Adopting langchain generations format

[krrishdholakia]

Wondering if it might make more sense to adopt something similar to the langchain Generations format

• LiteLLM output
• original model provider response

Reason I like this is because it gives me an easy way to tell if something went wrong - who was at fault (could maybe also include the exact input params to the model, for easy repro).

[krrishdholakia]

cc: @ishaan-jaff

[krrishdholakia]

https://github.com/refuel-ai/autolabel also has an interesting approach here

• response
• error (if any)
• cost