RepoContext

Ask plain-English questions about any GitHub repository and get cited answers that link back to the exact file and line.

"Where is authentication handled?"  →  src/auth/login.py#L42-L88
"What does the payment module depend on?"  →  src/payments/processor.py#L12-L67

---

How It Works

  GitHub repo
      │  git clone
      ▼
  ┌─────────────┐    ┌──────────────────┐    ┌──────────────────────┐
  │ File walker │───▶│  tree-sitter     │───▶│  Semantic chunks     │
  │ (skip junk) │    │  AST parser      │    │  + metadata          │
  └─────────────┘    └──────────────────┘    └──────────┬───────────┘
                                                          │  embed (Voyage AI)
                                                          ▼
                                               ┌──────────────────────┐
                                               │   Qdrant vector DB   │
                                               └──────────┬───────────┘
                                                          │
  User question ──▶ embed ──▶ retrieve top-k ──▶ LLM ──▶ cited answer

Ingestion (once per repo): clone → walk files → parse with tree-sitter → extract function/class/method chunks with metadata → embed → store in Qdrant.

Query (every question): embed question → vector search → retrieve top-k chunks → Llama 3.3 70B generates a cited answer.

The Key Design Decision

Naive RAG splits code every N characters — cutting functions in half and separating them from their context. RepoContext chunks by syntax tree (using tree-sitter), so every chunk is a complete function, class, or method with its metadata intact (file_path, start_line, end_line, language, parent_class). This makes retrieved chunks semantically meaningful.

---

Tech Stack

Layer	Choice
Backend	FastAPI (Python 3.12)
AST Parsing	tree-sitter (Python, JS, TS, Java, Go)
Embeddings	Voyage AI voyage-code-3
Vector DB	Qdrant Cloud
LLM	Groq — Llama 3.3 70B
Frontend	React + Tailwind CSS
Backend host	GCP Cloud Run
Frontend host	Vercel

---

Project Structure

ingestion/    # Repo cloning, file walking, AST chunking, embedding + Qdrant upsert
retrieval/    # Query embedding, Qdrant search, LLM answer generation
api/          # FastAPI app — /ingest, /query, /health
eval/         # Evaluation harness — recall@k, MRR, faithfulness scoring
frontend/     # React + Tailwind chat UI

---

Running Locally

1. Clone and install

git clone https://github.com/byreddy1303/RepoContext.git
cd RepoContext
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

2. Set up environment

cp .env.example .env

Fill in .env:

GROQ_API_KEY=...          # free at console.groq.com
VOYAGE_API_KEY=...        # free tier at dash.voyageai.com
QDRANT_URL=...            # free cluster at cloud.qdrant.io
QDRANT_API_KEY=...
GITHUB_TOKEN=...          # GitHub PAT (read-only)
QDRANT_COLLECTION=RepoContext

3. Start the backend

uvicorn api.main:app --reload --port 8000

4. Start the frontend

cd frontend
npm install
npm run dev

Open http://localhost:5173, paste a GitHub repo URL, and start asking questions.

CLI (no UI)

# Ingest a repo
python -c "from ingestion.pipeline import ingest_repo; ingest_repo('https://github.com/pallets/flask')"

# Ask a question
python cli_search.py "how does Flask handle request routing"

---

API

Method	Endpoint	Description
GET	/health	Health check
POST	/ingest	Ingest a repository {"repo_url": "..."}
POST	/query	Ask a question {"question": "...", "repo_url": "..."}

---

Running Tests

pytest ingestion/test_repo_loader.py ingestion/test_chunker.py -v

---

Roadmap

• Evaluation harness — recall@k, MRR, LLM-as-judge faithfulness scoring
• Cross-encoder re-ranking
• Incremental re-indexing (hash-based, only re-embed changed files)
• Deploy to GCP Cloud Run + Vercel