DocRAG - AI Documentation RAG System

A lightweight, installable Python package that provides RAG (Retrieval Augmented Generation) access to technical documentation through an MCP (Model Context Protocol) server. This enables LLMs to search and retrieve relevant documentation on-demand.

Features

• 🚀 Single pip-installable package with CLI and MCP server

• 📚 Project-based documentation collections (BrightSign, Venafi, Qumu, web frameworks)

• 🔍 Local vector database with efficient embedding using LanceDB

• 📥 Easy documentation ingestion from local files or scraped sources

• 🤖 Designed for use with Claude Code via MCP

Installation

Prerequisites

• Python 3.10+

• pipx (recommended) or pip

• git (for updates)

Recommended: Install globally with pipx

# Install globally with pipx in editable mode (keeps dependencies isolated)
pipx install -e /opt/claude-ops/doc-rag

# Verify installation
docrag --help

# Optional: Install Playwright browsers (for scraping)
pipx runpip docrag install playwright
pipx run --spec docrag playwright install chromium

Note: The -e flag installs in "editable" mode, which means changes to the source code are immediately reflected without reinstalling.

Alternative: Install from source (development)

# Clone or navigate to the project directory
cd /opt/claude-ops/doc-rag

# Create and activate virtual environment
python3 -m venv venv
source venv/bin/activate

# Install in development mode
pip install -e ".[dev]"

# Install Playwright browsers (for scraping)
playwright install chromium

Updating DocRAG

Option 1: Using the Update Script (Recommended)

cd /opt/claude-ops/doc-rag
./update.sh

This script will:

• Pull latest changes from git

• Detect your installation method (pipx or pip)

• Reinstall only if necessary (non-editable installs)

• Handle editable installs automatically

Option 2: Using Make

cd /opt/claude-ops/doc-rag
make update

Option 3: Manual Update

For editable installs (installed with -e):

cd /opt/claude-ops/doc-rag
git pull origin main
# No reinstall needed - changes are already active!

For regular installs (installed without -e):

cd /opt/claude-ops/doc-rag
git pull origin main
pipx uninstall docrag && pipx install -e .
# or for pip: pip install -e . --force-reinstall

Verifying Updates

# Check git status
cd /opt/claude-ops/doc-rag
git log -1 --oneline

# Test the installation
docrag --version
docrag --help

Quick Start

1. Initialize DocRAG

docrag init

This creates the configuration directory at ~/.docrag/ with the following structure:

~/.docrag/
├── config.json           # Global configuration
├── collections/          # Documentation collections
└── vectordb/            # LanceDB storage

2. Add a Documentation Collection

# Add documentation from a local directory
docrag add brightsign --source /path/to/brightsign/docs --description "BrightSign player documentation"

# Or add without source initially
docrag add venafi --description "Venafi TPP API documentation"

3. List Collections

docrag list

4. Search Documentation (CLI Testing)

# Search across all active collections
docrag search "how to initialize the player"

# Search a specific collection
docrag search "authentication methods" --collection venafi --limit 10

5. Start the MCP Server

docrag serve

The server will listen on stdio for connections from Claude Code.

CLI Commands

docrag init

Initialize DocRAG configuration directory.

docrag add <name>

Add a new documentation collection.

Options:

• -s, --source PATH - Source directory containing documentation

• -d, --description TEXT - Description of the collection

Example:

docrag add qumu --source ~/docs/qumu --description "Qumu video platform docs"

docrag list

List all documentation collections with their status.

docrag update <name> <source>

Update an existing collection with new documents.

Example:

docrag update brightsign ~/docs/brightsign/updated

docrag remove <name>

Remove a documentation collection (with confirmation).

docrag search <query>

Search documentation from the CLI for testing.

Options:

• -c, --collection TEXT - Specific collection to search

• -l, --limit INTEGER - Number of results (default: 5)

Example:

docrag search "websocket connection" --collection brightsign

docrag serve

Start the MCP server for Claude Code integration.

docrag scrape <url>

Scrape documentation from websites.

Options:

• -o, --output PATH - Output directory (required)

• --smart, --use-crawl4ai - Use AI-powered Crawl4AI scraper (recommended)

• --no-llm - Disable LLM extraction (faster, still better than basic)

• --llm-provider TEXT - LLM provider (default: openai/gpt-4o-mini)

• --playwright - Use Playwright for dynamic content (basic scraper)

• --max-pages INTEGER - Maximum pages to scrape (default: 1000)

Examples:

# Basic scraping
docrag scrape https://docs.example.com --output ./docs

# Smart scraping with AI (recommended)
docrag scrape https://docs.example.com --output ./docs --smart

# Smart scraping without LLM (faster, no API key needed)
docrag scrape https://docs.example.com --output ./docs --smart --no-llm

# Limit pages
docrag scrape https://docs.example.com --output ./docs --max-pages 100

Smart Scraping Features:

• ✨ AI-powered content extraction

• 🎯 Automatically removes navigation and boilerplate

• 📊 Better handling of complex layouts

• 🧠 Semantic understanding of documentation structure

• ⚡ Faster and more accurate than basic scraping

To enable smart scraping:

# Install Crawl4AI
pipx inject docrag crawl4ai

# Optional: Set OpenAI API key for LLM-powered extraction
export OPENAI_API_KEY='your-key-here'

Using with Claude Code

1. Configure Claude Code MCP Settings

Add DocRAG to your Claude Code MCP configuration (~/.config/claude-code/mcp_settings.json or similar):

{
  "mcpServers": {
    "docrag": {
      "command": "docrag",
      "args": ["serve"],
      "env": {}
    }
  }
}

If using the full path:

{
  "mcpServers": {
    "docrag": {
      "command": "/home/claude-admin/.local/bin/docrag",
      "args": ["serve"],
      "env": {}
    }
  }
}

2. Restart Claude Code

After adding the configuration, restart Claude Code to load the MCP server.

3. Use in Claude Code

Once connected, Claude Code can use two tools:

search_docs: Search through indexed documentation collections

Query: "how to handle authentication in BrightSign"
Collection: (optional) "brightsign"
Limit: (optional) 5

list_collections: List all available documentation collections

Claude will automatically use these tools when working on projects that need documentation access.

Architecture

Core Components

1. ConfigManager (config.py) - Manages configuration and collection metadata

2. EmbeddingGenerator (embeddings.py) - Generates embeddings using sentence-transformers

3. VectorDB (vectordb.py) - LanceDB wrapper for vector storage and search

4. DocumentIndexer (indexer.py) - Intelligent document chunking and indexing

5. DocRAGServer (server.py) - MCP server implementation

6. CLI (cli.py) - Command-line interface

Technical Stack

• MCP Framework: Official Anthropic MCP package

• Vector Database: LanceDB (lightweight, file-based, performant)

• Embeddings: sentence-transformers with all-MiniLM-L6-v2 model (384 dims, fast, local)

• Text Processing: langchain-text-splitters for intelligent chunking

• CLI: Click for user-friendly commands

• Web Scraping: Playwright + BeautifulSoup4 for scraping

Data Structure

~/.docrag/
├── config.json                 # Global configuration
│   └── {
│         "active_collections": ["brightsign", "venafi"],
│         "embedding_model": "sentence-transformers/all-MiniLM-L6-v2",
│         "chunk_size": 512,
│         "chunk_overlap": 50
│       }
├── collections/
│   ├── brightsign/
│   │   ├── metadata.json       # Collection metadata
│   │   └── source_docs/        # Original documents
│   ├── venafi/
│   └── qumu/
└── vectordb/
    └── lancedb/                # Vector storage (one table per collection)

Configuration

Global configuration is stored in ~/.docrag/config.json:

{
  "active_collections": ["brightsign", "venafi"],
  "embedding_model": "sentence-transformers/all-MiniLM-L6-v2",
  "chunk_size": 512,
  "chunk_overlap": 50
}

Collection metadata is stored in ~/.docrag/collections/<name>/metadata.json:

{
  "name": "brightsign",
  "source_type": "local",
  "source_path": "/path/to/docs",
  "created_at": "2025-10-28T10:00:00",
  "updated_at": "2025-10-28T10:00:00",
  "doc_count": 150,
  "description": "BrightSign player documentation"
}

Development

Project Structure

docrag/
├── docrag/
│   ├── __init__.py
│   ├── cli.py              # CLI commands
│   ├── server.py           # MCP server
│   ├── indexer.py          # Document indexing
│   ├── vectordb.py         # Vector database
│   ├── embeddings.py       # Embeddings
│   ├── config.py           # Configuration
│   └── scrapers/           # Web scrapers
│       ├── __init__.py
│       ├── base.py
│       └── generic.py
├── tests/
├── pyproject.toml
├── README.md
└── DOCRAG_MVP_BUILD_GUIDE.md

Running Tests

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

Code Formatting

# Format with black
black docrag/

# Lint with ruff
ruff check docrag/

Troubleshooting

"DocRAG not initialized"

Run docrag init first to create the configuration directory.

"No collections found"

Add a collection with docrag add <name> --source <path>.

"Model download fails"

The first time you run DocRAG, it will download the sentence-transformers model (~100MB). Ensure you have internet connectivity.

"Playwright not installed"

If using scrapers, run playwright install chromium.

Future Enhancements

• Web scraper CLI commands

• Support for more file types (PDF, HTML, RST)

• Incremental indexing (only index changed files)

• Collection activation/deactivation

• Collection statistics and health checks

• Export/import collections

• Cloud sync for collections

• Advanced search filters

License

MIT

Author

Ryan - Built for homelab and Claude Code integration