Stopping benefit answers that the plan documents do not support
MLA-C02
Operating, Monitoring, and Securing ML and AI Solutions
Medium
Question 15 of 30
A health insurer runs a member assistant that answers benefit questions from passages retrieved from its plan documents. Auditors found responses that state coverage limits that appear in no retrieved passage. The insurer must block responses that the retrieved passages do not support, or that do not address the member's question, before the responses reach members.
Which solution will meet these requirements?
Show answer and explanation
DConfigure an Amazon Bedrock guardrail contextual grounding check with grounding and relevance thresholds
A contextual grounding check meets the requirements. Its grounding score measures whether a response is supported by the reference source, and its relevance score measures whether the response addresses the query; responses below the thresholds that the insurer sets are filtered before they are returned. A sensitive information filter masks PII, which is a different problem from unsupported claims. A temperature of 0 makes output less random, but a consistent answer can still state a limit that no passage supports. Invocation logs reviewed weekly find problems after members have already seen them.
Source: docs.aws.amazon.com/
Anonymous, no account needed. Your tally stays in this browser.