Availability
Know what is reachable.
orag_up, HTTP request/error rate, and HTTP p95 show whether the API can serve traffic. Use /readyz to investigate dependencies.
Observability · Prometheus · Grafana
Operate the RAG path
with evidence.
Import the versioned ORAG overview dashboard to see API availability, request errors and latency, RAG outcomes, dependency readiness, and trace-store health.
grafana — import
1. Configure Prometheus to scrape /metrics 2. Dashboards → New → Import 3. Upload orag-overview.json 4. Select the Prometheus datasource ✓ Observe before changing anything
RAG quality path
Track outcomes and latency.
RAG outcomes/cache status and RAG p95 are intentionally aggregate views. A single request remains diagnosable through its trace ID, not a metric label.
Safe response
Observe, diagnose, then approve.
Dashboard and alerts are read-only. Follow the runbook, generate a dry-run plan, and require explicit human approval before any write action.
Data boundary
Low-cardinality by design.
The dashboard uses only documented Prometheus series. It never includes trace IDs, tenants, user input, prompts, documents, model output, or raw errors. Metrics are process-local and reset on restart; Prometheus retention is configured by the operator. OTLP trace and core metrics export are optional; sampling, persisted metrics, and cross-service topology remain future work.