A citation must support the sentence beside it A document assistant can attach a legitimate source link to an unsupported answer. The presence of a citatio...
A citation must support the sentence beside it
A document assistant can attach a legitimate source link to an unsupported answer. The presence of a citation does not prove the answer was grounded. Evaluation must check whether the cited passage exists, is accessible to the user, is current, and actually supports the specific claim.
Build an evaluation set with questions, approved source passages, expected answer elements, and cases that should be refused. Include conflicting versions of a document and questions requiring multiple sources. A simple answer string match is not enough when several valid phrasings exist.
Split the pipeline into checks
First measure retrieval: did the correct passage enter the context? Then assess answer grounding: can each material statement be traced to one or more cited spans? Finally assess citation presentation: does the link open the expected document and section for an authorized user?
If retrieval failed, change ingestion, chunking, or ranking. If the right source was present but the model cited the wrong one, investigate prompt, context layout, and answer generation. Keep these failure categories separate in reports.
Include negative examples
Ask about a fact absent from approved sources and verify the assistant says it cannot determine the answer. Ask about another tenant's private document and verify it never enters retrieval results. Insert a source passage with an instruction that tries to override the assistant's behavior and verify the application refuses unauthorized actions.
Citation scoring should penalize a correct sounding answer with a fabricated or irrelevant source. Human review remains important for subtle claims and business policies.
Monitor after release
Sample answers for periodic review. Record source document IDs, versions, retrieval scores, and feedback without retaining more personal data than necessary. When a cited document changes, mark prior answers as potentially stale if the product keeps conversation history.
Define a correction workflow: update the source, reindex it, rerun the relevant evaluation cases, and communicate a high impact correction when required.
Conclusion
Useful citations are evidence links, not decoration. Test retrieval, grounding, and link accuracy separately, and include refusal behavior. Our AI automation service covers the surrounding permissions and review workflow.
References