RAG Documentation System

A Retrieval-Augmented Generation (RAG) system that enhances LLM responses with relevant document context.

Features

• Efficient document chunking with context preservation
• Markdown-aware processing that maintains document structure
• Semantic search using multilingual embeddings
• Incremental document updates using content hashing
• Integration with Google's Gemini LLM

Setup

1. Install dependencies:

uv sync

2. Set up environment variables: You can use any model, along with the correct key. For a list of all models and providers, refer to litellm docs.

GEMINI_API_KEY=your_api_key_here

3. Place your documentation in the docs/ directory as markdown files.

Usage

Basic Usage

uv run agent.py "Describe the caves of Xylos." # optional -web to add websearch tool to agent.

Command Line Interface

The embedding creation and database management is done through the CLI:

# Add/update documents and run the test query
uv run src/embed.py

# List all stored passages
uv run src/embed.py list

# Clear the database
uv run src/embed.py clear

# Perform a custom search query for testing
uv run src/embed.py "your search query"

Architecture

• text_chunker.py: Handles intelligent document splitting
• embed.py: Manages document processing and vector database
• retriever.py: Implements semantic search functionality
• agent.py: Integrates components with LLM using smolagents

Technical Details

• Uses Alibaba's multilingual embedding model for semantic search
• ChromaDB for vector storage
• Google Gemini for LLM responses
• Smart chunking preserves document context and header hierarchy
• Incremental updates avoid re-embedding unchanged content

Contributing

1. Fork the repository
2. Create your feature branch
3. Commit your changes
4. Push to the branch
5. Create a new Pull Request