Answer:

Correct: B. This is the same pattern as Q22's customer-isolation fix: a hard metadata pre-filter applied before ANN search runs, so superseded chunks are structurally excluded from candidates rather than merely discouraged. It satisfies both constraints at once — audit retention (nothing is deleted) and citation safety (nothing non-current is ever a retrieval candidate). Option A breaks the audit requirement outright by removing the very records compliance needs to keep. Option C is a similarity-ranking nudge, not a guarantee — a highly relevant superseded filing can still outscore a less relevant current one, especially if wording barely changed between versions. Option D relies on the LLM reliably following an instruction buried in a long prompt, which is exactly the kind of soft constraint that fails silently under load — the question explicitly asks for a guarantee, and only a structural retrieval-time filter provides one.

RAG Interview Self-Assessment