Shortens each context to the parts relevant to its question, returning results in input order.

Results are cached per (context, question, prompt version), so repeated requests do not re-call the model.