this is, I think, one of the inherent problems with LLMs using natural language interfaces: they are a statistical tool *that looks like* a logical tool statistical approaches can be a useful way to find related information! BUT for it to be useful you have to *know that's what it's doing*