An HR agent's retriever returns the top-3 most-relevant passages by similarity score. When two versions of a remote-work policy exist in the index — a 2023 doc saying "fully remote available" and a 2024 leadership memo saying "hybrid required" — the retriever usually returns just the higher-scoring one. The agent presents whichever version was retrieved as canonical. Multiple versions never both surface, so the agent never sees the conflict.
What's the architectural fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
If you want the agent to surface conflicts, the retriever has to surface them first. A standard top-K retriever ranking by similarity will return one version of a policy and silently drop the others. Topic-aware deduplication — "if multiple passages cover the same policy topic, return all distinct versions with metadata" — is the architectural fix. The agent can then say "these two policies appear to conflict; please clarify" instead of confidently picking one.
Returning 10 instead of 3 means more random extras. The 8th-most-similar passage might be on a different topic entirely. Conflict surfacing requires structural awareness of when two passages are about the same policy, not just more volume.
Re-querying with different phrasings is a luck-based workaround that may or may not surface the alternate version. It also doubles latency and cost. Topic-aware retrieval is the deterministic fix.
Similarity scores measure how close a passage is to the query in embedding space — not which version is most current, most authoritative, or most applicable. A higher score on an older policy doesn't make it correct.