.isEmpty());
+            ChatItem *thinkingItem = new ChatItem(this, ChatItem::think_tag, split.second);
+            thinkingItem->isCurrentResponse = true;
+            newResponse->subItems.push_back(thinkingItem);
+
+            // Add new response and reset our value
+            currentResponse->subItems.push_back(newResponse);
+            currentResponse->value = QString();
+        }
+
+        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {ChildItemsRole, ContentRole});
+    }
+
+    Q_INVOKABLE void endThinking(const QPair<QString, QString> &split, int thinkingTime)
+    {
+        qsizetype index;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (m_chatItems.isEmpty() || m_chatItems.cend()[-1]->type() != ChatItem::Type::Response)
+                throw std::logic_error("can only end thinking on a chat that ends with a response");
+
+            index = m_chatItems.count() - 1;
+            ChatItem *currentResponse = m_chatItems.back();
+            Q_ASSERT(currentResponse->isCurrentResponse);
+
+            ChatItem *subResponse = currentResponse->subItems.back();
+            Q_ASSERT(subResponse->type() == ChatItem::Type::Response);
+            Q_ASSERT(subResponse->isCurrentResponse);
+            subResponse->setCurrentResponse(false);
+
+            ChatItem *thinkingItem = subResponse->subItems.back();
+            Q_ASSERT(thinkingItem->type() == ChatItem::Type::Think);
+            thinkingItem->setCurrentResponse(false);
+            thinkingItem->setValue(split.first);
+            thinkingItem->setThinkingTime(thinkingTime);
+
+            currentResponse->setValue(split.second);
+        }
+
+        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {ChildItemsRole, ContentRole});
+    }
+
     Q_INVOKABLE void splitToolCall(const QPair<QString, QString> &split)
     {
         qsizetype index;
@@ -1013,10 +1133,12 @@ public:
         // A flattened version of the chat item tree used by the backend and jinja
         QMutexLocker locker(&m_mutex);
         std::vector<MessageItem> chatItems;
-        for (const ChatItem *item : m_chatItems) {
-            chatItems.reserve(chatItems.size() + item->subItems.size() + 1);
-            ranges::copy(item->subItems | views::transform(&ChatItem::asMessageItem), std::back_inserter(chatItems));
-            chatItems.push_back(item->asMessageItem());
+        for (qsizetype i : views::iota(0, m_chatItems.size())) {
+            auto *parent = m_chatItems.at(i);
+            chatItems.reserve(chatItems.size() + parent->subItems.size() + 1);
+            ranges::copy(parent->subItems | views::transform([&](auto *s) { return s->asMessageItem(i); }),
+                         std::back_inserter(chatItems));
+            chatItems.push_back(parent->asMessageItem(i));
         }
         return chatItems;
     }

gpt4all-chat/src/codeinterpreter.cpp

@@ -7,8 +7,15 @@
 
 using namespace Qt::Literals::StringLiterals;
 
+CodeInterpreter::CodeInterpreter()
+    : Tool()
+    , m_error(ToolEnums::Error::NoError)
+{
+    m_worker = new CodeInterpreterWorker;
+    connect(this, &CodeInterpreter::request, m_worker, &CodeInterpreterWorker::request, Qt::QueuedConnection);
+}
 
-QString CodeInterpreter::run(const QList<ToolParam> &params, qint64 timeout)
+void CodeInterpreter::run(const QList<ToolParam> &params)
 {
     m_error = ToolEnums::Error::NoError;
     m_errorString = QString();
@@ -18,27 +25,24 @@ QString CodeInterpreter::run(const QList<ToolParam> &params, qint64 timeout)
           && params.first().type == ToolEnums::ParamType::String);
 
     const QString code = params.first().value.toString();
-
-    QThread workerThread;
-    CodeInterpreterWorker worker;
-    worker.moveToThread(&workerThread);
-    connect(&worker, &CodeInterpreterWorker::finished, &workerThread, &QThread::quit, Qt::DirectConnection);
-    connect(&workerThread, &QThread::started, [&worker, code]() {
-        worker.request(code);
+    connect(m_worker, &CodeInterpreterWorker::finished, [this, params] {
+        m_error = m_worker->error();
+        m_errorString = m_worker->errorString();
+        emit runComplete({
+        ToolCallConstants::CodeInterpreterFunction,
+            params,
+            m_worker->response(),
+            m_error,
+            m_errorString
+        });
     });
-    workerThread.start();
-    bool timedOut = !workerThread.wait(timeout);
-    if (timedOut) {
-        worker.interrupt(timeout); // thread safe
-        m_error = ToolEnums::Error::TimeoutError;
-    }
-    workerThread.quit();
-    workerThread.wait();
-    if (!timedOut) {
-        m_error = worker.error();
-        m_errorString = worker.errorString();
-    }
-    return worker.response();
+
+    emit request(code);
+}
+
+bool CodeInterpreter::interrupt()
+{
+    return m_worker->interrupt();
 }
 
 QList<ToolParamInfo> CodeInterpreter::parameters() const
@@ -89,20 +93,20 @@ QString CodeInterpreter::exampleReply() const
 
 CodeInterpreterWorker::CodeInterpreterWorker()
     : QObject(nullptr)
+    , m_engine(new QJSEngine(this))
 {
-}
+    moveToThread(&m_thread);
 
-void CodeInterpreterWorker::request(const QString &code)
-{
-    JavaScriptConsoleCapture consoleCapture;
-    QJSValue consoleInternalObject = m_engine.newQObject(&consoleCapture);
-    m_engine.globalObject().setProperty("console_internal", consoleInternalObject);
+    QJSValue consoleInternalObject = m_engine->newQObject(&m_consoleCapture);
+    m_engine->globalObject().setProperty("console_internal", consoleInternalObject);
 
     // preprocess console.log args in JS since Q_INVOKE doesn't support varargs
-    auto consoleObject = m_engine.evaluate(uR"(
+    auto consoleObject = m_engine->evaluate(uR"(
         class Console {
             log(...args) {
-                if (args.length && typeof args[0] === 'string')
+                if (args.length == 0)
+                    return;
+                if (args.length >= 2 && typeof args[0] === 'string')
                     throw new Error('console.log string formatting not supported');
                 let cat = '';
                 for (const arg of args) {
@@ -114,15 +118,28 @@ void CodeInterpreterWorker::request(const QString &code)
 
         new Console();
     )"_s);
-    m_engine.globalObject().setProperty("console", consoleObject);
+    m_engine->globalObject().setProperty("console", consoleObject);
+    m_thread.start();
+}
 
-    const QJSValue result = m_engine.evaluate(code);
+void CodeInterpreterWorker::reset()
+{
+    m_response.clear();
+    m_error = ToolEnums::Error::NoError;
+    m_errorString.clear();
+    m_consoleCapture.output.clear();
+    m_engine->setInterrupted(false);
+}
 
+void CodeInterpreterWorker::request(const QString &code)
+{
+    reset();
+    const QJSValue result = m_engine->evaluate(code);
     QString resultString;
 
-    if (m_engine.isInterrupted()) {
-        resultString = QString("Error: code execution was timed out as it exceeded %1 ms. Code must be written to ensure execution does not timeout.").arg(m_timeout);
-    } else if (result.isError()) {
+    if (m_engine->isInterrupted()) {
+        resultString = QString("Error: code execution was interrupted or timed out.");
+   } else if (result.isError()) {
         // NOTE: We purposely do not set the m_error or m_errorString for the code interpreter since
         // we *want* the model to see the response has an error so it can hopefully correct itself. The
         // error member variables are intended for tools that have error conditions that cannot be corrected.
@@ -143,9 +160,16 @@ void CodeInterpreterWorker::request(const QString &code)
     }
 
     if (resultString.isEmpty())
-        resultString = consoleCapture.output;
-    else if (!consoleCapture.output.isEmpty())
-        resultString += "\n" + consoleCapture.output;
+        resultString = m_consoleCapture.output;
+    else if (!m_consoleCapture.output.isEmpty())
+        resultString += "\n" + m_consoleCapture.output;
     m_response = resultString;
     emit finished();
 }
+
+bool CodeInterpreterWorker::interrupt()
+{
+    m_error = ToolEnums::Error::TimeoutError;
+    m_engine->setInterrupted(true);
+    return true;
+}

gpt4all-chat/src/codeinterpreter.h

@@ -8,6 +8,7 @@
 #include <QObject>
 #include <QString>
 #include <QtGlobal>
+#include <QThread>
 
 class JavaScriptConsoleCapture : public QObject
 {
@@ -39,32 +40,37 @@ public:
     CodeInterpreterWorker();
     virtual ~CodeInterpreterWorker() {}
 
+    void reset();
     QString response() const { return m_response; }
-
-    void request(const QString &code);
-    void interrupt(qint64 timeout) { m_timeout = timeout; m_engine.setInterrupted(true); }
     ToolEnums::Error error() const { return m_error; }
     QString errorString() const { return m_errorString; }
+    bool interrupt();
+
+public Q_SLOTS:
+    void request(const QString &code);
 
 Q_SIGNALS:
     void finished();
 
 private:
-    qint64 m_timeout = 0;
-    QJSEngine m_engine;
     QString m_response;
     ToolEnums::Error m_error = ToolEnums::Error::NoError;
     QString m_errorString;
+    QThread m_thread;
+    JavaScriptConsoleCapture m_consoleCapture;
+    QJSEngine *m_engine = nullptr;
 };
 
 class CodeInterpreter : public Tool
 {
     Q_OBJECT
 public:
-    explicit CodeInterpreter() : Tool(), m_error(ToolEnums::Error::NoError) {}
+    explicit CodeInterpreter();
     virtual ~CodeInterpreter() {}
 
-    QString run(const QList<ToolParam> &params, qint64 timeout = 2000) override;
+    void run(const QList<ToolParam> &params) override;
+    bool interrupt() override;
+
     ToolEnums::Error error() const override { return m_error; }
     QString errorString() const override { return m_errorString; }
 
@@ -77,9 +83,13 @@ public:
     QString exampleCall() const override;
     QString exampleReply() const override;
 
+Q_SIGNALS:
+    void request(const QString &code);
+
 private:
     ToolEnums::Error m_error = ToolEnums::Error::NoError;
     QString m_errorString;
+    CodeInterpreterWorker *m_worker;
 };
 
 #endif // CODEINTERPRETER_H

gpt4all-chat/src/config.h.in

@@ -0,0 +1,7 @@
+#pragma once
+
+#define APP_VERSION "@APP_VERSION@"
+
+#define G4A_CONFIG(name) (1/G4A_CONFIG_##name == 1)
+
+#define G4A_CONFIG_force_d3d12 @GPT4ALL_CONFIG_FORCE_D3D12@

gpt4all-chat/src/database.cpp

@@ -1209,7 +1209,6 @@ public:
             FPDF_ClosePage(m_page);
         if (m_doc)
             FPDF_CloseDocument(m_doc);
-        FPDF_DestroyLibrary();
     }
 
     int page() const override { return m_currentPage; }
@@ -1224,7 +1223,7 @@ private:
                     return std::nullopt;
 
                 if (m_page)
-                    FPDF_ClosePage(m_page);
+                    FPDF_ClosePage(std::exchange(m_page, nullptr));
                 m_page = FPDF_LoadPage(m_doc, m_currentPage++);
                 if (!m_page)
                     throw std::runtime_error("Failed to load page.");

gpt4all-chat/src/jinja_helpers.cpp

@@ -7,111 +7,75 @@
 #include <QString>
 #include <QUrl>
 
-#include <memory>
+#include <iterator>
+#include <map>
+#include <ranges>
 #include <vector>
 
-using namespace std::literals::string_view_literals;
+namespace views  = std::views;
+using json = nlohmann::ordered_json;
 
 
-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
+json::object_t JinjaResultInfo::AsJson() const
 {
-    std::vector<std::string> result;
-    auto &keys = this->keys();
-    result.reserve(keys.size());
-    result.assign(keys.begin(), keys.end());
-    return result;
+    return {
+        { "collection", m_source->collection.toStdString() },
+        { "path",       m_source->path      .toStdString() },
+        { "file",       m_source->file      .toStdString() },
+        { "title",      m_source->title     .toStdString() },
+        { "author",     m_source->author    .toStdString() },
+        { "date",       m_source->date      .toStdString() },
+        { "text",       m_source->text      .toStdString() },
+        { "page",       m_source->page                     },
+        { "file_uri",   m_source->fileUri() .toStdString() },
+    };
 }
 
-auto JinjaMessage::keys() const -> const std::unordered_set<std::string_view> &
+json::object_t JinjaPromptAttachment::AsJson() const
 {
-    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 MessageItem::Type;
-    case System:
-    case Response:
-    case ToolResponse:
-        return baseKeys;
-    case Prompt:
-        return userKeys;
-        break;
-    }
-    Q_UNREACHABLE();
+    return {
+        { "url",               m_attachment->url.toString()    .toStdString() },
+        { "file",              m_attachment->file()            .toStdString() },
+        { "processed_content", m_attachment->processedContent().toStdString() },
+    };
 }
 
-bool operator==(const JinjaMessage &a, const JinjaMessage &b)
+json::object_t JinjaMessage::AsJson() const
 {
-    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.content() != ib.content())
-        return false;
-
-    switch (type) {
-        using enum MessageItem::Type;
-    case System:
-    case Response:
-    case ToolResponse:
-        return true;
-    case Prompt:
-        return ia.sources() == ib.sources() && ia.promptAttachments() == ib.promptAttachments();
-        break;
-    }
-    Q_UNREACHABLE();
-}
-
-const JinjaFieldMap<JinjaMessage> JinjaMessage::s_fields = {
-    { "role", [](auto &m) {
-        switch (m.item().type()) {
+    json::object_t obj;
+    {
+        json::string_t role;
+        switch (m_item->type()) {
             using enum MessageItem::Type;
-            case System:   return "system"sv;
-            case Prompt:   return "user"sv;
-            case Response: return "assistant"sv;
-            case ToolResponse: return "tool"sv;
-                break;
+            case System:       role = "system";    break;
+            case Prompt:       role = "user";      break;
+            case Response:     role = "assistant"; break;
+            case ToolResponse: role = "tool";      break;
         }
-        Q_UNREACHABLE();
-    } },
-    { "content", [](auto &m) {
-        if (m.version() == 0 && m.item().type() == MessageItem::Type::Prompt)
-            return m.item().bakedPrompt().toStdString();
-        return m.item().content().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());
-    } },
-};
+        obj.emplace_back("role", std::move(role));
+    }
+    {
+        QString content;
+        if (m_version == 0 && m_item->type() == MessageItem::Type::Prompt) {
+            content = m_item->bakedPrompt();
+        } else {
+            content = m_item->content();
+        }
+        obj.emplace_back("content", content.toStdString());
+    }
+    if (m_item->type() == MessageItem::Type::Prompt) {
+        {
+            auto sources = m_item->sources() | views::transform([](auto &r) {
+                return JinjaResultInfo(r).AsJson();
+            });
+            obj.emplace("sources", json::array_t(sources.begin(), sources.end()));
+        }
+        {
+            auto attachments = m_item->promptAttachments() | views::transform([](auto &pa) {
+                return JinjaPromptAttachment(pa).AsJson();
+            });
+            obj.emplace("prompt_attachments", json::array_t(attachments.begin(), attachments.end()));
+        }
+    }
+    return obj;
+}

gpt4all-chat/src/jinja_helpers.h

@@ -3,47 +3,17 @@
 #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 <nlohmann/json.hpp>
 
 #include <QtGlobal>
 
-namespace views = std::views;
+using json = nlohmann::ordered_json;
 
 
-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 {
+class JinjaHelper {
 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() }; }
+    json::object_t AsJson() const { return static_cast<const Derived *>(this)->AsJson(); }
 };
 
 class JinjaResultInfo : public JinjaHelper<JinjaResultInfo> {
@@ -51,18 +21,10 @@ 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; }
+    json::object_t AsJson() const;
 
 private:
-    static const JinjaFieldMap<ResultInfo> s_fields;
     const ResultInfo *m_source;
-
-    friend class JinjaHelper<JinjaResultInfo>;
 };
 
 class JinjaPromptAttachment : public JinjaHelper<JinjaPromptAttachment> {
@@ -70,18 +32,10 @@ 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; }
+    json::object_t AsJson() const;
 
 private:
-    static const JinjaFieldMap<PromptAttachment> s_fields;
     const PromptAttachment *m_attachment;
-
-    friend class JinjaHelper<JinjaPromptAttachment>;
 };
 
 class JinjaMessage : public JinjaHelper<JinjaMessage> {
@@ -89,28 +43,9 @@ public:
     explicit JinjaMessage(uint version, const MessageItem &item) noexcept
         : m_version(version), m_item(&item) {}
 
-    const JinjaMessage &value  () const { return *this;     }
-    uint                version() const { return m_version; }
-    const MessageItem      &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;
+    json::object_t AsJson() const;
 
 private:
-    auto keys() const -> const std::unordered_set<std::string_view> &;
-
-private:
-    static const JinjaFieldMap<JinjaMessage> s_fields;
-    uint            m_version;
+    uint               m_version;
     const MessageItem *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

@@ -1,17 +0,0 @@
-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/jinja_replacements.cpp

@@ -5,26 +5,9 @@
 // This is a list of prompt templates known to GPT4All and their associated replacements which are automatically used
 // instead when loading the chat template from GGUF. These exist for two primary reasons:
 // - HuggingFace model authors make ugly chat templates because they do not expect the end user to see them;
-// - and our Jinja2Cpp-based template parsing is not fully compatible with HuggingFace transformers and jinja2.
-
-// Below is a list of known incompatibilities with the official HF jinja2 implementation. These are not all necessarily
-// reflected in the below substitution list, and this cannot be an exhaustive list because there are a plethora of edge
-// cases in template parsing in which jinja2 and Jinja2Cpp differ. These are differences that could be reasonably
-// expected to affect chat templates that could be seen in the wild, or that cause a crash:
-// - Jinja2Cpp crashes (in debug builds) if given the template `a[""(`
-// - Jinja2Cpp does not support these jinja2 constructs:
-//   - `is not none`
-//   - list slicing, e.g. `messages[1:]`
-//   - the jinja2.ext.loopcontrols extension, which HF enables by default
-//   - a missing space after a quote in substitution (e.g. `{{ 'foo'}}`), which *has* been seen in the wild
-// - GPT4All does not currently support these HuggingFace template features:
-//   - customized "tojson" filter (we provide the built-in Jinja2Cpp one)
-//   - the AssistantTracker extension
-
+// - and chat templates occasionally use features we do not support. This is less true now that we use minja.
 
 // The substitution list.
-// For templates that apply to models listed in models3.json, these should be copied there as well for best
-// compatibility with older versions of GPT4All.
 
 const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBSTITUTIONS {
     // calme-2.1-phi3.5-4b.Q6_K.gguf (reported by ThilotE on Discord), Phi-3.5-mini-instruct-Q4_0.gguf (nomic-ai/gpt4all#3345)
@@ -52,6 +35,30 @@ const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBST
     {{- '<|assistant|>\n' }}
 {%- else %}
     {{- eos_token }}
+{%- endif %})TEMPLATE",
+    },
+    // DeepSeek-R1-Distill-Qwen-7B-Q4_0.gguf
+    {
+        // original
+        R"TEMPLATE({% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>'}}{% endif %})TEMPLATE",
+        // replacement
+        R"TEMPLATE({%- if not add_generation_prompt is defined %}
+    {%- set add_generation_prompt = false %}
+{%- endif %}
+{%- if messages[0]['role'] == 'system' %}
+    {{- messages[0]['content'] }}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message['role'] == 'user' %}
+        {{- '<｜User｜>' + message['content'] }}
+    {%- endif %}
+    {%- if message['role'] == 'assistant' %}
+        {%- set content = message['content'] | regex_replace('^[\\s\\S]*</think>', '') %}
+        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}
+    {%- endif %}
+{%- endfor -%}
+{%- if add_generation_prompt %}
+    {{- '<｜Assistant｜>' }}
 {%- endif %})TEMPLATE",
     },
     // gemma-2-9b-it-Q4_0.gguf (nomic-ai/gpt4all#3282)
@@ -626,7 +633,7 @@ const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBST
         // replacement
         R"TEMPLATE({%- for message in messages %}
     {%- if message['role'] == 'system' %}
-        {{-'<|system|>\n' + message['content'] + '<|end|>\n'}}
+        {{- '<|system|>\n' + message['content'] + '<|end|>\n' }}
     {%- elif message['role'] == 'user' %}
         {{- '<|user|>\n' + message['content'] + '<|end|>\n' }}
     {%- elif message['role'] == 'assistant' %}

gpt4all-chat/src/main.cpp

@@ -25,6 +25,10 @@
 #include <QVariant>
 #include <Qt>
 
+#if G4A_CONFIG(force_d3d12)
+#   include <QSGRendererInterface>
+#endif
+
 #ifndef GPT4ALL_USE_QTPDF
 #   include <fpdfview.h>
 #endif
@@ -83,24 +87,27 @@ int main(int argc, char *argv[])
         return 0;
     }
 
+#if G4A_CONFIG(force_d3d12)
+    QQuickWindow::setGraphicsApi(QSGRendererInterface::Direct3D12);
+#endif
+
 #ifdef Q_OS_LINUX
     app.setWindowIcon(QIcon(":/gpt4all/icons/gpt4all.svg"));
 #endif
 
     // set search path before constructing the MySettings instance, which relies on this
-    QString llmodelSearchPaths = QCoreApplication::applicationDirPath();
-    const QString libDir = QCoreApplication::applicationDirPath() + "/../lib/";
-    if (LLM::directoryExists(libDir))
-        llmodelSearchPaths += ";" + libDir;
-#if defined(Q_OS_MAC)
-    const QString binDir = QCoreApplication::applicationDirPath() + "/../../../";
-    if (LLM::directoryExists(binDir))
-        llmodelSearchPaths += ";" + binDir;
-    const QString frameworksDir = QCoreApplication::applicationDirPath() + "/../Frameworks/";
-    if (LLM::directoryExists(frameworksDir))
-        llmodelSearchPaths += ";" + frameworksDir;
+    {
+        auto appDirPath = QCoreApplication::applicationDirPath();
+        QStringList searchPaths {
+#ifdef Q_OS_DARWIN
+            u"%1/../Frameworks"_s.arg(appDirPath),
+#else
+            appDirPath,
+            u"%1/../lib"_s.arg(appDirPath),
 #endif
-    LLModel::Implementation::setImplementationsSearchPath(llmodelSearchPaths.toStdString());
+        };
+        LLModel::Implementation::setImplementationsSearchPath(searchPaths.join(u';').toStdString());
+    }
 
     // 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.
@@ -174,5 +181,9 @@ int main(int argc, char *argv[])
     // Otherwise, we can get a heap-use-after-free inside of llama.cpp.
     ChatListModel::globalInstance()->destroyChats();
 
+#ifndef GPT4ALL_USE_QTPDF
+    FPDF_DestroyLibrary();
+#endif
+
     return res;
 }

gpt4all-chat/src/modellist.cpp

@@ -1621,7 +1621,6 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
         QString requiresVersion = obj["requires"].toString();
         QString versionRemoved = obj["removedIn"].toString();
         QString url = obj["url"].toString();
-        QByteArray modelHash = obj["md5sum"].toString().toLatin1();
         bool isDefault = obj.contains("isDefault") && obj["isDefault"] == u"true"_s;
         bool disableGUI = obj.contains("disableGUI") && obj["disableGUI"] == u"true"_s;
         QString description = obj["description"].toString();
@@ -1632,6 +1631,16 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
         QString type = obj["type"].toString();
         bool isEmbeddingModel = obj["embeddingModel"].toBool();
 
+        QByteArray modelHash;
+        ModelInfo::HashAlgorithm hashAlgorithm;
+        if (auto it = obj.find("sha256sum"_L1); it != obj.end()) {
+            modelHash = it->toString().toLatin1();
+            hashAlgorithm = ModelInfo::Sha256;
+        } else {
+            modelHash = obj["md5sum"].toString().toLatin1();
+            hashAlgorithm = ModelInfo::Md5;
+        }
+
         // Some models aren't supported in the GUI at all
         if (disableGUI)
             continue;
@@ -1660,7 +1669,7 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
             { ModelList::FilenameRole, modelFilename },
             { ModelList::FilesizeRole, modelFilesize },
             { ModelList::HashRole, modelHash },
-            { ModelList::HashAlgorithmRole, ModelInfo::Md5 },
+            { ModelList::HashAlgorithmRole, hashAlgorithm },
             { ModelList::DefaultRole, isDefault },
             { ModelList::DescriptionRole, description },
             { ModelList::RequiresVersionRole, requiresVersion },

gpt4all-chat/src/modellist.h

@@ -269,7 +269,7 @@ private:
             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_chatNamePrompt          = "Describe the above conversation. Your entire response must be three 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;

gpt4all-chat/src/tool.cpp

@@ -1,12 +1,13 @@
 #include "tool.h"
 
-#include <jinja2cpp/value.h>
-
 #include <string>
 
-jinja2::Value Tool::jinjaValue() const
+using json = nlohmann::ordered_json;
+
+
+json::object_t Tool::jinjaValue() const
 {
-    jinja2::ValuesList paramList;
+    json::array_t paramList;
     const QList<ToolParamInfo> p = parameters();
     for (auto &info : p) {
         std::string typeStr;
@@ -20,26 +21,24 @@ jinja2::Value Tool::jinjaValue() const
         case Boolean:  typeStr = "boolean"; break;
         case Null:     typeStr = "null"; break;
         }
-        jinja2::ValuesMap infoMap {
-            { "name", info.name.toStdString() },
-            { "type", typeStr},
+        paramList.emplace_back(json::initializer_list_t {
+            { "name",        info.name.toStdString()        },
+            { "type",        typeStr                        },
             { "description", info.description.toStdString() },
-            { "required", info.required }
-        };
-        paramList.push_back(infoMap);
+            { "required",    info.required                  },
+        });
     }
 
-    jinja2::ValuesMap params {
-        { "name", name().toStdString() },
-        { "description", description().toStdString() },
-        { "function", function().toStdString() },
-        { "parameters", paramList },
+    return {
+        { "name",           name().toStdString()           },
+        { "description",    description().toStdString()    },
+        { "function",       function().toStdString()       },
+        { "parameters",     paramList                      },
         { "symbolicFormat", symbolicFormat().toStdString() },
-        { "examplePrompt", examplePrompt().toStdString() },
-        { "exampleCall", exampleCall().toStdString() },
-        { "exampleReply", exampleReply().toStdString() }
+        { "examplePrompt",  examplePrompt().toStdString()  },
+        { "exampleCall",    exampleCall().toStdString()    },
+        { "exampleReply",   exampleReply().toStdString()   },
     };
-    return params;
 }
 
 void ToolCallInfo::serialize(QDataStream &stream, int version)

gpt4all-chat/src/tool.h

@@ -1,13 +1,16 @@
 #ifndef TOOL_H
 #define TOOL_H
 
+#include <nlohmann/json.hpp>
+
 #include <QList>
 #include <QObject>
 #include <QString>
 #include <QVariant>
 #include <QtGlobal>
 
-#include <jinja2cpp/value.h>
+using json = nlohmann::ordered_json;
+
 
 namespace ToolEnums
 {
@@ -25,6 +28,7 @@ namespace ToolEnums
 
     enum class ParseState {
         None,
+        InTagChoice,
         InStart,
         Partial,
         Complete,
@@ -87,7 +91,8 @@ public:
     Tool() : QObject(nullptr) {}
     virtual ~Tool() {}
 
-    virtual QString run(const QList<ToolParam> &params, qint64 timeout = 2000) = 0;
+    virtual void run(const QList<ToolParam> &params) = 0;
+    virtual bool interrupt() = 0;
 
     // Tools should set these if they encounter errors. For instance, a tool depending upon the network
     // might set these error variables if the network is not available.
@@ -121,7 +126,10 @@ public:
 
     bool operator==(const Tool &other) const { return function() == other.function(); }
 
-    jinja2::Value jinjaValue() const;
+    json::object_t jinjaValue() const;
+
+Q_SIGNALS:
+    void runComplete(const ToolCallInfo &info);
 };
 
 #endif // TOOL_H

gpt4all-chat/src/toolcallparser.cpp

@@ -1,16 +1,29 @@
 #include "toolcallparser.h"
 
-#include <QDebug>
+#include "tool.h"
+
+#include <QSet>
 #include <QtGlobal>
-#include <QtLogging>
 
-#include <cstddef>
+#include <stdexcept>
 
-static const QString ToolCallStart = ToolCallConstants::CodeInterpreterTag;
-static const QString ToolCallEnd = ToolCallConstants::CodeInterpreterEndTag;
 
 ToolCallParser::ToolCallParser()
+    : ToolCallParser(ToolCallConstants::AllTagNames)
+    {}
+
+ToolCallParser::ToolCallParser(const QStringList &tagNames)
 {
+    QSet<QChar> firstChars;
+    for (auto &name : tagNames) {
+        if (name.isEmpty())
+            throw std::invalid_argument("ToolCallParser(): tag names must not be empty");
+        if (firstChars.contains(name.at(0)))
+            throw std::invalid_argument("ToolCallParser(): tag names must not share any prefix");
+        firstChars << name.at(0);
+        m_possibleStartTags << makeStartTag(name).toUtf8();
+        m_possibleEndTags   << makeEndTag  (name).toUtf8();
+    }
     reset();
 }
 
@@ -20,36 +33,68 @@ void ToolCallParser::reset()
     resetSearchState();
 
     // These are global states maintained between update calls
-    m_buffer.clear();
-    m_hasSplit = false;
+    m_buffers.clear();
+    m_buffers << QByteArray();
 }
 
 void ToolCallParser::resetSearchState()
 {
-    m_expected = ToolCallStart.at(0);
+    m_expected = {'<'};
     m_expectedIndex = 0;
     m_state = ToolEnums::ParseState::None;
+
     m_toolCall.clear();
+    m_startTagBuffer.clear();
     m_endTagBuffer.clear();
+
+    m_currentTagIndex = -1;
     m_startIndex = -1;
+    m_endIndex = -1;
+}
+
+bool ToolCallParser::isExpected(char c) const
+{
+    return m_expected.isEmpty() || m_expected.contains(c);
+}
+
+void ToolCallParser::setExpected(const QList<QByteArray> &tags)
+{
+    m_expected.clear();
+    for (const auto &tag : tags) {
+        Q_ASSERT(tag.size() > m_expectedIndex);
+        m_expected << tag.at(m_expectedIndex);
+    }
+}
+
+QByteArray ToolCallParser::startTag() const
+{
+    if (m_currentTagIndex < 0)
+        return {};
+    return m_possibleStartTags.at(m_currentTagIndex);
+}
+
+QByteArray ToolCallParser::endTag() const
+{
+    if (m_currentTagIndex < 0)
+        return {};
+    return m_possibleEndTags.at(m_currentTagIndex);
+}
+
+QByteArray &ToolCallParser::currentBuffer()
+{
+    return m_buffers.last();
 }
 
 // This method is called with an arbitrary string and a current state. This method should take the
 // current state into account and then parse through the update character by character to arrive at
 // the new state.
-void ToolCallParser::update(const QString &update)
+void ToolCallParser::update(const QByteArray &update)
 {
-    Q_ASSERT(m_state != ToolEnums::ParseState::Complete);
-    if (m_state == ToolEnums::ParseState::Complete) {
-        qWarning() << "ERROR: ToolCallParser::update already found a complete toolcall!";
-        return;
-    }
+    currentBuffer().append(update);
 
-    m_buffer.append(update);
-
-    for (size_t i = m_buffer.size() - update.size(); i < m_buffer.size(); ++i) {
-        const QChar c = m_buffer[i];
-        const bool foundMatch = m_expected.isNull() || c == m_expected;
+    for (qsizetype i = currentBuffer().size() - update.size(); i < currentBuffer().size(); ++i) {
+        const char c = currentBuffer()[i];
+        const bool foundMatch = isExpected(c);
         if (!foundMatch) {
             resetSearchState();
             continue;
@@ -59,34 +104,58 @@ void ToolCallParser::update(const QString &update)
         case ToolEnums::ParseState::None:
             {
                 m_expectedIndex = 1;
-                m_expected = ToolCallStart.at(1);
-                m_state = ToolEnums::ParseState::InStart;
+                setExpected(m_possibleStartTags);
+                m_state = ToolEnums::ParseState::InTagChoice;
                 m_startIndex = i;
                 break;
             }
+        case ToolEnums::ParseState::InTagChoice:
+            {
+                for (int i = 0; i < m_possibleStartTags.size(); ++i) {
+                    const auto &tag = m_possibleStartTags.at(i);
+                    if (c == tag.at(1)) m_currentTagIndex = i;
+                }
+                if (m_currentTagIndex >= 0) {
+                    m_expectedIndex = 2;
+                    setExpected({m_possibleStartTags.at(m_currentTagIndex)});
+                    m_state = ToolEnums::ParseState::InStart;
+                } else
+                    resetSearchState();
+                break;
+            }
         case ToolEnums::ParseState::InStart:
             {
-                if (m_expectedIndex == ToolCallStart.size() - 1) {
+                m_startTagBuffer.append(c);
+
+                const auto startTag = this->startTag();
+                Q_ASSERT(!startTag.isEmpty());
+                if (m_expectedIndex == startTag.size() - 1) {
                     m_expectedIndex = 0;
-                    m_expected = QChar();
+                    setExpected({});
                     m_state = ToolEnums::ParseState::Partial;
                 } else {
                     ++m_expectedIndex;
-                    m_expected = ToolCallStart.at(m_expectedIndex);
+                    Q_ASSERT(m_currentTagIndex >= 0);
+                    setExpected({startTag});
                 }
                 break;
             }
         case ToolEnums::ParseState::Partial:
             {
+                Q_ASSERT(m_currentTagIndex >= 0);
+                const auto endTag = this->endTag();
+                Q_ASSERT(!endTag.isEmpty());
                 m_toolCall.append(c);
                 m_endTagBuffer.append(c);
-                if (m_endTagBuffer.size() > ToolCallEnd.size())
+                if (m_endTagBuffer.size() > endTag.size())
                     m_endTagBuffer.remove(0, 1);
-                if (m_endTagBuffer == ToolCallEnd) {
-                    m_toolCall.chop(ToolCallEnd.size());
+                if (m_endTagBuffer == endTag) {
+                    m_endIndex = i + 1;
+                    m_toolCall.chop(endTag.size());
                     m_state = ToolEnums::ParseState::Complete;
                     m_endTagBuffer.clear();
                 }
+                break;
             }
         case ToolEnums::ParseState::Complete:
             {
@@ -97,15 +166,35 @@ void ToolCallParser::update(const QString &update)
     }
 }
 
-QPair<QString, QString> ToolCallParser::split()
+bool ToolCallParser::splitIfPossible()
 {
-    Q_ASSERT(m_state == ToolEnums::ParseState::Partial
-        || m_state == ToolEnums::ParseState::Complete);
+    // The first split happens when we're in a partial state
+    if (m_buffers.size() < 2 && m_state == ToolEnums::ParseState::Partial) {
+        Q_ASSERT(m_startIndex >= 0);
+        const auto beforeToolCall = currentBuffer().left(m_startIndex);
+        const auto toolCall       = currentBuffer().mid (m_startIndex);
+        m_buffers = { beforeToolCall, toolCall };
+        return true;
+    }
 
-    Q_ASSERT(m_startIndex >= 0);
-    m_hasSplit = true;
-    const QString beforeToolCall = m_buffer.left(m_startIndex);
-    m_buffer = m_buffer.mid(m_startIndex);
-    m_startIndex = 0;
-    return { beforeToolCall, m_buffer };
+    // The second split happens when we're in the complete state
+    if (m_buffers.size() < 3 && m_state == ToolEnums::ParseState::Complete) {
+        Q_ASSERT(m_endIndex >= 0);
+        const auto &beforeToolCall = m_buffers.first();
+        const auto toolCall        = currentBuffer().left(m_endIndex);
+        const auto afterToolCall   = currentBuffer().mid (m_endIndex);
+        m_buffers = { beforeToolCall, toolCall, afterToolCall };
+        return true;
+    }
+
+    return false;
+}
+
+QStringList ToolCallParser::buffers() const
+{
+    QStringList result;
+    result.reserve(m_buffers.size());
+    for (const auto &buffer : m_buffers)
+        result << QString::fromUtf8(buffer);
+    return result;
 }

gpt4all-chat/src/toolcallparser.h

@@ -1,47 +1,73 @@
 #ifndef TOOLCALLPARSER_H
 #define TOOLCALLPARSER_H
 
-#include "tool.h"
-
-#include <QChar>
+#include <QByteArray>
+#include <QList>
 #include <QString>
-#include <QPair>
+#include <QStringList>
+
+namespace ToolEnums { enum class ParseState; }
+
+using namespace Qt::Literals::StringLiterals;
 
-namespace ToolCallConstants
-{
-    const QString CodeInterpreterFunction = R"(javascript_interpret)";
-    const QString CodeInterpreterTag = R"(<)" + CodeInterpreterFunction + R"(>)";
-    const QString CodeInterpreterEndTag = R"(</)" + CodeInterpreterFunction + R"(>)";
-    const QString CodeInterpreterPrefix = CodeInterpreterTag + "\n```javascript\n";
-    const QString CodeInterpreterSuffix = "```\n" + CodeInterpreterEndTag;
-}
 
 class ToolCallParser
 {
 public:
     ToolCallParser();
+    ToolCallParser(const QStringList &tagNames);
+
     void reset();
-    void update(const QString &update);
-    QString buffer() const { return m_buffer; }
-    QString toolCall() const { return m_toolCall; }
+    void update(const QByteArray &update);
+    QString toolCall() const { return QString::fromUtf8(m_toolCall); }
     int startIndex() const { return m_startIndex; }
     ToolEnums::ParseState state() const { return m_state; }
+    QByteArray startTag() const;
+    QByteArray endTag() const;
 
-    // Splits
-    QPair<QString, QString> split();
-    bool hasSplit() const { return m_hasSplit; }
+    bool splitIfPossible();
+    QStringList buffers() const;
+    int numberOfBuffers() const { return m_buffers.size(); }
+
+    static QString makeStartTag(const QString &name) { return u"<%1>"_s .arg(name); }
+    static QString makeEndTag  (const QString &name) { return u"</%1>"_s.arg(name); }
 
 private:
+    QByteArray &currentBuffer();
     void resetSearchState();
+    bool isExpected(char c) const;
+    void setExpected(const QList<QByteArray> &tags);
 
-    QChar m_expected;
+    QList<QByteArray> m_possibleStartTags;
+    QList<QByteArray> m_possibleEndTags;
+    QByteArray m_startTagBuffer;
+    QByteArray m_endTagBuffer;
+    int m_currentTagIndex;
+
+    QList<char> m_expected;
     int m_expectedIndex;
     ToolEnums::ParseState m_state;
-    QString m_buffer;
-    QString m_toolCall;
-    QString m_endTagBuffer;
+    QList<QByteArray> m_buffers;
+    QByteArray m_toolCall;
     int m_startIndex;
-    bool m_hasSplit;
+    int m_endIndex;
 };
 
+namespace ToolCallConstants
+{
+    // NB: the parsing code assumes the first char of the various tags differ
+
+    inline const QString CodeInterpreterFunction = u"javascript_interpret"_s;
+    inline const QString CodeInterpreterStartTag = ToolCallParser::makeStartTag(CodeInterpreterFunction);
+    inline const QString CodeInterpreterEndTag   = ToolCallParser::makeEndTag  (CodeInterpreterFunction);
+    inline const QString CodeInterpreterPrefix   = u"%1\n```javascript\n"_s.arg(CodeInterpreterStartTag);
+    inline const QString CodeInterpreterSuffix   = u"```\n%1"_s            .arg(CodeInterpreterEndTag  );
+
+    inline const QString ThinkTagName  = u"think"_s;
+    inline const QString ThinkStartTag = ToolCallParser::makeStartTag(ThinkTagName);
+    inline const QString ThinkEndTag   = ToolCallParser::makeEndTag  (ThinkTagName);
+
+    inline const QStringList AllTagNames { CodeInterpreterFunction, ThinkTagName };
+}
+
 #endif // TOOLCALLPARSER_H