Before you can use Amazon Bedrock Guardrails, you must fulfill the following prerequisites:
-
Request access to the model or models with which you want to use guardrails.
-
Ensure that your IAM role has the necessary permissions to perform actions related to Amazon Bedrock Guardrails.
To prepare for the creation of your guardrail, consider preparing the following components of the guardrail in advance:
-
Look at the available content filters and determine the strength that you want to apply to each filter for prompts and model responses.
-
Determine the topics to block, consider how to define them, and decide which sample phrases to include. Describe and define the topic in a precise and concise manner. When you define denied topics, avoid using instructions or negative definitions.
-
Prepare a list of words and phrases (each up to three words) to block with word filters. Your list can contain up to 10,000 items and be up to 50 KB. Save the list in a .txt or .csv file. If you prefer, you can import it from an Amazon S3 bucket using the Amazon Bedrock console.
-
Look at the list of personally identifiable information in Remove PII from conversations by using sensitive information filters and consider which ones your guardrail should block or mask.
-
Consider regex expressions that might match sensitive information and consider which ones your guardrail should block or mask with the use of Sensitive information filters.
-
Develop the messages to send to users when the guardrail blocks a prompt or model response.
