Previewing retrieval: what the assistant would say, and why

Retrieval preview is a diagnostic tool, not a chat window: you arrive with a specific question already in mind, and the screen is built around answering exactly that — what would come back, and on what grounds.

Three things this shows that a live chat never would

Which passages, in fused order — each with its own citation naming the article, version and section it came from, never just a domain. Why each one ranked — the vector-search arm's rank and the keyword-search arm's rank are shown separately rather than collapsed into one score, because "why did this rank?" is the first question asked of any retrieval problem, and it's unanswerable once the two arms have been merged away. Whether the vector arm ran at all — if it didn't, that's normal for a language whose embedding backfill is behind publication, and the screen says so directly rather than leaving you to wonder if ranking itself is broken.

The override toggle is a diagnostic, not a resident-facing switch

By default, an override that matches your question fires here exactly as it would for a real resident — retrieval doesn't run, the same as it wouldn't in production. A separate toggle lets you see past it, to what retrieval would have found instead. That's for you, debugging why an article isn't ranking; a resident never gets this choice, because letting them see around a curated answer would defeat the reason the override exists.

Reading the diagnostics panel

Beneath the answer: the normalized form of your question, which glossary terms fired (linked straight to the glossary entry that caused it), how many candidates each search arm considered, and — if any overrides were in the running — which ones matched and which didn't, with the reason stated for each miss.