Skip to main content
Omnia accepts OpenTelemetry traces over OTLP/HTTP at:
If your application already emits OTel (Vercel AI SDK telemetry, Pydantic AI, an existing OTel SDK), integration is configuration only:
No OTel yet? Use the tracing SDK instead.

​
Accepted formats and dialects

  • Protocols: OTLP/HTTP in protobuf (application/x-protobuf) or JSON. Batches are capped at 4 MB; keep your exporter’s default batch size.
  • Attribute dialects: model-call spans are recognized and normalized from the official gen_ai.* semantic conventions (including the structured message form and the older event form), the Traceloop/OpenLLMetry indexed form (gen_ai.prompt.N.*), the Vercel AI SDK’s ai.* attributes, and OpenInference (llm.*). Whatever arrives, one canonical shape is stored.
  • Spans in a grammar Omnia doesn’t recognize are kept as trace structure, and the response’s unrecognized_attr_namespaces names the attribute-key namespaces (never values) so you can tell us what to map next.

​
What lands where

  • Model-call spans become log entries (gradeable, judgeable, comparable) when your workspace has request logging enabled. Content is stored under your retention window with the same scrubbing and size caps as gateway traffic.
  • Everything else (tool calls, retrieval, your application logic) is kept as span structure: name, kind, status, timing, including failed model calls, which are stored with their ERROR status rather than dropped.
  • Population and grouping: the omnia.tag resource attribute names the traffic population (falls back to service.name); trace ids group multi-step agent runs.

​
Delivery semantics

Exporters retry, so ingest is idempotent: each span’s log id is derived deterministically from its trace and span ids, and retried batches deduplicate instead of double-counting. Ingest has its own rate-limit bucket; telemetry can never consume your inference budget.

​
The response tells you what happened

accepted counts gradeable log rows; spans_stored counts structure rows; stored_as_structure_only counts model calls kept as structure (logging off, redacting exporter, or a failed call with nothing to store); content_logging says whether your workspace stores content at all; the difference between “your instrumentation redacts” and “logging is off” is visible here instead of producing an identical-looking empty timeline.