templates: Add Ollama multi-modal templates (#14868)

Templates for [local multi-modal
LLMs](https://llava-vl.github.io/llava-interactive/) using -
* Image summaries
* Multi-modal embeddings

---------

Co-authored-by: Erick Friis <erick@langchain.dev>

templates/rag-chroma-multi-modal-multi-vector/README.md

@@ -1,53 +1,72 @@
 
 # rag-chroma-multi-modal-multi-vector
 
-Presentations (slide decks, etc) contain visual content that challenges conventional RAG.
+Multi-modal LLMs enable text-to-image retrieval and question-answering over images. 
 
-Multi-modal LLMs unlock new ways to build apps over visual content like presentations.
+You can ask questions in natural language about a collection of photos, retrieve relevant ones, and have a multi-modal LLM answer questions about the retrieved images.
- 
-This template performs multi-modal RAG using Chroma with the multi-vector retriever (see [blog](https://blog.langchain.dev/multi-modal-rag-template/)):
 
-* Extracts the slides as images
+This template performs text-to-image retrieval for question-answering about a slide deck, which often contains visual elements that are not captured in standard RAG.
-* Uses GPT-4V to summarize each image
+
-* Embeds the image summaries with a link to the original images
+This will use GPT-4V for image captioning and answer synthesis.
-* Retrieves relevant image based on similarity between the image summary and the user input
+
-* Finally pass those images to GPT-4V for answer synthesis
+## Input
+
+Supply a slide deck as pdf in the `/docs` directory. 
+
+By default, this template has a slide deck about Q3 earnings from DataDog, a public techologyy company.
+
+Example questions to ask can be:
+```
+How many customers does Datadog have?
+What is Datadog platform % Y/Y growth in FY20, FY21, and FY22?
+```
+
+To create an index of the slide deck, run:
+```
+poetry install
+python ingest.py
+```
 
 ## Storage
 
+Here is the process the template will use to create an index of the slides (see [blog](https://blog.langchain.dev/multi-modal-rag-template/)):
+
+* Extract the slides as a collection of images
+* Use GPT-4V to summarize each image
+* Embed the image summaries using text embeddings with a link to the original images
+* Retrieve relevant image based on similarity between the image summary and the user input question
+* Pass those images to GPT-4V for answer synthesis
+
+By default, this will use [LocalFileStore](https://python.langchain.com/docs/integrations/stores/file_system) to store images and Chroma to store summaries.
+
+For production, it may be desirable to use a remote option such as Redis.
+
+You can set the `local_file_store` flag in `chain.py` and `ingest.py` to switch between the two options.
+
+For Redis, the template will use [UpstashRedisByteStore](https://python.langchain.com/docs/integrations/stores/upstash_redis).
+
 We will use Upstash to store the images, which offers Redis with a REST API.
 
 Simply login [here](https://upstash.com/) and create a database.
 
 This will give you a REST API with:
 
-* UPSTASH_URL
+* `UPSTASH_URL`
-* UPSTASH_TOKEN
+* `UPSTASH_TOKEN`
-
+ 
 Set `UPSTASH_URL` and `UPSTASH_TOKEN` as environment variables to access your database.
 
 We will use Chroma to store and index the image summaries, which will be created locally in the template directory.
 
-## Input
-
-Supply a slide deck as pdf in the `/docs` directory. 
-
-Create your vectorstore (Chroma) and populae Upstash with: 
-
-```
-poetry install
-python ingest.py
-```
-
 ## LLM
 
-The app will retrieve images using multi-modal embeddings, and pass them to GPT-4V.
+The app will retrieve images based on similarity between the text input and the image summary, and pass the images to GPT-4V.
 
 ## Environment Setup
 
 Set the `OPENAI_API_KEY` environment variable to access the OpenAI GPT-4V.
 
-Set `UPSTASH_URL` and `UPSTASH_TOKEN` as environment variables to access your database.
+Set `UPSTASH_URL` and `UPSTASH_TOKEN` as environment variables to access your database if you use `UpstashRedisByteStore`.
 
 ## Usage
 

templates/rag-chroma-multi-modal-multi-vector/ingest.py

@@ -11,7 +11,7 @@ from langchain.embeddings import OpenAIEmbeddings
 from langchain.retrievers.multi_vector import MultiVectorRetriever
 from langchain.schema.document import Document
 from langchain.schema.messages import HumanMessage
-from langchain.storage import UpstashRedisByteStore
+from langchain.storage import LocalFileStore, UpstashRedisByteStore
 from langchain.vectorstores import Chroma
 from PIL import Image
 
@@ -126,20 +126,31 @@ def convert_to_base64(pil_image):
     return img_str
 
 
-def create_multi_vector_retriever(vectorstore, image_summaries, images):
+def create_multi_vector_retriever(
+    vectorstore, image_summaries, images, local_file_store
+):
     """
     Create retriever that indexes summaries, but returns raw images or texts
 
     :param vectorstore: Vectorstore to store embedded image sumamries
     :param image_summaries: Image summaries
     :param images: Base64 encoded images
+    :param local_file_store: Use local file storage
     :return: Retriever
     """
 
-    # Initialize the storage layer for images
+    # File storage option
-    UPSTASH_URL = os.getenv("UPSTASH_URL")
+    if local_file_store:
-    UPSTASH_TOKEN = os.getenv("UPSTASH_TOKEN")
+        store = LocalFileStore(
-    store = UpstashRedisByteStore(url=UPSTASH_URL, token=UPSTASH_TOKEN)
+            str(Path(__file__).parent / "multi_vector_retriever_metadata")
+        )
+    else:
+        # Initialize the storage layer for images using Redis
+        UPSTASH_URL = os.getenv("UPSTASH_URL")
+        UPSTASH_TOKEN = os.getenv("UPSTASH_TOKEN")
+        store = UpstashRedisByteStore(url=UPSTASH_URL, token=UPSTASH_TOKEN)
+
+    # Doc ID
     id_key = "doc_id"
 
     # Create the multi-vector retriever
@@ -194,4 +205,5 @@ retriever_multi_vector_img = create_multi_vector_retriever(
     vectorstore_mvr,
     image_summaries,
     images_base_64_processed_documents,
+    local_file_store=True,
 )

templates/rag-chroma-multi-modal-multi-vector/rag_chroma_multi_modal_multi_vector/chain.py

@@ -11,7 +11,7 @@ from langchain.schema.document import Document
 from langchain.schema.messages import HumanMessage
 from langchain.schema.output_parser import StrOutputParser
 from langchain.schema.runnable import RunnableLambda, RunnablePassthrough
-from langchain.storage import UpstashRedisByteStore
+from langchain.storage import LocalFileStore, UpstashRedisByteStore
 from langchain.vectorstores import Chroma
 from PIL import Image
 
@@ -101,6 +101,9 @@ def multi_modal_rag_chain(retriever):
     return chain
 
 
+# Flag
+local_file_store = True
+
 # Load chroma
 vectorstore_mvr = Chroma(
     collection_name="image_summaries",
@@ -108,10 +111,17 @@ vectorstore_mvr = Chroma(
     embedding_function=OpenAIEmbeddings(),
 )
 
-# Load redis
+if local_file_store:
-UPSTASH_URL = os.getenv("UPSTASH_URL")
+    store = LocalFileStore(
-UPSTASH_TOKEN = os.getenv("UPSTASH_TOKEN")
+        str(Path(__file__).parent.parent / "multi_vector_retriever_metadata")
-store = UpstashRedisByteStore(url=UPSTASH_URL, token=UPSTASH_TOKEN)
+    )
+else:
+    # Load redis
+    UPSTASH_URL = os.getenv("UPSTASH_URL")
+    UPSTASH_TOKEN = os.getenv("UPSTASH_TOKEN")
+    store = UpstashRedisByteStore(url=UPSTASH_URL, token=UPSTASH_TOKEN)
+
+#
 id_key = "doc_id"
 
 # Create the multi-vector retriever

templates/rag-chroma-multi-modal/README.md

@@ -1,32 +1,41 @@
 
 # rag-chroma-multi-modal
 
-Presentations (slide decks, etc) contain visual content that challenges conventional RAG.
+Multi-modal LLMs enable text-to-image retrieval and question-answering over images. 
 
-Multi-modal LLMs unlock new ways to build apps over visual content like presentations.
+You can ask questions in natural language about a collection of photos, retrieve relevant ones, and have a multi-modal LLM answer questions about the retrieved images.
- 
+
-This template performs multi-modal RAG using Chroma with multi-modal OpenCLIP embeddings and OpenAI GPT-4V.
+This template performs text-to-image retrieval for question-answering about a slide deck, which often contains visual elements that are not captured in standard RAG.
+
+This will use OpenCLIP embeddings and GPT-4V for answer synthesis.
 
 ## Input
 
 Supply a slide deck as pdf in the `/docs` directory. 
 
-Create your vectorstore with: 
+By default, this template has a slide deck about Q3 earnings from DataDog, a public techologyy company.
 
+Example questions to ask can be:
+```
+How many customers does Datadog have?
+What is Datadog platform % Y/Y growth in FY20, FY21, and FY22?
+```
+
+To create an index of the slide deck, run:
 ```
 poetry install
 python ingest.py
 ```
 
-## Embeddings
+## Storage
 
-This template will use [OpenCLIP](https://github.com/mlfoundations/open_clip) multi-modal embeddings.
+This template will use [OpenCLIP](https://github.com/mlfoundations/open_clip) multi-modal embeddings to embed the images.
 
-You can select different options (see results [here](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv)).
+You can select different embedding model options (see results [here](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv)).
 
 The first time you run the app, it will automatically download the multimodal embedding model.
 
-By default, LangChain will use an embedding model with reasonably strong performance, `ViT-H-14`.
+By default, LangChain will use an embedding model with moderate performance but lower memory requirments, `ViT-H-14`.
 
 You can choose alternative `OpenCLIPEmbeddings` models in `rag_chroma_multi_modal/ingest.py`:
 ```
@@ -41,7 +50,7 @@ vectorstore_mmembd = Chroma(
 
 ## LLM
 
-The app will retrieve images using multi-modal embeddings, and pass them to GPT-4V.
+The app will retrieve images based on similarity between the text input and the image, which are both mapped to multi-modal embedding space. It will then pass the images to GPT-4V.
 
 ## Environment Setup
 

templates/rag-gemini-multi-modal/README.md

@@ -1,34 +1,43 @@
 
 # rag-gemini-multi-modal
 
-Presentations (slide decks, etc) contain visual content that challenges conventional RAG.
+Multi-modal LLMs enable text-to-image retrieval and question-answering over images. 
 
-Multi-modal LLMs unlock new ways to build apps over visual content like presentations.
+You can ask questions in natural language about a collection of photos, retrieve relevant ones, and have a multi-modal LLM answer questions about the retrieved images.
+
+This template performs text-to-image retrieval for question-answering about a slide deck, which often contains visual elements that are not captured in standard RAG.
  
-This template performs multi-modal RAG using Chroma with multi-modal OpenCLIP embeddings and [Google Gemini](https://deepmind.google/technologies/gemini/#introduction).
+This will use OpenCLIP embeddings and [Google Gemini](https://deepmind.google/technologies/gemini/#introduction) for answer synthesis.
 
 ## Input
 
 Supply a slide deck as pdf in the `/docs` directory. 
 
-Create your vectorstore with: 
+By default, this template has a slide deck about Q3 earnings from DataDog, a public techologyy company.
 
+Example questions to ask can be:
+```
+How many customers does Datadog have?
+What is Datadog platform % Y/Y growth in FY20, FY21, and FY22?
+```
+
+To create an index of the slide deck, run:
 ```
 poetry install
 python ingest.py
 ```
 
-## Embeddings
+## Storage
 
-This template will use [OpenCLIP](https://github.com/mlfoundations/open_clip) multi-modal embeddings.
+This template will use [OpenCLIP](https://github.com/mlfoundations/open_clip) multi-modal embeddings to embed the images.
 
-You can select different options (see results [here](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv)).
+You can select different embedding model options (see results [here](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv)).
 
 The first time you run the app, it will automatically download the multimodal embedding model.
 
-By default, LangChain will use an embedding model with reasonably strong performance, `ViT-H-14`.
+By default, LangChain will use an embedding model with moderate performance but lower memory requirments, `ViT-H-14`.
 
-You can choose alternative `OpenCLIPEmbeddings` models in `ingest.py`:
+You can choose alternative `OpenCLIPEmbeddings` models in `rag_chroma_multi_modal/ingest.py`:
 ```
 vectorstore_mmembd = Chroma(
     collection_name="multi-modal-rag",
@@ -45,7 +54,7 @@ The app will retrieve images using multi-modal embeddings, and pass them to Goog
 
 ## Environment Setup
 
-Set the `GOOGLE_API_KEY` environment variable to access Gemini.
+Set your `GOOGLE_API_KEY` environment variable in order to access Gemini.
 
 ## Usage
 

templates/rag-multi-modal-local/.gitignore

@@ -0,0 +1,2 @@
+docs/img_*.jpg
+chroma_db_multi_modal

templates/rag-multi-modal-local/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 LangChain, Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

templates/rag-multi-modal-local/README.md

@@ -0,0 +1,122 @@
+
+# rag-multi-modal-local
+
+Visual search is a famililar application to many with iPhones or Android devices: use natural language to search across your photo collection. 
+  
+With the release of open source, multi-modal LLMs it's possible to build this kind of application for yourself and have it run on your personal laptop. 
+
+This template demonstrates how to perform visual search and question-answering over a collection of photos.
+ 
+Given a set of photos, it will use OpenCLIP embeddings to index them, retrieve photos relevant to user question, and use Ollama to run a local, open-source multi-modal LLM to answer questions about the retrieved photos.
+
+## Input
+
+Supply a set of photos in the `/docs` directory. 
+
+By default, this template has a toy collection of 3 food pictures.
+
+Example questions to ask can be:
+```
+What kind of soft serve did I have?
+```
+
+In practice, a larger corpus of images can be tested.
+
+To create an index of the images, run:
+```
+poetry install
+python ingest.py
+```
+
+## Storage
+
+This template will use [OpenCLIP](https://github.com/mlfoundations/open_clip) multi-modal embeddings to embed the images.
+
+You can select different embedding model options (see results [here](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv)).
+
+The first time you run the app, it will automatically download the multimodal embedding model.
+
+By default, LangChain will use an embedding model with moderate performance but lower memory requirments, `ViT-H-14`.
+
+You can choose alternative `OpenCLIPEmbeddings` models in `rag_chroma_multi_modal/ingest.py`:
+```
+vectorstore_mmembd = Chroma(
+    collection_name="multi-modal-rag",
+    persist_directory=str(re_vectorstore_path),
+    embedding_function=OpenCLIPEmbeddings(
+        model_name="ViT-H-14", checkpoint="laion2b_s32b_b79k"
+    ),
+)
+```
+
+## LLM
+
+This template will use [Ollama](https://python.langchain.com/docs/integrations/chat/ollama#multi-modal).
+
+Download the latest version of Ollama: https://ollama.ai/
+
+Pull the an open source multi-modal LLM: e.g., https://ollama.ai/library/bakllava
+
+```
+ollama pull bakllava
+```
+
+The app is by default configured for `bakllava`. But you can change this in `chain.py` and `ingest.py` for different downloaded models.
+
+## Usage
+
+To use this package, you should first have the LangChain CLI installed:
+
+```shell
+pip install -U langchain-cli
+```
+
+To create a new LangChain project and install this as the only package, you can do:
+
+```shell
+langchain app new my-app --package rag-chroma-multi-modal
+```
+
+If you want to add this to an existing project, you can just run:
+
+```shell
+langchain app add rag-chroma-multi-modal
+```
+
+And add the following code to your `server.py` file:
+```python
+from rag_chroma_multi_modal import chain as rag_chroma_multi_modal_chain
+
+add_routes(app, rag_chroma_multi_modal_chain, path="/rag-chroma-multi-modal")
+```
+
+(Optional) Let's now configure LangSmith. 
+LangSmith will help us trace, monitor and debug LangChain applications. 
+LangSmith is currently in private beta, you can sign up [here](https://smith.langchain.com/). 
+If you don't have access, you can skip this section
+
+```shell
+export LANGCHAIN_TRACING_V2=true
+export LANGCHAIN_API_KEY=<your-api-key>
+export LANGCHAIN_PROJECT=<your-project>  # if not specified, defaults to "default"
+```
+
+If you are inside this directory, then you can spin up a LangServe instance directly by:
+
+```shell
+langchain serve
+```
+
+This will start the FastAPI app with a server is running locally at 
+[http://localhost:8000](http://localhost:8000)
+
+We can see all templates at [http://127.0.0.1:8000/docs](http://127.0.0.1:8000/docs)
+We can access the playground at [http://127.0.0.1:8000/rag-chroma-multi-modal/playground](http://127.0.0.1:8000/rag-chroma-multi-modal/playground)  
+
+We can access the template from code with:
+
+```python
+from langserve.client import RemoteRunnable
+
+runnable = RemoteRunnable("http://localhost:8000/rag-chroma-multi-modal")
+```

templates/rag-multi-modal-local/ingest.py

@@ -0,0 +1,35 @@
+import os
+from pathlib import Path
+
+from langchain.vectorstores import Chroma
+from langchain_experimental.open_clip import OpenCLIPEmbeddings
+
+# Load images
+img_dump_path = Path(__file__).parent / "docs/"
+rel_img_dump_path = img_dump_path.relative_to(Path.cwd())
+image_uris = sorted(
+    [
+        os.path.join(rel_img_dump_path, image_name)
+        for image_name in os.listdir(rel_img_dump_path)
+        if image_name.endswith(".jpg")
+    ]
+)
+
+# Index
+vectorstore = Path(__file__).parent / "chroma_db_multi_modal"
+re_vectorstore_path = vectorstore.relative_to(Path.cwd())
+
+# Load embedding function
+print("Loading embedding function")
+embedding = OpenCLIPEmbeddings(model_name="ViT-H-14", checkpoint="laion2b_s32b_b79k")
+
+# Create chroma
+vectorstore_mmembd = Chroma(
+    collection_name="multi-modal-rag",
+    persist_directory=str(Path(__file__).parent / "chroma_db_multi_modal"),
+    embedding_function=embedding,
+)
+
+# Add images
+print("Embedding images")
+vectorstore_mmembd.add_images(uris=image_uris)

templates/rag-multi-modal-local/pyproject.toml

@@ -0,0 +1,38 @@
+[tool.poetry]
+name = "rag-multi-modal-local"
+version = "0.1.0"
+description = "Multi-modal RAG using Chroma"
+authors = [
+    "Lance Martin <lance@langchain.dev>",
+]
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = ">=3.8.1,<4.0"
+langchain = ">=0.0.351"
+openai = "<2"
+tiktoken = ">=0.5.1"
+chromadb = ">=0.4.14"
+open-clip-torch = ">=2.23.0"
+torch = ">=2.1.0"
+langchain-experimental = "^0.0.43"
+langchain-community = ">=0.0.4"
+
+[tool.poetry.group.dev.dependencies]
+langchain-cli = ">=0.0.15"
+
+[tool.langserve]
+export_module = "rag_multi_modal_local"
+export_attr = "chain"
+
+[tool.templates-hub]
+use-case = "rag"
+author = "LangChain"
+integrations = ["Ollama", "Chroma"]
+tags = ["multi-modal"]
+
+[build-system]
+requires = [
+    "poetry-core",
+]
+build-backend = "poetry.core.masonry.api"

templates/rag-multi-modal-local/rag_multi_modal_local.ipynb

@@ -0,0 +1,52 @@
+{
+ "cells": [
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "id": "681a5d1e",
+   "metadata": {},
+   "source": [
+    "## Run Template\n",
+    "\n",
+    "In `server.py`, set -\n",
+    "```\n",
+    "add_routes(app, chain_rag_conv, path=\"/rag-multi-modal-local\")\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d774be2a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langserve.client import RemoteRunnable\n",
+    "\n",
+    "rag_app = RemoteRunnable(\"http://localhost:8001/rag-multi-modal-local\")\n",
+    "rag_app.invoke(\" < keywords here > \")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.16"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

templates/rag-multi-modal-local/rag_multi_modal_local/__init__.py

@@ -0,0 +1,3 @@
+from rag_multi_modal_local.chain import chain
+
+__all__ = ["chain"]

templates/rag-multi-modal-local/rag_multi_modal_local/chain.py

@@ -0,0 +1,122 @@
+import base64
+import io
+from pathlib import Path
+
+from langchain.chat_models import ChatOllama
+from langchain.vectorstores import Chroma
+from langchain_core.documents import Document
+from langchain_core.messages import HumanMessage
+from langchain_core.output_parsers import StrOutputParser
+from langchain_core.pydantic_v1 import BaseModel
+from langchain_core.runnables import RunnableLambda, RunnablePassthrough
+from langchain_experimental.open_clip import OpenCLIPEmbeddings
+from PIL import Image
+
+
+def resize_base64_image(base64_string, size=(128, 128)):
+    """
+    Resize an image encoded as a Base64 string.
+
+    :param base64_string: A Base64 encoded string of the image to be resized.
+    :param size: A tuple representing the new size (width, height) for the image.
+    :return: A Base64 encoded string of the resized image.
+    """
+    img_data = base64.b64decode(base64_string)
+    img = Image.open(io.BytesIO(img_data))
+    resized_img = img.resize(size, Image.LANCZOS)
+    buffered = io.BytesIO()
+    resized_img.save(buffered, format=img.format)
+    return base64.b64encode(buffered.getvalue()).decode("utf-8")
+
+
+def get_resized_images(docs):
+    """
+    Resize images from base64-encoded strings.
+
+    :param docs: A list of base64-encoded image to be resized.
+    :return: Dict containing a list of resized base64-encoded strings.
+    """
+    b64_images = []
+    for doc in docs:
+        if isinstance(doc, Document):
+            doc = doc.page_content
+        # Optional: re-size image
+        # resized_image = resize_base64_image(doc, size=(1280, 720))
+        b64_images.append(doc)
+    return {"images": b64_images}
+
+
+def img_prompt_func(data_dict, num_images=1):
+    """
+    GPT-4V prompt for image analysis.
+
+    :param data_dict: A dict with images and a user-provided question.
+    :param num_images: Number of images to include in the prompt.
+    :return: A list containing message objects for each image and the text prompt.
+    """
+    messages = []
+    if data_dict["context"]["images"]:
+        for image in data_dict["context"]["images"][:num_images]:
+            image_message = {
+                "type": "image_url",
+                "image_url": f"data:image/jpeg;base64,{image}",
+            }
+            messages.append(image_message)
+    text_message = {
+        "type": "text",
+        "text": (
+            "You are a helpful assistant that gives a description of food pictures.\n"
+            "Give a detailed summary of the image.\n"
+            "Give reccomendations for similar foods to try.\n"
+        ),
+    }
+    messages.append(text_message)
+    return [HumanMessage(content=messages)]
+
+
+def multi_modal_rag_chain(retriever):
+    """
+    Multi-modal RAG chain,
+
+    :param retriever: A function that retrieves the necessary context for the model.
+    :return: A chain of functions representing the multi-modal RAG process.
+    """
+    # Initialize the multi-modal Large Language Model with specific parameters
+    model = ChatOllama(model="bakllava", temperature=0)
+
+    # Define the RAG pipeline
+    chain = (
+        {
+            "context": retriever | RunnableLambda(get_resized_images),
+            "question": RunnablePassthrough(),
+        }
+        | RunnableLambda(img_prompt_func)
+        | model
+        | StrOutputParser()
+    )
+
+    return chain
+
+
+# Load chroma
+vectorstore_mmembd = Chroma(
+    collection_name="multi-modal-rag",
+    persist_directory=str(Path(__file__).parent.parent / "chroma_db_multi_modal"),
+    embedding_function=OpenCLIPEmbeddings(
+        model_name="ViT-H-14", checkpoint="laion2b_s32b_b79k"
+    ),
+)
+
+# Make retriever
+retriever_mmembd = vectorstore_mmembd.as_retriever()
+
+# Create RAG chain
+chain = multi_modal_rag_chain(retriever_mmembd)
+
+
+# Add typing for input
+class Question(BaseModel):
+    __root__: str
+
+
+chain = chain.with_types(input_type=Question)

templates/rag-multi-modal-mv-local/.gitignore

@@ -0,0 +1,3 @@
+docs/img_*.jpg
+chroma_db_multi_modal
+multi_vector_retriever_metadata

templates/rag-multi-modal-mv-local/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 LangChain, Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

templates/rag-multi-modal-mv-local/README.md

@@ -0,0 +1,118 @@
+
+# rag-multi-modal-mv-local
+
+Visual search is a famililar application to many with iPhones or Android devices: use natural language to search across your photo collection. 
+  
+With the release of open source, multi-modal LLMs it's possible to build this kind of application for yourself and have it run on your personal laptop. 
+
+This template demonstrates how to perform visual search and question-answering over a collection of photos.
+ 
+Given a set of photos, it will produce image summaries and index them, retrieve photos relevant to user question using the summaries, and use Ollama to run a local, open-source multi-modal LLM to answer questions about the retrieved photos.
+
+## Input
+
+Supply a set of photos in the `/docs` directory. 
+
+By default, this template has a toy collection of 3 food pictures.
+
+The app will look up and summarize photos based upon provided keywords or questions:
+```
+What kind of ice cream did I have?
+```
+
+In practice, a larger corpus of images can be tested.
+
+To create an index of the images, run:
+```
+poetry install
+python ingest.py
+```
+
+## Storage
+
+Here is the process the template will use to create an index of the slides (see [blog](https://blog.langchain.dev/multi-modal-rag-template/)):
+
+* Given a set of images
+* It uses a local multi-modal LLM ([bakllava](https://ollama.ai/library/bakllava)) to summarize each image
+* Embeds the image summaries with a link to the original images
+* Given a user question, it will relevant image(s) based on similarity between the image summary and user input (using Ollama embeddings)
+* It will pass those images to bakllava for answer synthesis
+
+By default, this will use [LocalFileStore](https://python.langchain.com/docs/integrations/stores/file_system) to store images and Chroma to store summaries.
+
+## LLM and Embedding Models
+
+We will use [Ollama](https://python.langchain.com/docs/integrations/chat/ollama#multi-modal) for generating image summaries, embeddings, and the final image QA.
+
+Download the latest version of Ollama: https://ollama.ai/
+
+Pull an open source multi-modal LLM: e.g., https://ollama.ai/library/bakllava
+
+Pull an open source embedding model: e.g., https://ollama.ai/library/llama2:7b
+
+```
+ollama pull bakllava
+ollama pull llama2:7b
+```
+
+The app is by default configured for `bakllava`. But you can change this in `chain.py` and `ingest.py` for different downloaded models.
+
+The app will retrieve images based on similarity between the text input and the image summary, and pass the images to `bakllava`.
+
+## Usage
+
+To use this package, you should first have the LangChain CLI installed:
+
+```shell
+pip install -U langchain-cli
+```
+
+To create a new LangChain project and install this as the only package, you can do:
+
+```shell
+langchain app new my-app --package rag-multi-modal-mv-local
+```
+
+If you want to add this to an existing project, you can just run:
+
+```shell
+langchain app add rag-multi-modal-mv-local
+```
+
+And add the following code to your `server.py` file:
+```python
+from rag_multi_modal_mv_local import chain as rag_multi_modal_mv_local_chain
+
+add_routes(app, rag_multi_modal_mv_local_chain, path="/rag-multi-modal-mv-local")
+```
+
+(Optional) Let's now configure LangSmith. 
+LangSmith will help us trace, monitor and debug LangChain applications. 
+LangSmith is currently in private beta, you can sign up [here](https://smith.langchain.com/). 
+If you don't have access, you can skip this section
+
+```shell
+export LANGCHAIN_TRACING_V2=true
+export LANGCHAIN_API_KEY=<your-api-key>
+export LANGCHAIN_PROJECT=<your-project>  # if not specified, defaults to "default"
+```
+
+If you are inside this directory, then you can spin up a LangServe instance directly by:
+
+```shell
+langchain serve
+```
+
+This will start the FastAPI app with a server is running locally at 
+[http://localhost:8000](http://localhost:8000)
+
+We can see all templates at [http://127.0.0.1:8000/docs](http://127.0.0.1:8000/docs)
+We can access the playground at [http://127.0.0.1:8000/rag-multi-modal-mv-local/playground](http://127.0.0.1:8000/rag-multi-modal-mv-local/playground)  
+
+We can access the template from code with:
+
+```python
+from langserve.client import RemoteRunnable
+
+runnable = RemoteRunnable("http://localhost:8000/rag-multi-modal-mv-local")
+```

templates/rag-multi-modal-mv-local/ingest.py

@@ -0,0 +1,192 @@
+import base64
+import io
+import os
+import uuid
+from io import BytesIO
+from pathlib import Path
+
+from langchain.chat_models import ChatOllama
+from langchain.embeddings import OllamaEmbeddings
+from langchain.retrievers.multi_vector import MultiVectorRetriever
+from langchain.schema.document import Document
+from langchain.schema.messages import HumanMessage
+from langchain.storage import LocalFileStore
+from langchain.vectorstores import Chroma
+from PIL import Image
+
+
+def image_summarize(img_base64, prompt):
+    """
+    Make image summary
+
+    :param img_base64: Base64 encoded string for image
+    :param prompt: Text prompt for summarizatiomn
+    :return: Image summarization prompt
+
+    """
+    chat = ChatOllama(model="bakllava", temperature=0)
+
+    msg = chat.invoke(
+        [
+            HumanMessage(
+                content=[
+                    {"type": "text", "text": prompt},
+                    {
+                        "type": "image_url",
+                        "image_url": f"data:image/jpeg;base64,{img_base64}",
+                    },
+                ]
+            )
+        ]
+    )
+    return msg.content
+
+
+def generate_img_summaries(img_base64_list):
+    """
+    Generate summaries for images
+
+    :param img_base64_list: Base64 encoded images
+    :return: List of image summaries and processed images
+    """
+
+    # Store image summaries
+    image_summaries = []
+    processed_images = []
+
+    # Prompt
+    prompt = """Give a detailed summary of the image."""
+
+    # Apply summarization to images
+    for i, base64_image in enumerate(img_base64_list):
+        try:
+            image_summaries.append(image_summarize(base64_image, prompt))
+            processed_images.append(base64_image)
+        except Exception as e:
+            print(f"Error with image {i+1}: {e}")
+
+    return image_summaries, processed_images
+
+
+def get_images(img_path):
+    """
+    Extract images.
+
+    :param img_path: A string representing the path to the images.
+    """
+    # Get image URIs
+    pil_images = [
+        Image.open(os.path.join(img_path, image_name))
+        for image_name in os.listdir(img_path)
+        if image_name.endswith(".jpg")
+    ]
+    return pil_images
+
+
+def resize_base64_image(base64_string, size=(128, 128)):
+    """
+    Resize an image encoded as a Base64 string
+
+    :param base64_string: Base64 string
+    :param size: Image size
+    :return: Re-sized Base64 string
+    """
+    # Decode the Base64 string
+    img_data = base64.b64decode(base64_string)
+    img = Image.open(io.BytesIO(img_data))
+
+    # Resize the image
+    resized_img = img.resize(size, Image.LANCZOS)
+
+    # Save the resized image to a bytes buffer
+    buffered = io.BytesIO()
+    resized_img.save(buffered, format=img.format)
+
+    # Encode the resized image to Base64
+    return base64.b64encode(buffered.getvalue()).decode("utf-8")
+
+
+def convert_to_base64(pil_image):
+    """
+    Convert PIL images to Base64 encoded strings
+
+    :param pil_image: PIL image
+    :return: Re-sized Base64 string
+    """
+
+    buffered = BytesIO()
+    pil_image.save(buffered, format="JPEG")  # You can change the format if needed
+    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
+    # img_str = resize_base64_image(img_str, size=(831,623))
+    return img_str
+
+
+def create_multi_vector_retriever(vectorstore, image_summaries, images):
+    """
+    Create retriever that indexes summaries, but returns raw images or texts
+
+    :param vectorstore: Vectorstore to store embedded image sumamries
+    :param image_summaries: Image summaries
+    :param images: Base64 encoded images
+    :return: Retriever
+    """
+
+    # Initialize the storage layer for images
+    store = LocalFileStore(
+        str(Path(__file__).parent / "multi_vector_retriever_metadata")
+    )
+    id_key = "doc_id"
+
+    # Create the multi-vector retriever
+    retriever = MultiVectorRetriever(
+        vectorstore=vectorstore,
+        byte_store=store,
+        id_key=id_key,
+    )
+
+    # Helper function to add documents to the vectorstore and docstore
+    def add_documents(retriever, doc_summaries, doc_contents):
+        doc_ids = [str(uuid.uuid4()) for _ in doc_contents]
+        summary_docs = [
+            Document(page_content=s, metadata={id_key: doc_ids[i]})
+            for i, s in enumerate(doc_summaries)
+        ]
+        retriever.vectorstore.add_documents(summary_docs)
+        retriever.docstore.mset(list(zip(doc_ids, doc_contents)))
+
+    add_documents(retriever, image_summaries, images)
+
+    return retriever
+
+
+# Load images
+doc_path = Path(__file__).parent / "docs/"
+rel_doc_path = doc_path.relative_to(Path.cwd())
+print("Read images")
+pil_images = get_images(rel_doc_path)
+
+# Convert to b64
+images_base_64 = [convert_to_base64(i) for i in pil_images]
+
+# Image summaries
+print("Generate image summaries")
+image_summaries, images_base_64_processed = generate_img_summaries(images_base_64)
+
+# The vectorstore to use to index the images summaries
+vectorstore_mvr = Chroma(
+    collection_name="image_summaries",
+    persist_directory=str(Path(__file__).parent / "chroma_db_multi_modal"),
+    embedding_function=OllamaEmbeddings(model="llama2:7b"),
+)
+
+# Create documents
+images_base_64_processed_documents = [
+    Document(page_content=i) for i in images_base_64_processed
+]
+
+# Create retriever
+retriever_multi_vector_img = create_multi_vector_retriever(
+    vectorstore_mvr,
+    image_summaries,
+    images_base_64_processed_documents,
+)

templates/rag-multi-modal-mv-local/pyproject.toml

@@ -0,0 +1,38 @@
+[tool.poetry]
+name = "rag-multi-modal-mv-local"
+version = "0.1.0"
+description = "Multi-modal RAG using Chroma and multi-vector retriever"
+authors = [
+    "Lance Martin <lance@langchain.dev>",
+]
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = ">=3.8.1,<4.0"
+langchain = ">=0.0.351"
+openai = "<2"
+tiktoken = ">=0.5.1"
+chromadb = ">=0.4.14"
+pypdfium2 = ">=4.20.0"
+langchain-experimental = "^0.0.43"
+pillow = ">=10.1.0"
+langchain-community = ">=0.0.4"
+
+[tool.poetry.group.dev.dependencies]
+langchain-cli = ">=0.0.15"
+
+[tool.langserve]
+export_module = "rag_multi_modal_mv_local"
+export_attr = "chain"
+
+[tool.templates-hub]
+use-case = "rag"
+author = "LangChain"
+integrations = ["Ollama", "Chroma"]
+tags = ["multi-modal"]
+
+[build-system]
+requires = [
+    "poetry-core",
+]
+build-backend = "poetry.core.masonry.api"

templates/rag-multi-modal-mv-local/rag-multi-modal-mv-local.ipynb

@@ -0,0 +1,52 @@
+{
+ "cells": [
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "id": "681a5d1e",
+   "metadata": {},
+   "source": [
+    "## Run Template\n",
+    "\n",
+    "In `server.py`, set -\n",
+    "```\n",
+    "add_routes(app, chain_rag_conv, path=\"/rag-multi-modal-mv-local\")\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d774be2a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langserve.client import RemoteRunnable\n",
+    "\n",
+    "rag_app = RemoteRunnable(\"http://localhost:8001/rag-multi-modal-mv-local\")\n",
+    "rag_app.invoke(\" < keywords here > \")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.16"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

templates/rag-multi-modal-mv-local/rag_multi_modal_mv_local/__init__.py

@@ -0,0 +1,3 @@
+from rag_multi_modal_mv_local.chain import chain
+
+__all__ = ["chain"]

templates/rag-multi-modal-mv-local/rag_multi_modal_mv_local/chain.py

@@ -0,0 +1,131 @@
+import base64
+import io
+from pathlib import Path
+
+from langchain.chat_models import ChatOllama
+from langchain.embeddings import OllamaEmbeddings
+from langchain.pydantic_v1 import BaseModel
+from langchain.retrievers.multi_vector import MultiVectorRetriever
+from langchain.schema.document import Document
+from langchain.schema.messages import HumanMessage
+from langchain.schema.output_parser import StrOutputParser
+from langchain.schema.runnable import RunnableLambda, RunnablePassthrough
+from langchain.storage import LocalFileStore
+from langchain.vectorstores import Chroma
+from PIL import Image
+
+
+def resize_base64_image(base64_string, size=(128, 128)):
+    """
+    Resize an image encoded as a Base64 string.
+
+    :param base64_string: A Base64 encoded string of the image to be resized.
+    :param size: A tuple representing the new size (width, height) for the image.
+    :return: A Base64 encoded string of the resized image.
+    """
+    img_data = base64.b64decode(base64_string)
+    img = Image.open(io.BytesIO(img_data))
+    resized_img = img.resize(size, Image.LANCZOS)
+    buffered = io.BytesIO()
+    resized_img.save(buffered, format=img.format)
+    return base64.b64encode(buffered.getvalue()).decode("utf-8")
+
+
+def get_resized_images(docs):
+    """
+    Resize images from base64-encoded strings.
+
+    :param docs: A list of base64-encoded image to be resized.
+    :return: Dict containing a list of resized base64-encoded strings.
+    """
+    b64_images = []
+    for doc in docs:
+        if isinstance(doc, Document):
+            doc = doc.page_content
+        # Optional: re-size image
+        # resized_image = resize_base64_image(doc, size=(1280, 720))
+        b64_images.append(doc)
+    return {"images": b64_images}
+
+
+def img_prompt_func(data_dict, num_images=1):
+    """
+    Ollama prompt for image analysis.
+
+    :param data_dict: A dict with images and a user-provided question.
+    :param num_images: Number of images to include in the prompt.
+    :return: A list containing message objects for each image and the text prompt.
+    """
+    messages = []
+    if data_dict["context"]["images"]:
+        for image in data_dict["context"]["images"][:num_images]:
+            image_message = {
+                "type": "image_url",
+                "image_url": f"data:image/jpeg;base64,{image}",
+            }
+            messages.append(image_message)
+    text_message = {
+        "type": "text",
+        "text": (
+            "You are a helpful assistant that gives a description of food pictures.\n"
+            "Give a detailed summary of the image.\n"
+        ),
+    }
+    messages.append(text_message)
+    return [HumanMessage(content=messages)]
+
+
+def multi_modal_rag_chain(retriever):
+    """
+    Multi-modal RAG chain,
+
+    :param retriever: A function that retrieves the necessary context for the model.
+    :return: A chain of functions representing the multi-modal RAG process.
+    """
+    # Initialize the multi-modal Large Language Model with specific parameters
+    model = ChatOllama(model="bakllava", temperature=0)
+
+    # Define the RAG pipeline
+    chain = (
+        {
+            "context": retriever | RunnableLambda(get_resized_images),
+            "question": RunnablePassthrough(),
+        }
+        | RunnableLambda(img_prompt_func)
+        | model
+        | StrOutputParser()
+    )
+
+    return chain
+
+
+# Load chroma
+vectorstore_mvr = Chroma(
+    collection_name="image_summaries",
+    persist_directory=str(Path(__file__).parent.parent / "chroma_db_multi_modal"),
+    embedding_function=OllamaEmbeddings(model="llama2:7b"),
+)
+
+# Load file store
+store = LocalFileStore(
+    str(Path(__file__).parent.parent / "multi_vector_retriever_metadata")
+)
+id_key = "doc_id"
+
+# Create the multi-vector retriever
+retriever = MultiVectorRetriever(
+    vectorstore=vectorstore_mvr,
+    byte_store=store,
+    id_key=id_key,
+)
+
+# Create RAG chain
+chain = multi_modal_rag_chain(retriever)
+
+
+# Add typing for input
+class Question(BaseModel):
+    __root__: str
+
+
+chain = chain.with_types(input_type=Question)