AWS published a guide on the AWS Machine Learning Blog outlining eight best practices for writing Denied Topics in Amazon Bedrock Guardrails. Using a fictional bank assistant as an example, the authors showed how broad subject labels can inadvertently block legitimate user requests.
To prevent over-blocking, the authors recommend using action-scoped topic names and definitions, providing varied boundary examples, and carefully managing overlapping topics. They also advise using other Guardrails policies to handle signals such as prompt attacks or exact terms, noting that results can still vary for borderline inputs and when the topic list changes.
AI 根据报道生成 · 9 小时前更新