Yeah, I've been monkeying around with an AI tool to help me query piles and piles of PDF files - bazillions of specification documents and suchlike - and if the

Yeah, I've been monkeying around with an AI tool to help me query piles and piles of PDF files - bazillions of specification documents and suchlike - and if there's a single thing I've taken away from it, it's that I can't expect the same answer twice, even for questions that have just the one correct answer. That's all after preprocessing those PDFs, because there were some characters and sequences that blew up the works. That took some hunting. I did learn that FOSI processors love injecting x00 null characters into the PDF for some godforsaken reason. Didn't know that. Don't know why, though, and probably never will. Falling back on older school techniques - from stats, text mining, topic models and other arcana - for some tasks is probably going to be worth it, for a while anyway. Most users can't use any of this stuff, it's too CLI-ish for the vast majority of the people doing actual work. I suspect this is the gap the OP is hoping to insert themselves into, that gap between these . . these so-called "AI" things . . and the average user, but they're going to need to do a better job of scoping out just how big a job that is. Now take all this with the following boulder o salt: we can't use any of the big cloud things, everything we do here has to be brewed up on prem, by people like me who do not know what we're doing. So we're not exactly AI superstars. Odds are everything I say is irrelevant for the outside world, with all your Geminis and whatnot.