Question 430
At fictional Silver Bay Design Office, a Snowflake data and AI engineering team is conducting a regulated customer response automation exercise with regional service policies and accountable stakeholders. The separately composed technical incident is: A knowledge assistant passes an internal fluency review but its generated citations occasionally point to unrelated passages that do not support the claim being made. Assume only these documented facts and use safe, authorized production-change procedures; exact current Snowflake feature syntax and region or account availability require current review. Which repeatable assurance or control is most relevant to preventing this situation?
Explanation
Writer-proposed answer: B (UNVERIFIED)
Preliminary individual option explanations — NOT expert validated
Correct Answer
Require citation-level evaluation gates, negative-source tests and periodic review of production error reports. Citation presence is not sufficient proof of grounding; links must support the specific factual claims in context.
Incorrect Answers
This response concerns Unreviewed language model version drift, which is different from the stated failure. A knowledge assistant passes an internal fluency review but its generated citations occasionally point to unrelated passages that do not support the claim being made.
Agent tool execution requiring approvals is not the cause of the reported failure: A knowledge assistant passes an internal fluency review but its generated citations occasionally point to unrelated passages that do not support the claim being made.
Sensitive personal information in model prompts alone does not establish the needed evidence: Independent review shows each consequential answer statement is supported by the referenced passage or the model abstains clearly.