A first-line sweep of a conveyancing handbook: grounded answers whose citations are machine-checked against the retrieved text before you see them — so the citation leads you to retrieved text, on a page that text covers, to read and judge for yourself. (The check confirms the citation falls inside what was retrieved; it does not confirm the exact paragraph exists or that it supports the claim.)
Staff phrase questions colloquially; the handbook doesn't. Three quick rewrites bridge the vocabulary gap. The original question always keeps the loudest vote in retrieval.
Legal questions hinge on exact tokens ("Form VX-9", "s.72"); keyword search catches what semantic search fuzzes past — and vice versa. Reciprocal-rank fusion merges both rankings.
The model must answer from the retrieved extracts only, citing paragraph and page for every statement — or refuse with one exact sentence.
Each cited (paragraph, page) must match a chunk that was actually retrieved: a related section number, on a page that chunk covers. This is the fabricated-locator safeguard: a locator pointing outside what was retrieved has nothing to match, so it cannot pass. (It proves the citation falls inside retrieved text — not that the exact paragraph exists, nor that the text supports the claim; that is the reader's judgment, at the cited page.)