Ungithub

🧠 About the Project

🚀 Inspiration

As developers, we've all faced the pain of onboarding into unfamiliar or legacy codebases - slow, confusing, and time-consuming. We asked ourselves: What if a GitHub repo could explain itself?

That’s how ungithub was born, an AI-powered codebase insight explorer that acts like a smart mentor embedded into any open-source project.

🏗️ How We Built It

We broke the project into 6 clear phases:

1. Planning & Setup
   I scaffolded the backend (FastAPI) and optionally the frontend (Next.js), and connected to MongoDB Atlas with vector search enabled.

2. Repo Cloning & Parsing
   Using git and GitHub API, I cloned public repositories and recursively parsed files — filtering out noise (node_modules, .git, etc.) and storing structured metadata.

3. Embeddings + Vector Search
   Code chunks were vectorized using OpenAI or SentenceTransformers, then stored in MongoDB Atlas with a vector index for similarity search.

4. Natural Language QA
   I built an endpoint that accepts user questions, performs a vector search, and feeds the most relevant chunks to an LLM to generate accurate answers — with file-level citations.

5. Frontend (optional)
   A clean Next.js UI lets users paste a GitHub link, wait for indexing, and ask questions about the repo. Results are displayed in real time.

6. Polish & Deploy
   In the end, deployed the backend to Render and frontend to Vercel, added error handling, and optimized for minimal latency and good UX.

🧠 What I Learned

• MongoDB Atlas Vector Search is powerful and beginner-friendly for semantic search.
• LangChain + OpenAI/Gemini is a potent combo for contextual code summarization and Q&A.
• Optimizing cold starts, managing large repositories, and chunking logic was trickier than expected.
• Good UX is critical, even for technical tools, so I kept interactions minimal: paste, wait, ask.

🧱 Tech Stack

🧗 Challenges I Faced

• Handling large monorepos efficiently (chunking, memory usage)
• Dealing with cold backend starts due to free-tier hosting
• Creating a generic chunking algorithm that works across multiple languages
• Ensuring relevant vector search — avoiding false positives
• Avoiding token overflows when passing large context to LLMs

💡 Try it Yourself

🔗 Paste any repo at ungithub.vercel.app
⏳ Small repos index in 3–4 minutes, large ones take 5–7.
🤖 Once indexed, ask your questions and get answers in seconds!

🙌 Built For

AI in Action Hackathon — hosted by Google x MongoDB x GitLab

Built With

• fastapi
• gemini
• mongodb
• python