Semantic Code Search Engine

A state-of-the-art, AI-powered semantic code search engine. It allows you to ingest entire GitHub repositories, chunk their source code using AST parsing, generate high-dimensional embeddings, and perform incredibly accurate natural-language search and Retrieval-Augmented Generation (RAG) over your codebase.

🌟 Features

• Semantic & Hybrid Search: Combines Dense Vector Search (cosine similarity) with BM25 Keyword Search using Reciprocal Rank Fusion (RRF).
• Retrieval-Augmented Generation (RAG): Ask conversational questions about your codebase. The AI synthesizes answers using retrieved chunks and provides exact file/line citations.
• AST-Aware Chunking: Intelligently parses TypeScript, JavaScript, and Python into semantically meaningful chunks (functions, classes) using tree-sitter, ensuring embeddings capture full context.
• Ultra-Premium UI: A stunning, "Linear-style" frontend with frosted glassmorphism, dot-grid depth, and streaming markdown formatting.

🛠️ Tech Stack

• Runtime & Package Manager: Bun
• Monorepo: Turborepo / Bun Workspaces
• Frontend: Next.js (React), Tailwind CSS, Framer Motion, Lucide Icons, React Markdown
• API: Hono (Edge-ready)
• Vector Database: Qdrant
• Cache & Rate Limiting: Redis
• AI Models: Gemini (via OpenAI SDK compatibility layer) for embeddings and LLM generation.

📂 Project Structure

This is a monorepo structured as follows:

• packages/api: The Hono API server handling search, RAG, and health checks.
• packages/frontend: The Next.js web application.
• packages/ingestion: The worker library for cloning, AST chunking, and embedding GitHub repositories into Qdrant.
• packages/shared: Shared TypeScript types and core utilities (Observability, Caching).

🚀 Getting Started

Prerequisites

• Bun installed locally
• Docker & Docker Compose (for running Qdrant and Redis locally)

1. Install Dependencies

bun install

2. Start Infrastructure

Start the local Qdrant vector database and Redis cache:

docker-compose up -d

3. Environment Variables

Copy the example environment file and fill in your Gemini API key (or OpenAI key):

cp .env.example .env

Ensure you set OPENAI_API_KEY with your key from Google AI Studio.

4. Run the Development Servers

Start the frontend and API servers in parallel:

# Start the Next.js Frontend (Runs on port 3000)
npm run dev

# In a separate terminal, start the Hono API (Runs on port 3001)
bun run dev:api

🧠 How It Works

1. Ingestion: You submit a GitHub URL on the frontend. The API triggers the ingestion pipeline.
2. Chunking: The repository is cloned. Supported languages are parsed into an Abstract Syntax Tree (AST) to extract functions and classes. Other files use a sliding-window chunker.
3. Embedding: Chunks are embedded using text-embedding-004 (or your chosen model) and stored in Qdrant.
4. Retrieval: When you search, the query is embedded and searched against Qdrant. A parallel BM25 search is run. The results are merged via RRF.
5. Generation: If you ask a question, the top chunks are injected into a token-budgeted context window, and the LLM streams back a synthesized answer with precise file line citations.

🔒 CI/CD & Deployment

• The repository uses GitHub Actions for continuous integration (Typechecking, Linting, Testing).
• API is configured for deployment on Railway (via Dockerfile).
• Frontend is configured for deployment on Vercel.

License

MIT