# Public mode: the AI service is a paid upstream (a model call), so a write is limited to
# 2 r/s per client, burst 10; reads and preflights get their own 20 r/s, burst 80
# (http-limits.conf says why). A refused request is a 429 (not nginx's 503) so the
# client can tell "slow down" from "broken".
limit_req zone=public_ai burst=10 nodelay;
limit_req zone=public_ai_read burst=80 nodelay;
limit_req_status 429;