RAG citations

Two hosts walk through why a citation has to be a navigation target rather than a footnote, the three surfaces a retrieval system actually needs, and why streaming an answer is a trust signal rather than a UX flourish.

A: Most retrieval demos fail in exactly the same quiet way. The answer sounds fluent, sounds authoritative, and there's no way to actually check it.

B: In a consumer chat, that's a mild annoyance. In an HR or safety context inside a company, it's a liability — because whoever repeats the answer becomes responsible for it.

A: Which reframes the whole engineering problem. The failure mode buyers actually push back on isn't wrong. It's unverifiable.

B: A wrong answer that shows its source gets corrected in seconds. A plausible answer with no provenance has to be independently re-researched — which is exactly the work the system was supposed to remove in the first place.

A: So the design rule is that a citation isn't a footnote. It's a navigation target. Clicking it has to land on the actual page, with the passage visibly marked.

B: That sounds cosmetic, but it's not — it changes the data model entirely. If a citation only has to look right, a document ID is enough. If it has to resolve, you need the passage stored per document and per page, so the viewer can render it and place the highlight deterministically.

A: And retrofitting that later is genuinely unpleasant, because chunk boundaries and page boundaries aren't the same thing. Reconciling them after the fact means re-ingesting the whole corpus.

B: So — decide early whether citations need to resolve. If they do, that's schema, not a display feature you bolt on afterward.

A: There's a three-surface idea here too, that's easy to miss. A chat box is the smallest part of a system like this — three different people actually evaluate it.

B: The assistant, for end users — does it stream, cite, resolve to a page. The source viewer, for reviewers and compliance — does the claim exist in the original document. And the ops screens, for the team running it — what's indexed, what failed, what does re-indexing cost.

A: That third one's the one most demos skip, and it's usually the one that actually decides procurement. Knowledge-ops teams aren't evaluating answer quality in the abstract — they're asking who gets paged when ingestion silently drops twelve documents.

B: One more thing worth flagging — streaming. Answers rendering token by token isn't a UX flourish, it's a trust signal.

A: Text that appears instantly and complete reads as canned — people assume a lookup table. A stream reads as a model actually working through the question, right now, over this corpus.

B: Same content either way. The delivery changes whether people believe it's reasoning at all.

A: One honest limit — citation coverage isn't answer quality. A system can cite correctly and still retrieve the wrong passage. Provenance makes errors findable. It doesn't make them rare.

B: One thing to check this week.

A: If you're evaluating a retrieval system — yours or a vendor's — ask to see a wrong answer, and ask how you'd have caught it. If nobody can produce that, the citation piece probably isn't built yet.

B: Written version's at atinnovators.in, under insights.