One of the things that makes #LLM #Chatbots seem like they can carry on a conversation is that they adjust their state according to the conversation as it evolves, and respond in light of previous comments from both participants.
This may be good for a conversation, but naively implemented, it's quite a bad thing for a search tool / knowledge engine.
Here's an example. Here one answer Bard gives when you ask it what I think of the InfoMap algorithm that I co-developed with Martin Rosvall.
@ct_bergstrom Is this GPT-4?
@radehi Bard.
@ct_bergstrom Right, sorry.
@ct_bergstrom Oh, Bard, I see. I saw that someone had used similar tricks to get GPT-3.5 to invent a detailed story about Gerald Ford shooting a cabbage on the White House lawn.