By Vamsi Krishna
Engineering · · Discuss · updated
60 Days of Production AI Systems · chapter 5 of 10Production RAG Operations Day 29 · part 5 of 6

New here? Start this series at Day 25: Why production knowledge often lives in systems, not PDFs

  1. Day 25: Why production knowledge often lives in systems, not PDFs
  2. Day 26: Why evidence is becoming multimodal
  3. Day 27: Why freshness is part of correctness
  4. Day 28: Why weak evidence should trigger recovery, not confidence
  5. Day 29: Why reflection only matters when tied to evidence
  6. Day 30: Why RAG must be evaluated in parts

"Have the model check its own work" sounds circular, and sometimes it is. Whether it helps or just costs money comes down to one thing: what the check is anchored to.

Anchored versus unanchored

Ask a model "is this answer good?" and you get what you would expect: a thoughtful paragraph concluding that the answer is, on balance, quite good. Unanchored reflection is a model grading its own prose using the same intuitions that produced it.

Ask instead: does source 2 support claim 3?, and you have changed the operation entirely. That is verification against an external artifact rather than self-assessment, and it works because verification is a genuinely easier task than generation. The same model that confidently invented a detail can often catch that invention when forced to line the sentence up against the retrieved text.

One is a vibe check; the other is a comparison with a right answer available.

Make it produce a decision, not a paragraph

The most common way this pattern wastes money: the reflection step runs, produces a considered critique, and nothing downstream consumes it. The answer ships unchanged. You have added latency and a token bill for a second opinion nobody acts on.

Wire in consequences explicitly:

  • Unsupported claim → removed from the answer, or marked as unverified
  • Multiple unsupported claims → trigger another retrieval pass with a reformulated query
  • Core claim unsupported → refuse rather than ship

If you cannot name what changes as a result of the reflection verdict, do not add the step yet.

Budget it

Reflection loops have a natural tendency to sprawl. Generate, critique, revise, critique the revision, and so on, each iteration defensible, the total unbounded.

One pass, on answers that matter. Route by stakes: a customer-facing eligibility determination earns a verification pass; an internal search summary probably does not. A reflection loop with no exit condition is latency with good intentions.

Closing thought

Check the answer before sending it, against the evidence, with clear criteria, and with the authority to change what ships.

Reflection tied to sources is one of the highest-value additions to a RAG pipeline. Reflection tied to nothing is the model admiring its own work at your expense.

If your reflection step returned "unsupported" tomorrow, what would actually happen to the answer?

Checklist · · 12 checks

RAG production-readiness checklist

Twelve checks to pass before a retrieval system answers a real user. Tick them locally; progress stays in your browser.