Amazon Bedrock Guardrails: Automated Reasoning & Policy Refinement

Today, AWS announces new automated refinement workflows for Automated Reasoning checks in Amazon Bedrock Guardrails. Automated Reasoning checks use formal logic to mathematically validate the accuracy of generative AI responses against a policy you define, helping detect hallucinations and provide verifiable explanations. The quality of validation results depends on how well a policy is defined. The new workflows help customers improve their policies with less manual effort, leading to more reliable Guardrail validation results.

The launch introduces two refinement workflows. With the iterative policy improvement workflow, customers who have created natural language tests for a policy can start an iterative refinement run, letting the system deduce the changes needed for the policy to pass those tests. With the ambiguity reduction workflow, customers who frequently encounter ambiguous translation results can run the resolve policy ambiguities workflow to automatically refine variable descriptions and type definitions, reducing how often ambiguous translations occur. Both workflows are available through the Amazon Bedrock APIs and in the AWS Management Console, where customers can start a workflow by choosing Refine policy on the policy page.

These workflows are available in all AWS Regions where Automated Reasoning checks in Amazon Bedrock Guardrails are available. To learn more, visit the Amazon Bedrock Guardrails product page and the Automated Reasoning checks User Guide.

Categories: general:products/amazon-bedrock,marketing:marchitecture/artificial-intelligence

Source: Amazon Web Services