LLM observability: connect logs, traces, and outcomes
Trace retrieval, generation, validation, and retries without confusing model completion with application success.
CONNECTED OBSERVABILITY
LLM log APIFollow the entire model workflow.
Connect LLM events to retrieval, retries, validation, and delivery so a single request tells a coherent story.
{
"schema_version": "1.0",
"timestamp": "2026-09-21T09:41:02.184Z",
"severity": "INFO",
"event_id": "evt_demo_002",
"event_name": "generation.completed",
"service": "assistant-worker",
"attributes": {
"workflow_id": "wf_demo_042",
"model": "example-model",
"config_version": "prompt-v3",
"outcome": "completed",
"duration_ms": 842
}
}Synthetic fixture · nothing is sentDownload JSON ↗
DESIGN WITH INTENT
Keep the important distinctions.
Find the slow stage
Distinguish model time from retrieval, queueing, and application work.
Preserve retries
Keep failed attempts visible even when a later attempt succeeds.
Connect outcomes
Attach validation results without calling them guarantees of truth.
THE EVENT CONTRACT
Fields with a reason to exist.
Use supported context propagation across participating services. Where a dependency does not expose tracing context, preserve an approved application reference and the observed provider request reference. Missing instrumentation should appear as a gap, not as zero work.
These fields illustrate a design pattern. Map them deliberately to your own schema, source, and destination.
| Field | What it means |
|---|---|
trace_id | An approved correlation reference from your tracing system. |
span_id | The operation within a trace, when instrumentation provides it. |
duration_ms | A duration with an explicit measurement boundary. |
attempt | The individual execution attempt within the workflow. |
A trace explains observed operations. It does not establish that a generated answer is correct or that every dependency participated in collection.
Technical reference: W3C Trace Context . Verify the version and settings used by your own implementation.
FIELD NOTES
Keep following the signal.
AI logging in production: context without oversharing
Instrument AI workflows with useful metadata, configuration versions, and clear outcomes—not entire conversations.
Token usage logging: measurements, estimates, and cost
Keep reported usage separate from estimates, handle retries, and make cost calculations explainable.
ONE EVENT IS A GOOD START
Make your next log more useful.
Start with a clear schema. Keep the context that matters.