Reference · How-to · ~12 min

How to deploy a RAG app (checklist)

Last updated

Pre-flight before customer traffic hits retrieval + generation. Treat index and model versions as a pair.

Pre-flight before customer traffic hits retrieval + generation. Treat index and model versions as a pair.

#Prerequisites

  • Staging environment with production-like docs sample
  • Golden questions with expected citations
  • Rollback plan for both index and prompt/model

#Checklist

### Security & config

  • API keys server-side only; no keys in the browser
  • Prompt injection posture documented for retrieved text
  • PII retention rules for logs and traces
  • ### Retrieval

  • Embedding model version pinned (same at index + query)
  • Vector index version tagged in deploy notes
  • Chunking settings recorded (size, overlap, metadata filters)
  • ### Serving

  • `/health` checks DB + model reachability
  • Timeouts and max tokens set
  • Rate limits and backpressure tested
  • ### Quality gate

  • Golden questions pass in staging (faithfulness + citations)
  • Refusal behavior when evidence is missing
  • Rollback: prior index snapshot or model ID documented
  • ### After cutover

  • Monitor groundedness / citation rate
  • Alert on retrieval empty-hit spikes
  • Owner on-call for quality regressions
  • #Failure notes

  • Index rebuilt with a different embedding model than queries
  • Prompt updated without re-running golden tasks
  • Stale docs still dominating top-k
  • **Try the lesson:** `deploy-rag-app` in Lane D · How-to: [Build a 5-doc RAG](/reference/how-to/build-5-doc-rag)