AI logging in production: context without oversharing
Instrument AI workflows with useful metadata, configuration versions, and clear outcomes—not entire conversations.
AI WORKFLOWS
AI log APISee the context behind the answer.
Design AI logs for workflows, retrieval, tool actions, and application outcomes—with metadata at the center.
{
"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.
Follow the workflow
Separate retrieval, generation, validation, and persistence.
Explain configuration
Reference versioned settings without copying private documents.
Control content
Treat prompt and output capture as separate, restricted decisions.
THE EVENT CONTRACT
Fields with a reason to exist.
A successful model response is not necessarily a successful application workflow. Record stage-specific outcomes and distinguish a proposed tool action from an executed one. Preserve the configuration references needed to investigate behavior after a release.
These fields illustrate a design pattern. Map them deliberately to your own schema, source, and destination.
| Field | What it means |
|---|---|
workflow_id | The logical application operation across attempts. |
model | A configured model or deployment reference. |
config_version | The prompt, retrieval, or application configuration revision. |
outcome | What this stage actually completed or failed to do. |
Do not place private conversations, tool credentials, or unrestricted document content in routine telemetry.
Technical reference: OpenTelemetry GenAI semantic conventions . Verify the version and settings used by your own implementation.
FIELD NOTES
Keep following the signal.
LLM observability: connect logs, traces, and outcomes
Trace retrieval, generation, validation, and retries without confusing model completion with application success.
Log redaction: keep sensitive data out of your pipeline
Minimize fields, apply deliberate transformations, and test redaction through storage, viewers, and exports.
ONE EVENT IS A GOOD START
Make your next log more useful.
Start with a clear schema. Keep the context that matters.