Which Solution Detects Hallucinations in a Bedrock Clinical Knowledge Base Before Release?

Why the Answer Is Correct

Option B satisfies all three stated requirements. A reference dataset of validated question-answer pairs built from approved clinical guidelines supplies the verified medical information against which model responses must be analyzed. Output diffing with Amazon Bedrock Guardrails for response consistency analysis compares answers to the same clinical question phrased in different ways, which directly addresses the semantic-inconsistency requirement. Amazon Bedrock evaluation then systematically detects factual inaccuracies before release, making this a pre-release validation pipeline rather than runtime monitoring.

Why the Other Options Are Wrong

Option A depends on predefined keywords and Lambda comparisons against a DynamoDB table, which only match explicit strings and cannot judge semantic consistency when a question is paraphrased; it is also a runtime monitoring design, not a pre-release evaluation. Option C misuses two services: Guardrails custom rules are denied-topic and word filters that flag explicit patterns, not fabricated medical facts, and SageMaker Feature Store holds ML feature vectors, not a repository of clinical guideline documents. Option D's automatic model evaluation applies built-in metrics to a custom prompt dataset rather than scoring responses against validated clinical answers, and CloudWatch anomaly alarms with SNS notifications amount to post-release monitoring instead of analysis against verified information.

Community Comment Notes

TNT87's comment selects B, reasoning that you should "Create a reference dataset with validated question-answer pairs from clinical guidelines" and pair it with Guardrails output diffing and Bedrock evaluation — a mapping that matches the three requirements one-to-one. No commenter defends the keyword-based, pattern-rule, or anomaly-alarm alternatives, so the community consensus aligns with this verdict.