Harrison Chase: recommendation

Transcript context

…On one of the heuristics I wanted to get your thoughts on, hallucination is one of the big problems there. Do you have any recommendations on how people should reduce hallucinations? To reduce hallucinations, we did a webinar on like evaluating RAG this past week. And I think there's this great project called RAGOS that evaluates four different things across two different spectrums. So the two different spectrums are like, is the retrieval part right? Or is the generation, or sorry, like, is it messing up in retrieval or is it messing up in generation? And so I think to fix hallucination, it probably depends on where it's messing up. If it's messing up in generation, then you're getting the right information, but it's still hallucinating. Or you're getting like partially right information and hallucinating some bits, a lot of that's prompt engineering. And so that's what we would recommend kind of like focusing on the prompt engineering part. And then if you're getting it wrong in the, if you're just not retrieving the right stuff, then there's a lot of different things that you can probably do, or you should look at on the retrieval bit. And honestly, that's where it starts to become a bit like application specific as well. Maybe there's some temporal stuff going on. Maybe you're not parsing things correctly. Yeah. Okay. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.