Notes RAG MCP Server

A Model Context Protocol (MCP) server for semantic search over markdown notes and documentation.

This server indexes your markdown notes into a Qdrant vector database using embeddings (defaulting to LiteLLM), and provides semantic search capabilities as an MCP tool, allowing LLMs to seamlessly search and retrieve context from your personal knowledge base.

Features

• MCP Integration: Exposes search_notes, trigger_reindex, and index_status tools.

• Smart Chunking: Splits markdown files by headings, ensuring context is preserved in chunks.

• Incremental Indexing: Uses SQLite to cache file modification times and Qdrant points, only re-indexing changed files.

• Web Dashboard: Includes an admin dashboard at /admin for managing indexed paths, viewing stats, and triggering re-indexing manually.

• Parallel Processing: Speeds up indexing using concurrent thread pools.

• Frontmatter Support: Parses YAML frontmatter for tags, categories, and other metadata to enrich the vector payload.

Related MCP server: mcp-recall-md

Tools Exposed

• search_notes: Perform semantic search on markdown notes and documentation. Filters by query, folder, tag, and category.

• trigger_reindex: Force an immediate directory scan to index new or updated files.

• index_status: Get indexing statistics of the vault.

Environment Variables

• QDRANT_URL: URL to the Qdrant instance (default: http://qdrant:6333).

• LITELLM_URL: URL to LiteLLM instance for embeddings (default: http://litellm:4000/v1).

• LITELLM_API_KEY: API key for embeddings (default: dummy).

• EMBEDDING_MODEL: Embedding model to use (default: text-embedding-3-small).

• COLLECTION_NAME: Qdrant collection name (default: notes_rag).

• VAULT_PATH: Default path to the markdown notes vault (default: /containers/productivity/obsidian/shared).

• CACHE_DB_PATH: Path to SQLite index cache DB (default: /app/data/index_cache.db).

• CHUNK_SIZE: Max characters per chunk (default: 1500).

• CHUNK_OVERLAP: Overlap between chunks (default: 200).

Running

Build and run the Docker container. Make sure to mount your markdown notes directory and a path for the persistent cache database:

docker build -t notes-rag-mcp .

# Replace /path/to/notes with the actual path to your markdown files
docker run -d \
  -p 3000:3000 \
  --env-file .env \
  -v /path/to/notes:/notes:ro \
  -v ./data:/app/data \
  notes-rag-mcp

Note: Make sure to update the VAULT_PATH environment variable to match the internal mounted path (e.g. /notes).

Connecting MCP Clients

Since this server runs via HTTP with Server-Sent Events (SSE), you can configure your AI client (like Claude Desktop or Gemini) to connect to the /sse endpoint.

Example configuration for an MCP client settings.json:

{
  "mcpServers": {
    "notes-rag": {
      "url": "http://localhost:3000/sse",
      "type": "sse",
      "trust": true
    }
  }
}

Changelog

• v1.0.1: Updated Python MCP SDK SSE transport method from connect_retrying to connect_sse to fix AttributeError during initialization and ensure compatibility with modern MCP clients.