A full-stack AI application that ingests GitHub repositories and provides a grounded onboarding chatbot for new contributors.
The system clones a repository, parses and chunks source files, embeds them into a vector database, and enables semantic search and question answering over the codebase. It is designed to work even when a repository has little or no documentation by inferring purpose and structure directly from source code.
- Repository ingestion via GitHub URL
- Automatic cloning and deterministic updates
- Language-agnostic source parsing and chunking
- Dense vector search using Pinecone
- Code-aware embeddings via Hugging Face
- Retrieval-augmented responses with citations
- Multi-repo support using per-commit namespaces
- Interactive Streamlit frontend
- FastAPI backend with LangGraph orchestration
- Designed to function without README files
- Streamlit
- Handles repository ingestion and interactive chat
- Maintains per-repository chat sessions
- FastAPI REST API
- LangGraph for retrieval → answer orchestration
- GitPython for cloning and updates
- Hugging Face for embeddings
- Pinecone for vector storage
- User submits a GitHub repository URL
- Backend clones or updates the repository
- Files are parsed, chunked, and embedded
- Vectors are stored in Pinecone under a commit-scoped namespace
- User questions are embedded and matched against stored vectors
- Responses are generated strictly from retrieved code context
- Python 3.11+
- FastAPI
- Streamlit
- LangChain
- LangGraph
- Pinecone
- Hugging Face (Embeddings)
- GitPython
- Gunicorn + Uvicorn
github-onboarding-agent/
├── backend/
│ └── app/
│ ├── main.py
│ ├── config.py
│ ├── routers/
│ │ ├── ingest.py
│ │ └── chat.py
│ ├── services/
│ │ ├── github_loader.py
│ │ ├── chunker.py
│ │ ├── embeddings.py
│ │ ├── pinecone_store.py
│ │ ├── graph.py
│ │ └── openai_llm.py
│ └── models/
│ └── schemas.py
├── streamlit_app.py
├── requirements.txt
└── README.md
- Clone the Repository
git clone https://github.com/your-username/github-onboarding-agent.git
cd github-onboarding-agent
- Create a Virtual Environment
python -m venv venv
source venv/bin/activate
# Windows: venv\Scripts\activate
- Install Dependencies
pip install -r requirements.txt
Create a .env file in the project root:
PINECONE_API_KEY=your_pinecone_key
PINECONE_INDEX=github-onboarding
HF_EMBED_MODEL=sentence-transformers/all-MiniLM-L6-v2
REPOS_DIR=.repos
OPENAI_API_KEY=optional_for_llm_answers
OPENAI_MODEL=gpt-4o-mini
- Hugging Face embeddings do not require a token for public models
- OpenAI is optional; the app works in retrieval-only mode without it
Start Backend (FastAPI)
uvicorn backend.app.main:app --reload --port 8000
Backend will be available at:
http://127.0.0.1:8000
Start Frontend (Streamlit)
streamlit run streamlit_app.py
Frontend will be available at:
http://localhost:8501
- Paste a GitHub repository URL into the sidebar
- Click Ingest
- Wait for cloning, embedding, and indexing to complete
- Ask questions about the repository
- What does this repo do?
- What are the main entry points?
- How is the project structured?
- What are good first tasks for a new contributor?
- Each ingestion creates a new namespace tied to the commit hash, ensuring isolation between repositories.
- Namespaces are commit-scoped to prevent cross-repository contamination
- README files are optional; source code is the primary ground truth
- Chunk metadata stores snippet text for grounded answers
- Chat state resets automatically when switching repositories
- Incremental re-ingestion instead of re-embedding everything
- Language-specific parsing heuristics
- Repository summaries cached per namespace
- Graph-based dependency reasoning
- Authentication and multi-user support
MIT License