Vector Search and RAG Using MongoDB Atlas + Embedding Models + LLMs

About

This repo has sample code showcasing building Vector Search / RAG (Retrieval-Augmented Generation) applications using built-in Vector Search capablities of MongoDB Atlas, embedding models and LLMs (Large Language Models).

What is Vector Search?

vector search explained

Hackathon specific notes

• 2024-03-23 - Mistral AI hackathon

Labs

Setup: Setup Python Environment

Follow setup-python-env.md

Lab-1: Connect to MongoDB Atlas

Setup Atlas in the cloud and make sure we can connect to it.

Lab-1

Lab-2 - Vector Search Using OpenAI Embeddings

Perform vector search on an already indexed collection. This collection is pre-populated with embeddings using an OpenAI embedding model.

lab-2

Lab-3: Vector Search Using Custom Embeddings

We will populate collections data with custom embeddings, using open source embedding models and query them.

lab-3

Sample streamlit app

streamlit app

screencast | screenshot 1 | screenshot 2

Lab-4: RAG (Retrieval Augmentation Generation)

Index PDF files and store the index in Atlas with embeddings, and ask questions about the documents using LLMs

lab-4

Dockerizing and Deploying the App

dockerize.md

Some Fun Benchmarks

Vector search results using different embedding models

Local embedding models benchmark

LLMs performance on RAG

Useful Resources

• RAG Series Part 1: How to Choose the Right Embedding Model for Your Application by Apoorva Joshi