An insurance company has deployed a policy information chatbot using Amazon Bedrock with a Knowledge Base containing all their policy documents. The chatbot must NEVER provide incorrect coverage amounts or policy terms, as errors could constitute regulatory violations. During testing, the model occasionally "fills in" details not present in the retrieved policy documents — for example, stating a deductible amount when the retrieved chunk only mentions the policy type without specific numbers. Which Bedrock Guardrails configuration best prevents this type of hallucination?
Show answerHide answer
Bedrock Guardrails' contextual grounding check is specifically designed to detect when a model's response makes claims not supported by the provided reference material (source documents). By setting a high grounding threshold, responses where the model fabricates details (like deductible amounts not in the source) are blocked. This is the native Bedrock solution for preventing RAG hallucination — it compares each statement in the response against the retrieved context and scores whether the response is grounded.