KB response providing false information and answering things it does not know.
Hey, I’m having trouble regarding the KB response providing false information and answering things it does not know. This destroys so much value, especially when providing for clients wich need almost 100% source security. I tested the KB when I removed all of my web urls and only then did it respond with real Information and nothing more than the reference information. Is there a way to stop this?
3 Replies
Bad info in gives bad info out. Also, if you prompt the LLM to provide info that is not in the retrieved chunks, it will hallucinate.
This is not a VF issue. We have very little hallucinations. BTW, if you expect an LLM to be 100% accurate (see Google's latest search issues), it is hard. You will need to do a lot of follow-up checks of each response.
Yeah, the client didnt really provide me with good info. Here is my prompt. how can i improve it so it does not hallucinate?
Yeah I've seem those, kind of scary