Rate limit /chat/message: per-user + per-IP defence-in-depth

[vahid-ahmadi]

Summary

Adds two-layer rate limiting on `POST /chat/message` — the only expensive endpoint. Other routes are cheap reads and stay unlimited.

• Per user (or IP if anonymous) — `5/min` and `60/hour`. Authenticated clients send `X-User-Id`; the limiter keys by `user:` so signed-in users aren't blocked by anon flooders on shared NAT/proxy IPs.
• Per IP defence-in-depth — `30/min`, regardless of who is sending. Catches anon abuse and runaway scripts.
• Limits configurable via `RATE_LIMIT_CHAT_PER_MIN`, `RATE_LIMIT_CHAT_PER_HOUR`, `RATE_LIMIT_CHAT_IP_PER_MIN`.
• 429 handler returns JSON with `retry_after_seconds` and a matching `Retry-After` header.
• Frontend handles 429 with an inline assistant message ("you're sending messages a bit fast — please wait ~Ns") rather than the 402 paywall flow.

Closes #46.

Implementation notes

• New module `backend/rate_limit.py` owns the `Limiter`, the `chat_key_func`, and the env-tunable limit strings. Imported by both `main.py` (registration + handler) and `chatbot.py` (decorator).
• `backend/tests/conftest.py` bumps test limits to 10000/min — without it the existing `TestChatMessage` tests would trip the production cap on the 6th call.
• Storage is slowapi's default in-memory backend. Modal runs up to 10 containers with concurrency 100, so a determined attacker can spread requests across containers and bypass any single counter. The IP layer is approximate but adequate for the threat model — swap to Redis via `storage_uri` when we need cross-container accuracy.

Test plan

• `docker-compose up`, then hammer `/chat/message` 6 times in 30s with the same logged-in user → 6th request returns 429 with `Retry-After: 60`.
• Hammer 6 anonymous requests from one IP → same behaviour.
• In the UI, hit the limit → an inline assistant bubble appears with "you're sending messages a bit fast — please wait ~60s". Streaming state clears, paywall is not shown.
• `GET /health`, `GET /version`, `GET /conversations`, and the title generator continue to work under load.
• `pytest backend/tests/test_api.py -q` passes — the new `TestRateLimitConfig` is fast and offline.

Out of scope

• Cross-container accurate rate limiting (would need Redis).
• Per-org/team limits (no orgs yet).
• Rate-limiting cheap GET endpoints (deemed unnecessary for current threat model).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
policyengine-uk-chat	Ready	Preview, Comment	May 6, 2026 11:37am

[github-actions]

Beta preview is ready.

• Frontend: open preview
• Backend: open backend