Rate limits
The ceilings applied to the public API and the widget.
Limits are fixed windows of one minute. Ceilings keyed on your workspace or chatbot are the ones that actually bound spend, because a caller cannot change which workspace it is in.
Public API
Website widget
Handling a 429
Exceeding a limit returns 429 with code rate_limited, and the message states how long until the window resets. Retry with exponential backoff and jitter. If you consistently need more headroom than the per-key limit, spread traffic across keys per application rather than retrying harder.
Deployment note
Limits are enforced per application instance. A deployment running several instances behind a load balancer multiplies the effective ceiling, so an edge rate limiter should front the public API in production.