AI Center reads ordinary traces. Any span that follows the
OpenTelemetry GenAI semantic conventions,
version 1.37 or later, shows up there once it’s stored: send traces as for any service (see
Traces), with an ingest key that has obs.traces.ingest.
There’s nothing to turn on in Observability.
Only the OpenTelemetry GenAI conventions are read. Other formats, such as OpenInference (llm.*)
or indexed prompt attributes (gen_ai.prompt.0.content), are stored as ordinary span attributes
and don’t appear in AI Center. Content in the layout of conventions before 1.37, one event per
message (such as gen_ai.user.message), isn’t read either.
Turn on content capture
Instrumentations leave prompts and completions out of spans unless you turn content capture on. Without content, a run still has its steps, timing, tokens and errors, and counts in the overview, but its transcript has no text and search doesn’t find it.
AI Center reads content from span attributes: gen_ai.input.messages,
gen_ai.output.messages, gen_ai.system_instructions and the others in the
data reference, on the span or on its
gen_ai.client.inference.operation.details span event. Where an instrumentation can put content
on spans or in log records, choose spans. Content sent as log records is stored as ordinary logs
and isn’t shown in AI Center.
For example, OpenTelemetry’s Python GenAI instrumentations take:
OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental
OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_ONLYFor other instrumentations, see their documentation for content capture and the conventions version they follow.
To keep some content out of storage, redact it at ingest with transform rules: they see GenAI spans with their content, before it’s stored.
Mastra
A Mastra app that sends its traces with @ezghcloud/mastra-integration
sends GenAI spans with their content, so its agents’ runs appear in AI Center with nothing else to
set up. See Deploy a Mastra app.
Check it
Open Observability, then AI Center, Runs, and choose the environment your agents send to. The page is empty until a GenAI span arrives in the time range, and runs are searchable within seconds of their spans arriving (in a dataset with trace sampling, after its decision wait). A run whose transcript has steps but no text was sent without content.