In this article
Three links below an assistant’s answer do not automatically make it well supported. A working link can lead to a page that does not establish the quoted number or a policy for a different product. Citation quality depends on the relationship between a specific claim and its evidence.
The NIST Generative AI Profile recommends reviewing and verifying sources and citations during pre-deployment evaluation and ongoing monitoring. The process below is a proposed way for a business team to apply that principle.
Break answers into checkable claims
Begin with figures, dates, conditions and exceptions that could influence a customer decision. Connect each material claim to a specific passage in an authorised source. A general product page is not evidence for every detail about the product.
In a hypothetical example, an assistant promises same-day maintenance and links to the support page. If that page only states contact-centre opening hours, it does not establish the maintenance promise even though the link works.
Assess coverage and support separately
Coverage asks whether each important claim has evidence. Support asks whether the cited passage actually establishes the statement. An answer can appear fully cited while its links do not support its sentences. Also record contradictions and sources that cannot be opened using the permitted access level.
Keep an audit trail
Record answer identifier, claim, source link, relevant passage, review date and reviewer verdict. For changing documents, retain an appropriate version reference under the organisation’s policy. Do not copy sensitive internal material into an unauthorised external reviewing service.
Turn verdicts into changes
When support is absent, remove the claim, narrow it to what the evidence establishes or send it for human review. Then determine whether retrieval selected the wrong document or generation stretched the meaning of a correct one. Those failures require different corrections.
Start with twenty answers, deliberately including questions whose mistakes would matter most. Recheck after changes to knowledge or retrieval. Combine this audit with knowledge-freshness testing: a relevant source can still be out of date. The acceptance test is whether a reviewer can trace a sentence to evidence and understand the limits of that evidence.
Contact on WhatsApp
