Richer ChatModel.invoke,stream,batch etc outputs

[baskaryan]

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

Richer ChatModel.invoke,stream,batch etc outputs

Motivation

Current the main Runnable methods on ChatModels return a Message (or MessageChunks, or list of Messages, etc.) This loses some of the model response information. For example if there are multiple candidate generations, or if logprobs are returned. There should be an easy way to use ChatModel's runnable interface and get the full model response

Proposal (If applicable)

The two main ways I can think to do this are:

1. Have an invoke argument ChatModel.invoke(..., return_type="...") that let's you determine what the output type. You could then do ChatModel.bind(return_type="...") to create an object that always returns your chosen type. Here's a sketch of what this could look like: rfc: return full llm res #17188
2. Have some ChatModelLLMResult wrapper class that returns LLMResult instead of Message and a ChatModel.return_llm_results() class method that generates the ChatModelLLMResult for that ChatModel

Would love to hear other ideas and thoughts on the ideal solution