Support for Bedrock Application Inference Profiles

[bioshazard]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Please add support for use of Application Inference Profiles as model in Bedrock client.

Motivation

My organization has adopted Application Inference Profiles to track costs and usage within a shared AWS account. I therefore need to be able to supply the Application Inference Profile arn as the model.

Proposal (If applicable)

I see that support for Cross Region Inference was added in 29c5b8c but it expects a region-optional model ID as the model.

I need support for Application Inference Profiles if I am to continue using LangChain JS. Per the docs, this requires permitting the full ARN of the application inference profile as the model name. I therefore need the filter to include accommodation for inference profile ARN as model name. I will attempt to contribute this myself as I have urgent need for enabling this.

[bioshazard]

I would also accept a simple ENV var that makes it possible to simply bypass the model provider while this accommodation is pending.

[bioshazard]

I will try to just add a "or matches app inf regex" to the condition triggering... honestly not clear to me why we check at all... its not a typed check at param to be useful at init and backend will catch and error anyway.

[bioshazard]

Still trying to get it the additional condition working locally but I might pitch a separate PR where I just delete this check outright and let Bedrock handle its own exceptions. Curious to learn why that isn't a good idea especially since the restriction isn't revealed in the types as of now.

https://github.com/langchain-ai/langchainjs/blob/main/libs/langchain-community/src/llms/bedrock/web.ts#L113

[bioshazard]

Gah... looks like if I bypass that condition it just fails silently. So maybe a useful condition afterall. Also here is the correct file that actually gets used for this:

https://github.com/langchain-ai/langchainjs/blob/main/libs/langchain-community/src/chat_models/bedrock/web.ts

[bioshazard]

Delivered #7822 for this. Wound up adding a new field modelAlias to mediate between deriving metadata from the modelId and using application inference profile ARN in the final /invoke URL call.

[AllenFang]

You can consider to use ChatBedrockConverse from @langchain/aws, I try and it work well. the bedrock integration form @langchain/community is old.

[AllenFang]

if you can't use newest @langchain/aws for some reason, how about switch to chat_models, it also works to me

const { BedrockChat } = require("@langchain/community/chat_models/bedrock");

async function run() {
    const llm = new BedrockChat({
        model: "us.anthropic.claude-3-5-sonnet-20240620-v1:0",
        region:  "us-east-1",
        credentials: {
            ...
        },
        ...
      });
      
      const inputText = "...";

     const completion = await llm.invoke(inputText);
     console.log(completion);

[bioshazard]

Thanks, Cross region interface model works. Now try an application inference profile. Will check /aws tho if community is old