By using AWS re:Post, you agree to the AWS re:Post Terms of Use
AWS re:Postre:Post

Multiple guardrails

0

All, in accordance with my research only one guardrail can be used in knowledge base, agents and chats. Does anybody know why do we have this limitation? i see the same limitation in both console and API. Possible use case: an enterprise level guardrail for all enterprise app (PII etc) and additional one or more on application level. While It may be possible to implement this use case by using files etc for guardrails, but multiple guardrails may be better solution. Any information would be greatly appreciated Thank you YK

Topics
Machine Learning & AIGenerative AI on AWS
Tags
Amazon Bedrock
Language
English
YK
asked a month ago89 views
1 Answer
0

Hi,

The configuration of a guardrail is based on policies: see https://docs.aws.amazon.com/bedrock/latest/userguide/quotas.html

- A guardrail is a combination of multiple policies configured for prompts and response 
including; content filters, denied topics, sensitive information filters, and word filters.

- A guardrail can be configured with a single policy, or a combination of multiple policies.

So, can you define the common part that you want to share in a policy that you re-use across multiple guardrails?

Best,

Didier

profile pictureAWS
EXPERT
Didier_Durand
answered a month ago
  • YK
    a month ago

    Didier, Thank you for a quick answer and great explanations! Let’s consider example of a company “C” and two departments “A” and “B”. In API terms, Company security requirement is to use couple of filtersConfig (types like HATE, MISCONDUCT etc.) and sensitiveInformationPolicyConfig. https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-create.html#:~:text=To%20create%20a%20guardrail,Guardrails%20section%2C%20select%20Create%20guardrail.

    Departments “A” and “B” want to use additional separate wordsConfig. In that case “A” creates guardrail based on a copied “C” policy with additional its own wordsConfig. Same with “B”. They also need to consider versioning of changing “C” policy… Is it what you referring to?

    Maybe it is the matter of approaches, but would it not it easier to create guardrail based on “C” policy and individual department guardrails? And combine them as needed?

    However, maybe combine approach would be more costly and less quotas friendly? Thank you again YK

Relevant content