Trust needs traces
A grounded answer is only trustworthy if a user can check it. Citation and attribution make a RAG system tie each statement in its answer back to the specific retrieved passage that supports it, so claims can be verified rather than taken on faith.
How citations are produced
- Identifier tagging. Each retrieved passage is labeled with a short id, and the model is instructed to mark every claim with the id of its source.
- Span linking. The system maps a cited id back to the exact passage and document for display.
- Post hoc checking. A verifier can re embed the answer sentence and its cited passage to confirm the citation actually supports the claim.
Faithful versus plausible
A citation is faithful only if the cited passage genuinely entails the claim. A model can produce a plausible looking citation that does not actually support the statement, so attribution quality must be measured, not assumed.
Why it matters
Attribution turns an opaque answer into an auditable one. It lets users follow each claim to its source, exposes when the model drifted off its evidence, and is often a hard requirement in legal, medical, and enterprise settings.
Key idea
Citation and attribution tag each retrieved passage and require the model to mark every claim with its source, turning an opaque answer into one a user can audit, provided citations are faithful not just plausible.