A customer-service bot built on Amazon Bedrock must decline investment advice, block abusive language, and mask account numbers that appear in chat messages. Which Bedrock feature provides these safeguards?
Show answer and explanation
Answer: C
Bedrock Guardrails apply configurable safeguards to model inputs and outputs. They can deny specified topics such as investment advice, filter harmful content categories including abusive language, and use sensitive information filters to block or mask data like account numbers. Model evaluation compares model quality, provisioned throughput reserves capacity, and batch inference processes large offline datasets. None of those three enforces content and privacy rules during live conversations.
Why the other options are wrong
- A. Model evaluation jobs measure and compare how candidate models perform on test data, but they do not enforce rules on live chat inputs and outputs.
- B. Provisioned throughput reserves model capacity for predictable performance and cost; it does not inspect or filter what users or the model say.
- D. Batch inference runs model predictions over stored datasets asynchronously, which is useful for offline jobs but does not screen live customer conversations.