How to implement AWS Guardrails (or get Experimental_LanguageModelV1Middleware working)

[matthew-gill]

I have the following guardrail config I want to add to my request to AWS Bedrock:

guardrailConfig: {
	guardrailIdentifier: 'user_provided_content',
	guardrailVersion: '1',
	trace: 'enabled',
},

But I can't seem the right place to add it in createAmazonBedrock or streamText, I believe it needs to end in the body of the request.

I've tried all sorts of combinations of bedrockOptions and additionalModelRequestFields all with no luck.

Is this possible to do?

I've been looking into the Language Model Middleware documentation, but I get an error if I try to import the experimental_wrapLanguageModel: TS2305: Module "ai" has no exported member experimental_wrapLanguageModel

I think I'm up to date with the ai package: [email protected]

No doubt I've missed something basic, but any help would be really appreciated, thanks!

[matthew-gill]

I've just bumped to 3.3.31 and I can now export the experimental_wrapLanguageModel 🎉
I'll have a go at implementing the AWS guardrails with that, but if we can add the ability to define guardrails in the SDK, that would be brilliant.

Once I've got something useful, I'll raise a request in the repo 👍

[lgrammel]

You have seen #2935 right? I expect guardrail support to land in the next few days.

[matthew-gill]

Yay! Amazing, thank you!

[lgrammel]

Guardrails are available in @ai-sdk/[email protected]

https://sdk.vercel.ai/providers/ai-sdk-providers/amazon-bedrock#guardrails