PRODUCTION SCENARIO
A pharmacy's retrieval augmented generation assistant sometimes adds dosage details that do not appear in the retrieved leaflet. The company wants those responses blocked automatically and wants to tune how strict the check is.
Which Amazon Bedrock Guardrails policy will meet these requirements?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: A contextual grounding check with a configured threshold
A contextual grounding check scores each response against the supplied source, treats new information that is not in the source as ungrounded, blocks responses below the configured threshold, and gets stricter as the threshold rises. A regex in the sensitive information filter would mask dosage numbers whether or not the leaflet supports them. A denied topic would block every dosage answer, including correct ones. Word filters match fixed strings and cannot judge support.