A retrieval demo on a clean corpus looks easy. Production RAG breaks in quieter ways: wrong chunk boundaries, stale indexes, permission leaks across tenants, "confident" answers with weak grounding, and evaluations that never catch regressions.

This talk covers product and architecture lessons from shipping enterprise retrieval systems, including RAG, vector search, and LLM orchestration from pre-revenue through early enterprise customers, plus constraints learned working alongside high-stakes operational software. We focus on failure modes after the happy path: what to measure, what to gate before release, and how product managers should prioritize fixes when everything looks "fine" in a pilot.

Attendees get a production checklist: indexing and refresh, access control, evaluation harnesses, fallback behavior, and operator UX when the system should abstain. Examples stay vendor-agnostic. No FHIR. No tooling pitch.