A tax FAQ wrapper must not show ordinance answers that lack retrieved citation spans. Which behavior should the module implement?
Select an answer to reveal the explanation.
Short Explanation
No citation, no auto-answer—like a clerk who won’t quote a rule they can’t point to in the book. Either attach retrieved ordinance spans or hand it to a human.
Full Explanation
A tax FAQ wrapper must require citation spans from retrieved ordinance chunks or return 'needs human review' when those spans are missing. Grounding answers in retrieved municipal ordinance text reduces hallucinated section numbers and gives residents (and auditors) a verifiable trail. If retrieval returns nothing usable, refusing to invent citations and escalating to human review is the correct civic behavior. Always inventing plausible section numbers when retrieval returns nothing manufactures false legal authority and is a high-severity failure for tax guidance. Hiding the lack of citations behind a confident tone and a longer paragraph increases persuasive hallucination without adding evidence. Replacing citations with generic 'see tax code' footnotes without chunk IDs provides no retrievable span and fails auditability even when the prose sounds careful. Exam caveat: requiring citation spans assumes a working retrieval corpus; if ordinances are outdated or poorly chunked, citations can be real yet wrong on policy—pair wrapper rules with corpus freshness checks. Operational check: for each FAQ answer, assert presence of chunk IDs and character or sentence spans from retrieval; route null-retrieval cases to human review queues and measure how often the model attempts citation-free confident answers.