This article outlines best practices for configuring Denied Topics in Amazon Bedrock Guardrails to prevent common over-blocking or under-blocking issues. It provides strategies for reframing topic definitions from broad subject areas to specific user intentions, minimizing interference between topics, and establishing effective evaluation sets.
Key points
Define topics by specific user intent or action rather than broad subject domains to improve classification precision.
Avoid using keyword-based definitions, as semantic matching captures meaning better than list-based approaches.
Offload non-topic signals like jailbreak attempts or sensitive information patterns to their dedicated guardrail policy types.
Maintain a localized evaluation set comprising both clear matches and near-boundary non-matches to validate configuration changes.
Use a structured definition template to ensure topics have specific objects, qualifying constraints, and explicit sub-cases.