Observability
The ability to see inside your AI system as it runs — traces, metrics, logs.
Simple explanation
Observability means you can answer 'what happened?' for any request. It combines traces (step-by-step execution), metrics (latency, cost, error rate) and logs (structured events). Without it, you're flying blind.
Engineering explanation
Observability = tracing + metrics + logs. Tracing: OpenTelemetry spans per stage. Metrics: P50/P95/P99 latency, cost per request, error rate, eval scores. Logs: structured JSON with query, response, model, cost. Dashboards: latency, cost, quality, errors. Alerts: anomaly detection on all three.
Example
User reports 'slow response' → trace shows generate=8s (model issue) → metrics show P95 jumped → alert fires → rollback.
When to use it
Every production AI system. The foundation of all debugging and monitoring.
Common misconception
Observability is NOT just logging. Logging without traces and metrics can't attribute latency or cost to specific stages. You need all three.