Ask natural-language questions about any GitHub repository and get grounded, cited answers — not guesses. Connect a repo, CodeAtlas clones and chunks it respecting function/class boundaries, embeds the code, and answers your questions with real file/line citations pulled from an Atlas Vector Search index.
Most "chat with your codebase" tools either hallucinate confidently or bury you in raw search results. CodeAtlas is built around one constraint: every claim in an answer must trace back to a real, retrievable chunk of code — proven with an eval harness, not just eyeballed.
- GitHub OAuth — sign in with your GitHub account
- Multi-repo support — connect and switch between several of your own repos
- Structure-aware chunking — a regex-based parser that respects function/class/factory-call boundaries (e.g. Redux
createSlice), not fixed-size splits - Semantic search — Gemini embeddings + MongoDB Atlas Vector Search
- Grounded, cited chat — streaming answers with inline
[1]references that map to real file paths and line numbers, filtered to only what the model actually cited - Webhook-based re-indexing — a
git pushto the default branch automatically triggers a background re-index via a BullMQ/Redis job queue, verified with HMAC-signed GitHub webhooks - 3D dependency graph — visualizes real import relationships between files (React Three Fiber)
- Developer tools — dedicated Chunk Explorer and raw Vector Search pages for inspecting retrieval directly
Measured with a 10-query eval harness (pnpm eval) against a real indexed repo:
| Metric | Result |
|---|---|
| Hit rate @ top-8 | 100% (10/10) |
| Mean Reciprocal Rank | 0.792 |
graph LR
User -->|GitHub OAuth| Web[React + Vite]
Web -->|REST| API[Express API]
API -->|clone + chunk| GitHub[(GitHub API)]
API -->|embed| Gemini[Gemini API]
API -->|store chunks + vectors| Atlas[(MongoDB Atlas<br/>Vector Search)]
API -->|enqueue| Queue[(Redis / BullMQ)]
Worker[Background Worker] -->|consumes| Queue
Worker -->|re-index| Atlas
GitHub -->|webhook on push| API
| Layer | Choice |
|---|---|
| Monorepo | pnpm workspaces, TypeScript |
| Frontend | React, Vite, Tailwind CSS, shadcn/ui, React Three Fiber |
| Backend | Node.js, Express 5 |
| Database | MongoDB Atlas + Atlas Vector Search |
| Embeddings / LLM | Google Gemini (gemini-embedding-001, gemini-3.6-flash) |
| Job queue | BullMQ + Redis |
| Auth | GitHub OAuth |
pnpm install
pnpm build:sharedRequires .env files in apps/api for MongoDB, GitHub OAuth, Gemini, and Redis credentials — see apps/api/.env.example.
pnpm dev:api # backend
pnpm dev:web # frontend
pnpm dev:worker # background indexing workerpnpm test # unit tests (chunker, dependency graph resolution)
cd apps/api && pnpm eval <repoId> # retrieval quality eval- Chunking is regex-based, not full AST parsing — reliable for JS/TS, imprecise for other languages
- No conversation memory — each chat question is answered independently
- Dependency graph resolves relative imports only, not path aliases or external packages