🗄️ LanceDB MCP Server for LLMS

View original on GitHub 

A Model Context Protocol (MCP) server that enables LLMs to interact directly the documents that they have on-disk through agentic RAG and hybrid search in LanceDB. Ask LLMs questions about the dataset as a whole or about specific documents.

✨ Features

• 🔍 LanceDB-powered serverless vector index and document summary catalog.
• 📊 Efficient use of LLM tokens. The LLM itself looks up what it needs when it needs.
• 📈 Security. The index is stored locally so no data is transferred to the Cloud when using a local LLM.

🚀 Quick Start

To get started, create a local directory to store the index and add this configuration to your Claude Desktop config file:

MacOS: ~/Library/Application\ Support/Claude/claude_desktop_config.json
Windows: %APPDATA%/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "lancedb": {
      "command": "npx",
      "args": [
        "lance-mcp",
        "PATH_TO_LOCAL_INDEX_DIR"
      ]
    }
  }
}

Prerequisites

• Node.js 18+
• npx
• MCP Client (Claude Desktop App for example)
• Summarization and embedding models installed (see config.ts - by default we use Ollama models)
  • ollama pull snowflake-arctic-embed2
  • ollama pull llama3.1:8b

Demo

Local Development Mode:

{
  "mcpServers": {
    "lancedb": {
      "command": "node",
      "args": [
        "PATH_TO_LANCE_MCP/dist/index.js",
        "PATH_TO_LOCAL_INDEX_DIR"
      ]
    }
  }
}

Use npm run build to build the project.

Use npx @modelcontextprotocol/inspector dist/index.js PATH_TO_LOCAL_INDEX_DIR to run the MCP tool inspector.

Seed Data

The seed script creates two tables in LanceDB - one for the catalog of document summaries, and another one - for vectorized documents’ chunks. To run the seed script use the following command:

npm run seed -- --dbpath <PATH_TO_LOCAL_INDEX_DIR> --filesdir <PATH_TO_DOCS>

You can use sample data from the docs/ directory. Feel free to adjust the default summarization and embedding models in the config.ts file. If you need to recreate the index, simply rerun the seed script with the --overwrite option.

Catalog

• Document summary
• Metadata

Chunks

• Vectorized document chunk
• Metadata

🎯 Example Prompts

Try these prompts with Claude to explore the functionality:

"What documents do we have in the catalog?"
"Why is the US healthcare system so broken?"

📝 Available Tools

The server provides these tools for interaction with the index:

Catalog Tools

• catalog_search: Search for relevant documents in the catalog

Chunks Tools

• chunks_search: Find relevant chunks based on a specific document from the catalog
• all_chunks_search: Find relevant chunks from all known documents

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.