Can we get the same performance using less tokens ?
maybe ask_in_place() could provide a diff rather than full files ? that would dramatically reduce some queries
maybe ask_smart could use only the doc and not the full code, we could improve the doc until it works well enough, maybe that'd be a cleaner signal too. We would not have to copy the source too
we need a way to assess the performance ideally, to make sure these don't deteriorate the performance
Can we get the same performance using less tokens ?