RAG Codebase Assistant & AI Git Bot

A dual-purpose AI developer tool: a RAG-powered chat assistant that answers architectural and code queries using citations, combined with a GitHub App integration that indexes pushed files incrementally and performs automated, inline pull request reviews.

---

📸 Demo & Screenshots

Video Walkthrough

ragmp4.mp4

Screenshots

---

🏗️ Project Architecture Layout

The repository is divided into a self-contained Node/TypeScript backend/ and React frontend/ directories:

1. backend/
   • src/services/chunker.ts: A syntax-aware parser tailored to JS/TS (class, function, arrow declarations) and Python (def, class) that chunks files into semantic blocks (3-60 lines) carrying context headers.
   • src/services/queue.ts: BullMQ and Redis-backed background worker queues that handle async repository cloning, file traversal, and embedding generation.
   • src/services/embeddings.ts: Vector generation (768 dimensions) using Gemini's text-embedding-004 model.
   • src/services/qdrant.ts: Collection creator, semantic search coordinator, and cosine similarity distance engine.
   • src/routes/ask.ts: The query stream controller that pulls top-5 code matches, formats context prompts, and streams tokens using Groq's Llama models.
   • src/routes/webhook.ts: GitHub App callback handlers processing push events (incremental index updates) and pull requests (sending modified files to LLM and posting inline reviews).
2. frontend/
   • src/App.tsx: A premium React dashboard interface featuring code snippet drawer viewers, chat streams, indexing logs tables, and GitHub App installation guides.

---

🚀 Getting Started (Quick Start)

1. Launch Database & Infrastructure

Spin up PostgreSQL 15 (relational logs), Redis 7 (BullMQ queues), and Qdrant (vector store) via Docker:

docker-compose up -d

2. Configure Backend Server

Create a .env file in the backend/ folder:

PORT=3001
DATABASE_URL=postgresql://postgres:postgres@localhost:5432/rag_git_bot
REDIS_URL=redis://localhost:6379
QDRANT_URL=http://localhost:6333

# AI Credentials
GEMINI_API_KEY=your_gemini_api_key       # For text-embeddings-004
GROQ_API_KEY=your_groq_api_key           # For chat LLM streams
GROQ_MODEL=llama-3.1-70b-versatile

# GitHub App Secrets (Only needed for webhooks & reviews)
GITHUB_APP_ID=your_github_app_id
GITHUB_PRIVATE_KEY="-----BEGIN RSA PRIVATE KEY-----\n...\n-----END RSA PRIVATE KEY-----"
GITHUB_WEBHOOK_SECRET=your_webhook_hmac_secret

3. Launch Backend & Frontend Services

Start the Express backend API (automatically sets up Postgres tables on start):

cd backend
npm install
npm run dev

Start the Vite React web dashboard (runs on port 5173 or similar):

cd ../frontend
npm install
npm run dev

Note: By default, the frontend connects to the backend at http://localhost:3001/api. If you customize the backend port, you can configure it for the frontend by creating a frontend/.env file with VITE_API_BASE=http://localhost:your_port/api.

---

🛡️ Setting Up the GitHub App Integrations

To enable automated pull request reviews and push-to-index webhooks:

1. Register a GitHub App in your GitHub developer settings.
2. Set permissions:
   • Contents (Read): To read files on pushes and reviews.
   • Pull Requests (Read & Write): To post inline PR review reviews.
   • Metadata (Read).
3. Select Events: Subscribe to Push and Pull request callbacks.
4. Set Webhook URL: Point the webhook URL to your backend endpoint (use ngrok for local development: https://<ngrok-subdomain>.ngrok-free.app/api/webhook).
5. Private Key: Generate a private key .pem file. Convert the contents of the key into a single-line string replacing newlines with \n in your .env file.
6. Install the App on your desired target repositories.

---

🌟 Core Technical Highlights

• Syntax-Aware Chunking: Code chunks are split cleanly at function/class structures instead of arbitrary text windows. If a function is longer than 60 lines, the signature header is carried over to the next chunk as metadata context.
• Inline Pull Request Reviews: When a PR is opened/synced, the bot requests a structured JSON review from the LLM, filters out low-severity comments, and issues a combined review with inline code annotations using the GitHub pulls review API.
• Incremental Synchronization: The push webhook calculates changed files using commit SHA diffs, deletes outdated chunks from Qdrant, and indexes only the modified/added files without performing full repository clones.