A healthcare company is building a user support chat-based assistant on Amazon Bedrock.

Store approved health support guidelines in an Amazon Bedrock knowledge base. Configure system prompts that instruct the model not to provide diagnosis advice. Use an AWS Lambda function after inference to remove PII from the model response before the response is returned to users.

Fine-tune an FM on approved support conversations. Add prompt templates that prohibit diagnosis advice. Run a separate review process that scans transcripts for prohibited topics and sensitive information after the conversations end.

Build a custom moderation layer in the application to inspect prompts for prohibited topics and to redact sensitive information from user inputs. Invoke the model through the Converse API. Use separate post-processing logic to redact sensitive information from responses before displaying responses to users.

Create an Amazon Bedrock guardrail. Configure denied topics for medical diagnosis advice. Configure sensitive information filters to mask PII. Configure content filters. Call the ApplyGuardrail API on user prompts before inference. Include the same guardrail in the Converse API to evaluate model responses.