Document Sections: Better rendering of chunks for long documents

[SomebodySysop]

@stevenic

That’s one of the things I’m wanting to explore. I believe the technique of asking the model to make some conclusions about the various sub parts of the document should work. I may do some experiments this week just to work out what the prompt needs to be.

I’ve had some success using gpt-3.5-turbo-16k to do something similar: creating both rankings and categorizations of query log records. It is taking some back and forth with the prompting – mostly because you’ve got to tell gpt-3.5 what NOT to do almost as much as what TO do – but I am satisfied with the results so far.