How Amazon Bedrock Guardrails works
Amazon Bedrock Guardrails helps keep your generative AI applications safe by evaluating both user inputs and model responses.
You can configure guardrails for your applications based on the following considerations
An account can have multiple guardrails, each with a different configuration and customized to a specific use case.
A guardrail is a combination of multiple policies configured for prompts and response including; content filters, denied topics, sensitive information filters, word filters, and image content filters.
A guardrail can be configured with a single policy, or a combination of multiple policies.
A guardrail can be used with any text or image foundation model (FM) by referencing the guardrail during the model inference.
You can use guardrails with Amazon Bedrock Agents and Amazon Bedrock Knowledge Bases.
When using a guardrail in the InvokeModel, InvokeModelWithResponseStream, Converse, or ConverseStream APIs, it works as follows during the inference call:
The input is evaluated against the configured policies specified in the guardrail. Furthermore, for improved latency, the input is evaluated in parallel for each configured policy.
If the input evaluation results in a guardrail intervention, a configured blocked message response is returned and the foundation model inference is discarded.
If the input evaluation succeeds, the model response is then subsequently evaluated against the configured policies in the guardrail.
If the response results in a guardrail intervention or violation, it will be overridden with pre-configured blocked messaging or masking of the sensitive information.
If the response's evaluation succeeds, the response is returned to the application without any modifications.
For information on Amazon Bedrock Guardrails pricing, see the Amazon Bedrock pricing
How charges are calculated for Amazon Bedrock Guardrails
Charges for Amazon Bedrock Guardrails are incurred only for the policies configured in the guardrail.
The price for each policy type is available at
Amazon Bedrock Pricing
-
If a guardrail blocks the input prompt, you're charged for the guardrail evaluation. There are no charges for foundation model inference calls.
-
If a guardrail blocks the model response, you're charged for guardrail's evaluation of the input prompt and the model response. In this case, you're charged for the foundation model inference calls, in addition to the model response that was generated prior to the guardrail's evaluation.
-
If a guardrail doesn't block the input prompt and the model response, you're charged for guardrail's evaluation of the prompt and the model response, in addition to the foundation model inference.
