MongoDB RAG

A type-safe MongoDB implementation for Retrieval Augmented Generation with vector search

Overview

MongoDB RAG is a semantic memory system for AI applications that provides persistent storage and retrieval of context-aware knowledge using MongoDB Atlas. It enables AI systems to remember past interactions, learn from them, and provide more personalized and context-aware responses over time.

This package implements a RAG (Retrieval Augmented Generation) system that:

• Stores text content with vector embeddings in MongoDB
• Enables semantic search based on meaning, not just keywords
• Organizes memories by user, agent, and session contexts
• Allows rich metadata and categorization
• Supports automatic expiration for temporary memories
• Prevents duplicate content to maintain database efficiency

Check out the Changelog to see what changed in the last releases.

Features

• Vector Storage: Store and retrieve text with vector embeddings using MongoDB Atlas
• Semantic Search: Find semantically similar content using MongoDB's vector search
• Flexible Organization: Categorize memories with user, agent, session IDs and custom metadata
• Rich Filtering: Filter memories by various attributes including custom metadata
• Memory Expiration: Set automatic expiration dates for temporary memories
• Type Safety: Fully typed API with TypeScript
• Batch Operations: Support for batch updates and deletes
• Pagination: Handle large datasets with efficient pagination
• Duplicate Prevention: Automatically checks and prevents storing duplicate content
• Automatic Collections: Creates required collections and indexes if they don't exist

Installation

Prerequisites

• MongoDB Atlas account with vector search capabilities
• Node.js v16+ or Bun
• Google Gemini API key for embeddings and fact extraction

Install with npm:

npm install mongo-rag

Install with bun:

bun add mongo-rag

Environment Setup

Set up the required environment variables:

# MongoDB connection string
MONGO_URI=mongodb+srv://username:[email protected]/database

# Gemini API key for embeddings and fact extraction
GEMINI_API_KEY=your_gemini_api_key

Quick Start

1. Connect to MongoDB

import mongoose from 'mongoose'

// Connect to MongoDB
await mongoose.connect(process.env.MONGO_URI)

2. Initialize the Client

import { MongoRagClient } from 'mongo-rag'

const client = new MongoRagClient({
  gemini_api_key: process.env.GEMINI_API_KEY,
})

3. Add Memories

// Add a simple memory string
const stringMemory = await client.add('User prefers vegetarian food', {
  user_id: 'user123',
  categories: ['preferences', 'food'],
})

// Add from chat messages
const messages = [
  { role: 'user', content: 'I like science fiction books' },
  { role: 'assistant', content: "I'll recommend sci-fi titles then!" },
]

const messageMemory = await client.add(messages, {
  user_id: 'user123',
  agent_id: 'books_agent',
  categories: ['preferences', 'books'],
  metadata: { genre: 'science fiction' },
})

// Add with expiration date
const temporaryMemory = await client.add(
  'User is currently looking for a birthday gift',
  {
    user_id: 'user123',
    // Memory expires in 7 days
    expiration_date: new Date(Date.now() + 7 * 24 * 60 * 60 * 1000),
  }
)

4. Search for Memories

// Semantic search
const searchResults = await client.search('What books does the user like?', {
  user_id: 'user123',
})

// With filters
const filteredResults = await client.search('What food preferences?', {
  user_id: 'user123',
  categories: ['preferences'],
})

// Advanced filtering
const advancedResults = await client.search('User interests', {
  filters: {
    AND: [
      { user_id: 'user123' },
      {
        created_at: {
          gte: new Date(Date.now() - 30 * 24 * 60 * 60 * 1000),
        },
      },
    ],
  },
})

5. Retrieve All Memories

// Get all memories for a user
const allMemories = await client.getAll({ user_id: 'user123' })

// Get memories by category
const foodMemories = await client.getAll({
  user_id: 'user123',
  categories: ['food'],
})

// With pagination
const paginatedMemories = await client.getAll({
  user_id: 'user123',
  page: 1,
  page_size: 10,
})

API Overview

The MongoRagClient provides these primary methods:

• add(content, options): Add a new memory
• search(query, options): Find semantically similar memories
• getAll(options): Retrieve memories with filters and pagination
• get(id): Get a specific memory by ID
• update(id, content, options): Update an existing memory
• delete(id): Delete a specific memory
• deleteAll(options): Delete memories matching criteria
• batchUpdate(updates): Update multiple memories at once
• batchDelete(deletes): Delete multiple memories at once

See the API Documentation for complete details.

Development

Setup

# Clone the repository
git clone https://github.com/mguleryuz/mongo-rag.git
cd mongo-rag

# Install dependencies
bun i

Testing

# Set environment variables for testing
export MONGO_URI=mongodb+srv://...
export GEMINI_API_KEY=your_gemini_api_key
export NODE_ENV=test

# Run tests
bun test

Watching TS Problems:

bun watch

How to make a release

For the Maintainer: Add NPM_TOKEN to the GitHub Secrets.

1. PR with changes
2. Merge PR into main
3. Checkout main
4. git pull
5. bun release: '' | alpha | beta optionally add -- --release-as minor | major | 0.0.1
6. Make sure everything looks good (e.g. in CHANGELOG.md)
7. Lastly run bun release:pub
8. Done

License

This package is licensed - see the LICENSE file for details.