What is answer attribution?
Answer attribution names the source behind each part of an answer rather than the answer as a whole. Where three sentences derive from three different documents, each sentence carries its own reference. The reader can therefore check one clause without rereading everything the system retrieved.
Also called: Source attribution
Why a list of sources at the end of an answer is not attribution
Attribution is routinely confused with the footer of links that most retrieval systems already produce. A footer tells a reader which documents were consulted, leaving them to work out which document supports the one sentence they doubt. That work is exactly what a compliance officer or a clinician will not do, so the answer is either accepted whole or rejected whole. Per-claim attribution changes the unit of verification from the answer to the clause.
- Attribution per clause lets a reader verify the single statement they question.
- Mixed-source answers hide which document contributed the contested sentence.
- A clause with no available source becomes visible rather than absorbed.
- Attribution gives a reviewer a way to locate an incorrect source document.
How answer attribution works
- Retrieved passages are held with a stable identifier for each one.
- The answer is composed so that each sentence declares the passage behind it.
- Sentences drawing on two passages carry both references rather than the stronger one.
- Any sentence without a supporting passage is flagged before the answer is returned.
- References resolve to the passage location in the original document.
- The attribution map is retained with the answer for later inspection.
How answer attribution is evaluated
Answer attribution is assessed by checking each reference against the sentence it is attached to. Examine the following on answers already given to real questions:
- proportion of sentences carrying at least one resolvable reference
- share of references a reviewer judges to support the sentence attached
- frequency of attribution to the wrong passage within the retrieved set
- count of sentences returned with no reference at all
- reference resolution failures after a source document is reindexed
Limitations, risks and common failure modes
- A reference is attached to the nearest retrieved passage rather than the supporting one.
- Synthesis across passages produces a sentence no single source actually states.
- Attribution survives inside the interface and disappears when the answer is copied.
- Readers treat the presence of a reference as evidence that it was checked.