MCP Document Search System

A vector search system for document retrieval using MongoDB Atlas Vector Search and Voyage AI embeddings.

Sample data included is for Atlas Vector Search!

Features

• Ingests and chunks markdown documents with hierarchical headers
• Generates embeddings using Voyage AI's contextual embeddings API
• Stores documents and embeddings in MongoDB with parent-child relationships
• Provides a FastMCP server for semantic document search
• Supports configurable vector dimensions and chunking strategies

Available MCP Tools

The document search server provides these tools:

1. search_documents_vector(query: str, limit: int = 5)

• Primary search method using vector similarity
• Returns document chunks with metadata and similarity scores
• Best for semantic/meaning-based queries

2. search_documents_lexicaly(query: str, limit: int = 1)

• Fallback search using lexical/text matching
• Returns full parent documents with search scores
• Useful when vector search doesn't find good matches

3. get_parent_document(parent_id: str)

• Retrieves the complete parent document by ID
• Returns original content and file path
• Use after search to get full context for a chunk

Prerequisites

• Python 3.10+
• MongoDB Atlas cluster with vector search enabled
• Voyage AI API key

Installation

1. Clone the repository:

git clone https://github.com/patw/avs-document-search.git
cd avs-document-search

2. Install dependencies:

pip install -r requirements.txt

3. Create a .env file based on sample.env with your credentials

Usage

1. Ingest documents in the docs/ directory:

python ingest_docs.py

2. Run the search server:

python avs-mcp.py

Running the search server won't do much, other than verify your MongoDB URI is correct, you will need to plug this MCP server into an MCP client like Claude Desktop. Here's a sample config:

{
  "mcpServers": {
    "Atlas Vector Search Docs": {
      "command": "uv",
      "args": [
        "run",
        "--with",
        "fastmcp, pymongo, requests",
        "fastmcp",
        "run",
        "/avs-docs-mcp/avs-mcp.py"
      ]
    }
  }
}

Configuration

Copy sample.env to .env and Edit to configure:

• MongoDB connection string
• Database and collection names
• Voyage AI API key
• Vector dimensions (256 default)

Future Improvements

• Implement hybrid search combining vector and text search using $rankFusion (when MongoDB 8.1 is GA on Atlas)
• Support additional file formats (PDF, Word, etc.) with Docling

Contributing

Pull requests are welcome! For major changes, please open an issue first.

Author

Pat Wendorf

pat.wendorf@mongodb.com

GitHub: patw

License

MIT