I'm curious about this sort of use case - how long does it take for a GPT-based system to process a bunch of documents it's never seen before so that you can pe

I'm curious about this sort of use case - how long does it take for a GPT-based system to process a bunch of documents it's never seen before so that you can perform NL searches on them? Assuming something ike a million total pages of text are we talking minutes? Hours? Days? And is it at all feasible to ensure whatever factual information is returned is only sourced from said documents, vs being "hallucinated" by virtue of whatever weights exist based on the core training corpus?