Amazon Bedrock Guardrails can be invoked directly via the ApplyGuardrail API without making calls to foundation models. Additionally, its sensitive information filters support both probabilistic detection of entities (such as SSNs) and configuring custom regular expression patterns. Finally, reasoning content blocks are explicitly excluded from Guardrails protections.
With the Standard tier, Amazon Bedrock Guardrails extends its ability to detect undesirable content inside code elements, which specifically includes comments, variable and function names, and string literals. Reasoning content blocks are always excluded.
According to the official AWS documentation, the capability to evaluate only a specific section of an input prompt (such as in RAG or conversational setups) is exclusively available through the AWS SDK. This selective evaluation capability is not supported in the AWS Management Console, including the Bedrock Playground or the Bedrock Guardrails management console.
Contextual grounding checks are specifically designed to help detect hallucinations in model responses. They identify responses that are not grounded (either because they are factually inaccurate or add new information) in the source document, or are irrelevant to the user's query.
A company is designing a safety validation pipeline. They need to evaluate user inputs for sensitive information, such as social security numbers and custom proprietary patterns, and mask them. To minimize latency and cost, they want to perform this validation directly on the inputs without invoking any foundation models. How should they configure this solution using Amazon Bedrock Guardrails?
How confident are you in this answer?