Codebase RAG

AST-based code chunking, hybrid vector search (dense + sparse), and LLM-powered Q&A over your codebase.

Features

• AST chunking — parses code by functions, classes, and symbols (not line windows)
• Hybrid search — combines dense embeddings (BAAI/bge-small-en-v1.5) with sparse vectors (Splade_PP_en_v1) via Qdrant RRF fusion
• LLM agnostic — swap between Groq, OpenAI, Gemini, or Anthropic via LLMFactory
• Incremental sync — GitHub webhook receiver for real-time re-indexing

Quick Start

1. Setup

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Create .env:

QDRANT_HOST=http://localhost:6333
QDRANT_API_KEY=
COLLECTION_NAME=codebase_chunks

2. Ingest a repository

python cli.py index -r my-repo -p /path/to/my-repo

3. Ask questions

export GROQ_API_KEY="your-key"
python cli.py query "How does authentication work?"

Commands

Command	Description
index -r <name> -p <path>	Scan a local repo, AST chunk, and seed Qdrant
query <question>	Retrieve relevant chunks and answer via LLM

Optional flags for query: --top-k, --provider, --model, --api-key.

Architecture

Local repo → ASTChunker → CodeChunks → Embed (dense + sparse) → Qdrant Cloud
User query → HybridRetriever (RRF fusion) → context → LLM → answer

LLM Providers

Provider	Default model	Env var
groq	llama-3.3-70b-versatile	GROQ_API_KEY
openai	gpt-4o-mini	OPENAI_API_KEY
gemini	gemini-1.5-flash	GEMINI_API_KEY
anthropic	claude-3-5-sonnet-latest	ANTHROPIC_API_KEY

API

Run uvicorn api.main:app for the webhook receiver (GitHub push events → auto re-index).