Reference · Cheatsheet

Production ops signals

What to watch after shipping RAG or agents.

Dashboard minimum

| Signal | Typical action |

|--------|----------------|

| p95 latency | Page if 2× baseline |

| 5xx error rate | Page on spike |

| Citation / groundedness | Rollback if −30% DoD |

| Token spend | Budget alert |

1. **Alert** — investigate with trace ID

2. **Watch** — minor drift, no user impact yet

3. **Rollback** — quality or errors breach SLO

Root-cause: model, prompt, index, or traffic?

Add regression test or eval gate before re-ship

**Related:** glossary `slo`, `baseline-metric`, `rollback`, `observability`