Harrison Chase: evaluation

Transcript context

…What about the other side of open source? So on Acker News, there were a couple blog posts recently, like the problem with LangChain and LangChain is pointless, all these different things. So the TLDR of some of them were, the LangChain API is like kind of verbose and complicated versus like sometimes I can just do this in like 10 lines of code. How do you balance that in terms of allowing for the complex use cases versus making maybe the ergonomics like simpler, but then trading that off later? There's a lot to balance and there's a lot to do. And I think like posts like that are very valuable to hear basically what people are saying. And like, we have a lot of open issues. So it's not like these things hadn't been said before, but I think like that was a good emphasis on what people are saying. And I think there was a lot of things in there. I think part of it's kind of like around and we took all of it very seriously. And yeah, I think there's a lot to dive into there. There's like the documentation piece. And so I think we did a revamp of the documentation to address that. There's also like a comment in this, I think this was around, I think the top comment on the LangChain is pointless one was like basically like orchestration is like 5% of the work. And then like the other 95% is like prompt engineering and like data engineering. And those are the hard bits. I think maybe orchestration is a little bit more than 5%, but I like agree that those are like really big pain points that get exacerbated when you have these complex chains and agents where you can't really see what's going on inside of them. And I think that's partially why we built Langsmith to help out with exactly that. We also needed to do better things like make the prompts more visible and make it allow for more customizability around that. And so we've tried to add some stuff there. In terms of balancing, there's also LangChain is pointless. I don't need a wrapper. I can just call the underlying API. I think if all you're trying to do is call the underlying API, then like, yeah, that's gonna be the cleanest and simplest thing to do. And we try to get as close to that experience as possible, but we're not optimizing for calling the API. We're optimizing for helping people build context-aware reasoning applications as easily as possible. And so there's some level of abstractions that you need to add in order to assist in that. Yeah, that's definitely a balance that's tricky to strike, but I think there's also some aspect of it. Like, I do think one of the big benefits that LangChain provides is a standard interface for language models so that you can switch between them. And this kind of gets into like an ORM debate, like are ORMs generally kind of like useful or not? And so I think in this case they are. I think there's probably a larger kind of like philosophical kind of like question about that that people have strong opinions on. Just the prompts don't transfer like you also mentioned. Yeah, yeah, there's that, yeah.…

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