RepoMind — AI Codebase Assistant

Ask natural language questions about any Python codebase. Powered by function-level AST chunking, ChromaDB vector search, and a LangGraph agent loop with Claude.

Quick Start

1. Set up environment variables

Copy .env.example to .env and fill in your keys:

cp .env.example .env

Required keys:

• ANTHROPIC_API_KEY — from console.anthropic.com
• OPENAI_API_KEY — for text-embedding-3-small embeddings
• MONGODB_URI — MongoDB Atlas connection string
• DJANGO_SECRET_KEY — any random string for dev
• JWT_SECRET — random string signing login tokens; generate with python -c "import secrets; print(secrets.token_hex(32))". Required in production — the server refuses to start without it.

2. Start the backend

cd backend
pip install -r requirements.txt
python manage.py runserver

Backend runs at http://localhost:8000

3. Start the frontend

cd frontend
npm install
npm run dev

Frontend runs at http://localhost:5173

4. Use it

1. Drag-and-drop a .zip of your Python repo into the upload area
2. Wait for indexing to complete (you'll see chunk count)
3. Ask questions in the chat — answers include filename:start-end citations

API Reference

Endpoint	Method	Description
/api/upload/	POST	Upload ZIP or raw source (multipart or JSON)
/api/query/	POST	Ask a question, get answer + citations
/api/sessions/	GET/POST	List or create chat sessions
/api/sessions/<id>/	GET/DELETE	Get or delete a session

Architecture

User question
     │
     ▼
[Retrieve] ── top-5 chunks from ChromaDB
     │
     ▼
[Grade] ── relevant or irrelevant?
     │
     ├─ irrelevant ──► [Rewrite] ──► back to Retrieve (max 2 retries)
     │
     └─ relevant ───► [Generate] ── Claude answers with citations

Chunking is by function/class scope (Python ast module), not word count — so the LLM always receives complete, syntactically valid code.