not sure if I like this approach

what are you looking at in particular here?

the feature embedded into normal llms

and I think it would be maintainance headache down the road anyway if any api changes

and you have to reimplement everything again when existing llms add support

like some providers in bedrock already support multimodal, do you want to maintain the complex mechanisms for both, just for an extra parameter?

Ohhhh the mutli-modal LLM stuff?

Yea its super janky. Would like to reduce the code duplication there