Skip to content

Rate Limiting & Abuse Controls

  • Per-route rate limits are enforced server-side. The limit key is the authenticated user’s organization when a bearer token is present, and the IP address otherwise.
  • Rate-limited requests return HTTP 429 with a JSON body and a Retry-After header.
  • LLM-calling routes carry their own tighter per-organization limit so one tenant cannot drain the shared AI budget.
  • AI runs in mock mode by default (AI_LIVE_CALLS_ENABLED=False). Mock output is deterministic, labeled, and costs nothing.
  • Live calls require a human gate (PAID_EXTERNAL_CALLS_ENABLED) plus a kill switch (AI_ENABLED).
  • A hard monthly spend limit (default $60) is enforced in a shared ledger (Redis with a database fallback). Reaching it stops live calls until the window rolls over, and a web-process restart does not re-arm the cap.
  • Per-user token and prompt quotas cap individual usage inside the monthly limit.
  • Sync jobs are capped per source and per parse call, so an ingestion run cannot trigger unbounded model calls.