Prevent factual errors from LLM hallucinations with mathematically sound Automated Reasoning checks (preview)

AWS has introduced Automated Reasoning checks in Amazon Bedrock Guardrails to help prevent factual errors and hallucinations in generative AI responses. This new preview feature uses mathematical and logical verification to validate the accuracy of large language model outputs.

• Uses mathematical proofs to verify system behavior
• Allows creation of policies that encode organizational rules and guidelines
• Helps validate responses in domains requiring high accuracy like HR policies and operational workflows
• Provides explanations for why responses are correct or incorrect
• Currently available in preview in the US West (Oregon) region

The feature enables users to upload documents, automatically generate reasoning policies, and test responses against established guidelines, offering a new approach to ensuring AI-generated content reliability.