Search RAG Generation
Search RAG - Grounding and Uncertainty
A grounded answer is allowed to be incomplete.
Why this matters in production
Users do not need the model to sound certain. They need to know what the evidence supports. In high-risk workflows, an honest partial answer is better than a polished guess.
This works in demos but fails in production: forcing the model to always answer. That turns missing evidence into hallucination pressure.
What breaks
The model fills gaps, smooths contradictions, and turns weak evidence into strong claims. Citations then become decoration instead of proof.
The bad case is not 'I do not know.' The bad case is 'I know' from the wrong evidence.
What works
Require the model to classify evidence state: sufficient, partial, conflicting, missing, or unauthorized. The answer style should follow that state.
Do not hide conflicts. If two sources disagree, the answer should say so and cite both, or refuse to decide.
Practical guidance
Teach the system safe outputs: direct answer, answer with caveats, clarifying question, or refusal. Then evaluate whether it chose the right path.
What happened
A compliance assistant answered 'yes' to a retention question because one retrieved policy allowed it, while another newer policy restricted it.
Why retrieval failed
Retrieval returned both documents, but generation optimized for a clean answer instead of conflict handling.
Why it was hard to detect
The citations were valid individually. The failure was the missing conflict interpretation.
What fixed it
The fix was evidence-state detection, recency-aware conflict rules, and tests where refusing to choose counted as success.
Practical Guidance
Reward explicit uncertainty.
Treat conflict handling as grounding.
Evaluate refusal correctness.