RAG - Ollama, Chroma - multi-modal, multi-vector, local

Visual search is a familiar application to many with iPhones or Android devices. It allows user to search photos using natural language.

With the release of open source, multi-modal LLMs it's possible to build this kind of application for yourself for your own private photo collection.

This template demonstrates how to perform private visual search and question-answering over a collection of your photos.

It uses an open source multi-modal LLM of your choice to create image summaries for each photos, embeds the summaries, and stores them in Chroma.

Given a question, relevant photos are retrieved and passed to the multi-modal LLM for answer synthesis.

"Visual Search Process Diagram"

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):

• Given a set of images
• It uses a local multi-modal LLM (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 to store images and Chroma to store summaries.

LLM and Embedding Models

We will use Ollama 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:

pip install -U langchain-cli

To create a new LangChain project and install this as the only package, you can do:

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:

langchain app add rag-multi-modal-mv-local

And add the following code to your server.py file:

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. You can sign up for LangSmith here. If you don't have access, you can skip this section

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:

langchain serve

This will start the FastAPI app with a server is running locally at http://localhost:8000

We can see all templates at 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

We can access the template from code with:

from langserve.client import RemoteRunnable

runnable = RemoteRunnable("http://localhost:8000/rag-multi-modal-mv-local")