Observability

SLO / Error Budget

Availability SLI
target ≥ 99.5% (30d)
Latency SLI
target ≥ 90% under 10s
Eval Quality SLI
target ≥ 0.60
Error Budget Left
30-day window
SLO Description Current SLI Target Status
Loading…
Prometheus burn-rate alerts fire at 14.4× (fast, 1h+6h window) and 3× (slow, 6h+24h window). Rules live in monitoring/slo_rules.yml.

Drift Detection

Loading baseline…
Time Level Score Relevance Δ Quality Δ Centroid Dist Query Count Alert
Loading…

Distributed Trace Context (W3C)

Every request is assigned a W3C traceparent header. The trace ID propagates from FastAPI through Celery task kwargs into agent nodes, giving end-to-end correlation.

Format00-{trace_id_32hex}-{span_id_16hex}-01
Response headerX-Trace-Id, traceparent
Celery propagationtrace_id kwarg on process_ask
Agent propagationrestore_trace() at task entry
Langfusesession_id + user_id per span
Fetching…