179 questions
No questions match those filters.
← All questions
RAGHard· RAG & Retrieval
Your RAG system's answer-quality dashboards are green,...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansIf an AI evaluator scores the output of an AI generator with no external anchor, the two share failure modes and drift together — confidence rises while quality falls, because the judge is becoming more certain about the wrong things.
The prevention is an external anchor from day one:
- A fixed, human-labelled golden set that never changes.
- Retrieval quality measured independently of answer quality, so you know which layer broke.
- Periodic human review of a live sample, not just the automated score.
Never let a model be the sole judge of a model — a dashboard that’s all green while users complain is usually a sign the judge and the generator have started agreeing with each other, not that the system is healthy.