Open hustshawn opened 3 weeks ago
Thank you for your suggestion! While I think it's a good suggestion, I'm concerned that with innovative updates like the multimodal chat in claude3, the Converse API may lag behind in terms of support.
Currently, since there are only two models, Claude and Mistral, for model invocation, we're not having any maintenance issues. Particularly for streaming, we recently refactored it (Also unit test can be found here), so we don't see much value in migrating. However, if there is a significant need to add models other than Mistral, we may consider adopting the Converse API as you suggested. How do you think?
It looks a really neat refactoring, and I was not aware of it before. Thank you for mention it.
Yeah, I understand you may not consider to prioritize it at the moment. I can think about potential timeline to go with the Converse API could be either the bedrock invoke API is deprecated, or there is demand to support extra models like Llama3.
I suppose it's good to keep this ticket open, and see if people have same interest or requirement on extra models support.
The converse API is updated: supports document
https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html
Describe the solution you'd like
Amazon Bedrock just released the converse API, which provides a consistant way to invoke models and genereate response.
Why the solution needed
Currently the project using Anthropic client for Claude models, and boto3 bedrock client for Mistral models. Both of them are way different in handling invocation and response parsing. It would be ideal to leverage the Converse API to make the code base clean.
Additional context
Converse API: https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html
Implementation feasibility
Are you willing to discuss the solution with us, decide on the approach, and assist with the implementation?