Skip to main content
Usage records always contain request metadata, key/application attribution, requested and resolved models, provider, status, token counts, estimated cost, latency, and streaming state. Prompt and response bodies are stored only when deployment-owned data_governance.payload_storage is explicitly enabled.

​
Summary

Summary fields are requests, error_requests, input/output/total tokens, estimated_cost, average_latency_ms, and costed_requests.

​
Filters

Usage, breakdown, time-series, and recent-request endpoints support relevant combinations of:
  • start_at and end_at ISO-8601 timestamps;
  • api_key_id;
  • org_id;
  • app_id;
  • environment;
  • model;
  • provider;
  • status_code.

​
Breakdown

Dimensions are model, provider, api_key, org, app, environment, and status.

​
Time series

Intervals are hour and day.

​
Recent requests

The response includes total, limit, offset, and has_more. Advance offset by limit to page through the filtered result. Request logs additionally accept outcome=success|denied|error and a literal substring search over request/session IDs through q (maximum 128 characters). Request logs are buffered and retried outside the inference response path. A prolonged database outage can drop metadata records; monitor gateway_request_logs_dropped_total. Budget enforcement uses a separate transactional ledger and remains authoritative for admission. List rows include payload_available and payload_truncated but never inline the bodies. When available, fetch one exchange through the tenant-scoped detail endpoint:
The detail includes request_payload, response_payload, and payload_captured_at. Credential-bearing keys and gateway token patterns are redacted, but ordinary PII and customer content remain visible. See Data retention and privacy before enabling capture.