Reliability
Request IDs
Correlate API failures without logging sensitive content.
Generated API responses normally include the same opaque server-generated identifier in:
X-Request-ID: …
Inference-Id: …
Record it alongside model ID, endpoint, status, timing, and token counts. This is enough for support correlation without retaining prompts or generated content.
Do not assume every host-level probe includes a request ID. GET /health and HEAD / are minimal public checks and can omit it. Client-provided request IDs are not adopted as the authoritative billing or inference identifier.
Never log Authorization, x-api-key, raw request bodies, raw response bodies, or customer media. If a support case needs content-level reproduction, create a new synthetic prompt that demonstrates the behavior.