MCP-Markdown-RAG

<div align="center"> <img src="docs/banner.png" alt="MCP-Markdown-RAG" width="800" style="border-radius:10px;"/> <h1>MCP-Markdown-RAG</h1> <p> <img alt="GitHub forks" src="https://img.shields.io/github/forks/Zackriya-Solutions/MCP-Markdown-RAG"/> <img alt="GitHub Repo stars" src="https://img.shields.io/github/stars/Zackriya-Solutions/MCP-Markdown-RAG"> <img alt="GitHub last commit" src="https://img.shields.io/github/last-commit/Zackriya-Solutions/MCP-Markdown-RAG"> </p> <p> <a href="LICENSE"> <img src="https://img.shields.io/badge/License-Apache%202.0-green" alt="License" /> </a> <img src="https://img.shields.io/badge/MCP-Server-blue"/> </p> </div>

A Model Context Protocol (MCP) server that provides a local-first RAG engine for your markdown documents. This server uses a file-based Milvus vector database to index your notes, enabling Large Language Models (LLMs) to perform semantic search and retrieve relevant content from your local files.

[!NOTE] This project is in active development. The API and implementation are subject to change. We are exploring future enhancements, including a potential port to an Obsidian plugin for seamless vault integration.

🎯 Key Features

✅ Local-First & Private: All your data is processed and stored locally. Nothing is sent to a third-party service for indexing.

✅ Semantic Search for Markdown: Go beyond simple keyword search. Find document sections based on conceptual meaning.

✅ MCP Compatible: Integrates with any MCP-supported host application like Claude Desktop, Windsurf, or Cursor.

✅ Simple Tooling: Provides two straightforward tools (index_documents and search) for managing and querying your knowledge base.

⚙️ How It Works

The server operates in two main phases, exposing its functionality through MCP tools.

1. Indexing:

   • The index_documents tool is called with a path to your markdown files.
   • The server reads the documents, splits them into logical chunks (e.g., by headings), and converts each chunk into a vector embedding.
   • These embeddings, along with their metadata (original text, file path), are stored in a local Milvus vector database.
   • You can run it in two modes:
     • Full Reindex (force_reindex=True): Clears and rebuilds the entire index from scratch.
     • Incremental Update (force_reindex=False, default): Automatically detects and re-indexes only changed files by comparing them against a tracking log. Deleted or modified chunks are pruned and replaced to keep the index up-to-date.
     • Recursive Indexing (recursive=False, default): Recursively indexes all subdirectories.

2. Searching:

   • When you ask a question in a host application, it uses the search tool.
   • The server converts your query into a vector embedding.
   • It then performs a similarity search against the Milvus database to find the most semantically relevant document chunks.
   • The results are returned to the LLM, providing it with the context needed to answer your question accurately.

   <div align="center" > <img src="docs/mcp_search.png" alt="MCP Search" width="800" style="border-radius:10px;"/> </div>

🛠️ Available Tools

• index_documents

  • Description: Indexes Markdown documents for semantic search. Converts each file into structured vector chunks and inserts them into the Milvus database.
  • Incremental Indexing: Automatically reindexes only changed files unless force_reindex=True is passed.
  • Arguments:
    • directory (string, optional): The path to the folder containing .md files. Defaults to current directory.
    • force_reindex (boolean, optional): If True, clears and rebuilds the full index. Defaults to False.
    • recursive (boolean, optional): If True, recursively indexes all subdirectories. Defaults to False.

• search

  • Description: Searches the indexed documents using semantic similarity.
  • Arguments:
    • query (string, required): Your natural language query.
    • limit (integer, optional): Max number of chunks to return (default is usually 5–10).

🚀 Installation & Setup

This server requires UV (for running the Python server).

Step 1: Get the Server Code

Clone this repository to your local machine:

git clone https://github.com/Zackriya-Solutions/MCP-Markdown-RAG.git

Step 2: Configure Your Host App

Configure your MCP host application (e.g., Windsurf, Claude.app) to use the server. Add the following to your settings file:

{
  "mcpServers": {
    "markdown_rag": {
      "command": "uv",
      "args": [
        "--directory",
        "/ABSOLUTE/PATH/TO/MCP-Markdown-RAG",
        "run",
        "server.py"
      ]
    }
  }
}

Note: Replace /ABSOLUTE/PATH/TO/MCP-Markdown-RAG with the absolute path to where you cloned this repository.

Note: The first run will take a while and the same for the first indexing, as it needs to download the embedding model(~50MB).

📈 What's Next? (Roadmap)

We are actively working on improving the server. Future plans include:

• Performance Optimization: Improve indexing by encoding inputs in batches, which should better manage CPU usage.
• Flexible Embedding Models: Add support for other embedding models, such as the BGEM3-large model for potentially higher accuracy.
• Obsidian Plugin: Explore creating a dedicated Obsidian plugin for a fully integrated experience.

🐛 Debugging

You can use the MCP inspector to debug the server directly. Run the following command from the repository's root directory:

npx @modelcontextprotocol/inspector uv --directory /ABSOLUTE/PATH/TO/MCP-Markdown-RAG run server.py

🤝 Contributing

Contributions are welcome! Please feel free to open an issue or submit a pull request.

🙏 Acknowledgments

• The Model Context Protocol for the open standard that makes this possible.
• The Milvus Project for the powerful open-source vector database.