Grounding without hallucinated citations
Retrieval makes answers better. It also makes them more confidently wrong, unless you check that the citation actually supports the claim.
Nodexa team9 min read
Adding retrieval to an agent reliably improves accuracy. It also introduces a failure mode that is harder to spot than plain invention: an answer that cites a real document which does not say what the answer claims.
Require the span, not the document
Asking for a document-level citation is easy to satisfy and easy to fake. Asking for the specific passage forces the model to point at something checkable, and makes a bad citation obvious to a reviewer in seconds.
Measure it deliberately
We sampled 500 grounded answers and had two reviewers check whether the cited span supported the claim. Span-level citation moved the support rate from 82% to 96%. The remaining 4% were mostly answers that should have been refusals.