it is quite amazing how LLMs like claude v2 can in one turn give you solid responses about a document i uploaded and in the next turn when i ask about a specific person, the response is entirely made up. –> further to @bxchen recent piece, LLMs are useful when constrained to a