Request Tracing

Request tracing connects the operations involved in handling a request. A trace can contain spans, each representing an operation with timing and contextual information. Shared trace context allows work across services to be related rather than appearing as unrelated log messages.

Follow one assistant request

A trace might show policy retrieval, a model request, a tool call, and response construction. If retrieval returns no passages but the model still produces confident advice, the trace helps locate the missing evidence. Record useful version references and outcomes rather than assuming the final message describes what actually happened.

A span can contain child operations or overlap other spans. Summing every span duration can therefore double-count time; inspect the request timeline and concurrency when explaining latency.

A trace is an observation with limits

A technically successful span does not establish that the answer was correct or that a business action completed. Check the relevant service outcome and task criteria. Sampling and incomplete instrumentation can also leave gaps.

Use access controls, selective capture, redaction, and retention limits for sensitive inputs. Trace identifiers help connect observations; they are not authentication credentials or permission to inspect another user's data.

Reference: OpenTelemetry: Traces.