See how important each token of the context was for the LLM response

[LiquidGunay]

I think the ability to get something like an average attention score for each token of the context would be really useful to see what parts of the context did the LLM "focus" more on. This would be fairly useful for RAG and QA applications.