Local Docs MCP

A modular semantic search system with MCP (Model Context Protocol) integration for searching local documentation. The Retrieval-Augmented Generation (RAG) system not only lets you manage document chunks for knowledge retrieval but also gives AI assistants semantic search capabilities through the MCP.

Key Features

Related MCP server: RAG Database MCP Server

Quick Start

Installation

1. Clone and setup the project:

git clone git@github.com:nguyenchiencong/local-docs-mcp.git
cd local-docs-mcp

2. Start required services:

# Start Qdrant
docker run -d -p 6334:6334 -p 6333:6333 qdrant/qdrant

# Make sure Ollama is running with the embedding model
ollama pull hf.co/Qwen/Qwen3-Embedding-0.6B-GGUF:F16

# Setup postgres for cocoindex
docker compose -f <(curl -L https://raw.githubusercontent.com/cocoindex-io/cocoindex/refs/heads/main/dev/postgres.yaml) up -d

3. Configure environment:

# Edit .env with your specific configuration
cp .env.example .env

# Don't forget to setup your .cocoignore file
cp .cocoignore.example .cocoignore

4. Install dependencies:

uv sync

Usage

Before indexing, add your documents to the docs folder.

To index your documents:

uv run python -m src.indexing.main_flow
# To run the force reindex utility
uv run python -m src.indexing.force_reindex

To start the MCP server:

uv run python -m src.mcp_server.server

Make the CLI available on your PATH

If you want to run local-docs-mcp from any directory:

• Windows (PowerShell):

  setx PATH "path\to\local-docs-mcp\.venv\Scripts;$($env:Path)"

• Linux/macOS (bash/zsh):

  echo 'export PATH="/path/to/local-docs-mcp/.venv/bin:$PATH"' >> ~/.bashrc

To run MCP tools directly from the CLI (one-off calls):

# Start the server (default behavior)
local-docs-mcp

# Run a semantic search once and exit
local-docs-mcp semantic_search --query "vector search overview" --limit 5

# Run hybrid search
local-docs-mcp hybrid_search --query "async await" --semantic-weight 0.7 --limit 5

# Run a filtered search with metadata (JSON object)
local-docs-mcp search_with_metadata_filter --query "UI tutorial" --metadata-filter '{"filename": "ui.md"}' --limit 5

# Retrieve a specific document by ID
local-docs-mcp document_retrieval --document-id "doc-123"

# Fetch collection info
local-docs-mcp get_collection_info --json

Configuration

System Settings

All configuration is managed in pyproject.toml under the [tool.local-docs] section:

[tool.local-docs]
# Qdrant configuration
qdrant_url = "http://localhost:6334"
qdrant_collection = "local-docs-collection"

# Ollama configuration
ollama_url = "http://localhost:11434"
ollama_model = "hf.co/Qwen/Qwen3-Embedding-0.6B-GGUF:F16"
embedding_dimension = 1024

# Document configuration
docs_directory = "docs"
supported_extensions = [".md", ".rst", ".txt"]

# Search configuration
search_limit = 10

# Chunking configuration
chunk_size = 1200
chunk_overlap = 200

# Search configuration
search_limit = 10
similarity_threshold = 0.15
search_hnsw_ef = 256
hybrid_semantic_weight = 0.85
mmr_lambda = 0.75

Environment Variables: Override any setting with LOCAL_DOCS_* environment variables:

export LOCAL_DOCS_SEARCH_LIMIT=20
export LOCAL_DOCS_OLLAMA_MODEL="different-model"

MCP Client Setup

Add this to your MCP client configuration (e.g., Claude Code):

{
  "mcpServers": {
    "local-docs-mcp": {
      "command": "uv",
      "args": ["run", "--project", "/path/to/local-docs-mcp", "-m", "src.mcp_server.server"]
    }
  }
}

MCP Tools

The MCP server exposes the following semantic search tools to AI assistants:

Use cases

Documentation Research:

• "What are signals and how do they work in Godot?"

• "Find tutorials about character controllers"

• "Explain the difference between KinematicBody and RigidBody"

Problem Solving:

• "How do I fix 'node not found' errors?"

• "What are the best practices for performance optimization?"

• "Search for debugging techniques in Godot"

Learning Paths:

• "I'm a beginner, show me getting started content"

• "What should I learn after basic GDScript?"

• "Find intermediate tutorials about physics"

Specific Searches:

• "Show me the top 5 most relevant results about animations"

• "Find only tutorial files about UI design"

• "Look for performance optimization guides"

Development

Running Tests

uv run pytest tests/

Contributing

1. Fork the repository

2. Create a feature branch

3. Add tests for new functionality

4. Submit a pull request

License

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

Related Projects

• CocoIndex - Document indexing and processing

• Qdrant - Vector database for similarity search

• Ollama - Local AI model serving

• Model Context Protocol - Standard for AI tool integration