Citation Grounding
Also called Answer Attribution
Making the model point to the exact source for each claim, so answers are checkable.
Think of it like
A term paper with footnotes — you can flip to the reference and confirm the writer didn’t make it up.
Example
A research assistant answers "revenue grew 12%" with a clickable link to the sentence in the 10-K that says so.
How it actually works
Citation grounding ties each generated statement back to a retrieved passage, usually by having the model reference chunk IDs it was given. It boosts trust and makes hallucinations catchable, since a claim with no supporting source is a red flag. The hard part is faithfulness — models can cite a source that doesn’t actually support the claim, so citations reduce but do not eliminate the need to verify.
For product teams
The feature that turns "trust me" into "check for yourself"; a big driver of user trust in RAG.
For engineers
Claim-to-chunk linking via referenced IDs; improves verifiability but citation faithfulness itself must be checked.
Related
- Grounding — Builds on the broader idea of grounding.
- Source Attribution — Related to source attribution.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome