MongoDB Documentation MCP

MongoDB semantic documentation search server implementing the Model Context Protocol (MCP) for IDE integration with Claude and Cursor.

Transforms MongoDB documentation into searchable vector embeddings using Voyage AI's voyage-context-3 model (1024 dimensions) and MongoDB Atlas Vector Search, enabling natural language queries against comprehensive MongoDB documentation corpus.

Technical Architecture

• Embedding Model: voyage-context-3 (1024-dimensional vectors) - optimized for technical documentation
• Vector Database: MongoDB Atlas Vector Search with cosine similarity indexing
• Search Algorithms: RRF (Reciprocal Rank Fusion) hybrid + MMR (Maximum Marginal Relevance)
• Smart Indexing: Git commit hash tracking with incremental processing
• Document Coverage: 10,596 indexed documents from 5 MongoDB repositories
• MCP Protocol: Clean 2-tool architecture following Context7 patterns

Prerequisites

• Node.js 18+
• MongoDB Atlas (free M0 cluster sufficient)
• Voyage AI API key
• Claude Desktop or Cursor IDE

Quick Start

1. Get API Keys

MongoDB Atlas:

1. Create free cluster (M0 tier)
2. Get connection string: Database → Connect → Drivers
3. Copy URI starting with mongodb+srv://

Voyage AI:

1. Get API key
2. Copy key starting with pa-

2. Install & Index

# Install globally
npm install -g mongodocs-mcp

# Set credentials
export MONGODB_URI="mongodb+srv://your-connection-string"
export VOYAGE_API_KEY="pa-your-api-key"

# Initial indexing (10-15 minutes)
mongodocs-index

3. Configure IDE

For Cursor IDE - Add to .cursor/settings.json:

{
  "mcpServers": {
    "mongodocs": {
      "command": "npx",
      "args": ["mongodocs-mcp"],
      "env": {
        "MONGODB_URI": "mongodb+srv://your-connection-string",
        "VOYAGE_API_KEY": "pa-your-api-key"
      }
    }
  }
}

For Claude Desktop - Add to MCP settings:

{
  "mcpServers": {
    "mongodocs": {
      "command": "mongodocs-mcp",
      "env": {
        "MONGODB_URI": "mongodb+srv://your-connection-string", 
        "VOYAGE_API_KEY": "pa-your-api-key"
      }
    }
  }
}

Restart your IDE after configuration.

Usage

Query MongoDB documentation using natural language through your IDE:

How do I implement vector search in MongoDB?
Show me RAG examples with MongoDB Atlas
What's the syntax for compound indexes?
How to optimize aggregation performance?

MCP Tool Architecture:

• mongodb-search: RRF hybrid algorithm (vector + keyword fusion) - primary search method
• mongodb-mmr-search: Maximum Marginal Relevance algorithm - diverse result optimization

Commands

mongodocs-index           # Smart indexing (default)
mongodocs-index rebuild   # Force full rebuild  
mongodocs-index clean     # Clear database
mongodocs-index stats     # Show statistics
mongodocs-index test      # Test embedding service

Incremental Indexing

Initial Processing:

🔄 Smart indexing - checking for repository changes...
🆕 MongoDB Official Documentation first time indexing...

Delta Updates:

🔄 Smart indexing - checking for repository changes...
✅ MongoDB Official Documentation up to date (a1b2c3d4), skipping...
✅ MongoDB Manual up to date (e5f6g7h8), skipping...
✅ Smart update complete!

Git commit hash comparison enables differential processing - only changed documents trigger re-embedding and storage operations.

Implementation Details

Vector Embedding Pipeline:

• Model: voyage-context-3 (1024 dimensions, optimized for documentation)
• Chunking: Semantic content splitting (100-2000 characters)
• Normalization: Cosine similarity with L2 normalization

MongoDB Atlas Configuration:

• Vector Index: Cosine similarity, 40 numCandidates (7.5x speed optimization)
• Text Index: Lucene analyzers for keyword search
• Storage: Bulk upsert operations with graceful error handling

Search Algorithm Implementation:

• RRF (Reciprocal Rank Fusion): Combined vector + keyword ranking with tuned weights
• MMR (Maximum Marginal Relevance): Diversity optimization using cosine similarity calculations
• Performance: Sub-500ms response times with optimized candidate selection

Data Processing:

• Git-based incremental updates prevent redundant processing
• Token validation ensures embedding model compatibility
• Batch processing with p-limit concurrency control

Document Corpus

Repository Sources:

1. mongodb/docs - Official MongoDB documentation
2. mongodb/docs-generative-ai-showcase - Vector search examples
3. mongodb/atlas-search-playground-chatbot-starter - RAG implementations
4. mongodb/drivers-examples - Language-specific implementations
5. mongodb-developer/code-examples - Community patterns

Processing Statistics:

• Documents: 10,596 indexed chunks
• Storage: ~200MB in MongoDB Atlas
• Coverage: Complete MongoDB ecosystem documentation

Performance Metrics

Indexing Operations:

• Initial processing: 10-15 minutes (10,596 documents)
• Incremental updates: 30 seconds - 2 minutes
• Memory utilization: <100MB during batch processing

Search Performance:

• Query latency: <500ms (99th percentile)
• Vector search: 40 candidates vs 300 default (7.5x optimization)
• Storage footprint: ~200MB in MongoDB Atlas

Troubleshooting

No search results:

mongodocs-index stats  # Check if database is populated

Dimension errors:

mongodocs-index clean && mongodocs-index  # Fresh rebuild

Connection issues:

1. Verify MongoDB URI format
2. Check Atlas Network Access (add your IP)
3. Confirm database user permissions

Development

git clone https://github.com/romiluz/mongodocs-mcp.git
cd mongodocs-mcp
npm install
npm run build
npm start

License

MIT License - see LICENSE file

Support

• Issues: GitHub Issues
• Author: Rom Iluz
• Version: 10.1.3 - Production-Ready Vector Search MCP