RepoChat AI

AI that reads any GitHub repo so you don't have to.

Paste a public GitHub repository, ask anything, get an answer with citations to the actual source files.

▶ Watch the 2-minute demo

---

What it does

• Paste any public GitHub repository URL
• The backend fetches relevant files, chunks them, and stores embeddings in a local vector DB
• Ask natural-language questions about the repo
• Get AI-generated answers with clickable source citations to specific files on GitHub

How it works

                   ┌────────────────────┐
                   │  GitHub repo URL   │
                   └─────────┬──────────┘
                             │
                  ┌──────────▼───────────┐
                  │  GitHub REST API     │  ← fetch tree + raw files
                  └──────────┬───────────┘
                             │
                  ┌──────────▼───────────┐
                  │  Language-aware      │  ← RecursiveCharacterTextSplitter
                  │  chunker (LangChain) │
                  └──────────┬───────────┘
                             │
                  ┌──────────▼───────────┐
                  │  OpenAI embeddings   │  ← text-embedding-3-small
                  └──────────┬───────────┘
                             │
                  ┌──────────▼───────────┐
                  │  ChromaDB (on-disk)  │  ← stores vector + path + repo_id
                  └──────────┬───────────┘
                             │
                  ─────────── ask question ───────────
                             │
                  ┌──────────▼───────────┐
                  │  Retriever - top-K   │  ← nearest-neighbor on the question
                  └──────────┬───────────┘
                             │
                  ┌──────────▼───────────┐
                  │  Claude Sonnet 4.6   │  ← grounded synthesis from chunks
                  └──────────┬───────────┘
                             │
                  ┌──────────▼───────────┐
                  │  Answer + sources    │
                  └──────────────────────┘

Each chunk stores its file path and repo identifier as metadata, so the retriever can scope queries to a single repo and answers can cite specific files.

Tech stack

Frontend

• Next.js 16 (App Router)
• React 19
• Tailwind CSS v4 (+ typography plugin)
• react-markdown + rehype-highlight

Backend

• FastAPI (Python 3.11+)
• LangChain text splitters
• OpenAI Python SDK - embeddings
• ChromaDB - vector store
• Anthropic Python SDK - Claude Sonnet 4.6 for generation
• GitHub REST API

Running locally

Prerequisites

• Python 3.11+
• Node.js 20+
• An OpenAI API key (for embeddings)
• An Anthropic API key (for Q&A)
• (Optional) A GitHub personal access token - bumps API rate limit from 60/hr to 5000/hr

Backend

cd apps/api
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env
# Edit .env and add:
#   OPENAI_API_KEY=sk-...
#   ANTHROPIC_API_KEY=sk-ant-...
#   GITHUB_TOKEN=ghp_...   (optional)

uvicorn main:app --reload

The API serves at http://localhost:8000. Swagger UI at http://localhost:8000/docs.

Frontend

cd apps/web
npm install
npm run dev

Open http://localhost:3000.

API

POST /index-repo

{ "repo_url": "https://github.com/owner/repo" }

Fetches the repo, chunks supported files (.md, .py, .ts, .tsx, .js, .jsx, .json), and stores embeddings.

{ "repo_id": "owner/repo", "file_count": 47, "chunk_count": 148 }

POST /ask

{ "repo_id": "owner/repo", "question": "How does authentication work?" }

Returns an answer + the source chunks used to generate it:

{
  "answer": "Authentication uses JWT tokens issued by ... (src/auth.ts)",
  "sources": [
    { "path": "src/auth.ts", "chunk_index": 4, "distance": 0.187 },
    { "path": "src/middleware.ts", "chunk_index": 0, "distance": 0.213 }
  ]
}

Folder structure

repochat-ai/
├── apps/
│   ├── api/                  # FastAPI backend
│   │   ├── main.py           # routes + CORS + Pydantic models
│   │   ├── repo_loader.py    # GitHub URL → file list
│   │   ├── chunker.py        # files → chunks (language-aware)
│   │   ├── embeddings.py     # chunks → vectors → Chroma
│   │   └── rag.py            # retrieve + prompt + Claude
│   └── web/                  # Next.js frontend
│       ├── app/page.tsx      # single-page chat UI
│       └── lib/api.ts        # typed API client
└── data/
    └── chroma/               # local vector DB (gitignored)

The problem

Developers spend too much time understanding unfamiliar codebases.

RepoChat AI is an experiment in combining RAG, semantic search, code understanding, and AI-assisted developer workflows to make onboarding into repositories much faster.

Future improvements

• Stream the answer (Server-Sent Events) instead of waiting for the full response
• Persist conversation history server-side
• Line-range source links (#L42-L78) by carrying line numbers through chunking
• Support large repos that exceed GitHub's tree-truncation limit
• Distance-threshold gating so off-topic questions get a clear "I don't know" instead of low-confidence guesses
• Per-repo collections to support clean re-indexing

Built by

Mahima Thacker - @mahima_thacker on X