also, I'm curious what the last generation of models you used is and to what extent. blanket-level "hallucination" isn't really something I run into often anymore. inconsistencies or obfuscation, sure, but that's not the same.
Personally, I'm using Qwen, Phi, etc. to squeeze as much performance out of constrained hardware as possible. They're pretty good at searching my personal files which you mentioned in a previous comment.
edit: typos