Harrison Chase: evaluation

Transcript context

…Yeah. Do you have personal favorites that you want to shout out? The one that we have in the docs is the default is like the recursive text splitter. We added a playground for text splitters the other week because, yeah, we heard a lot that like, you know, and like these affect things like the chunk overlap and the chunks, they affect things in really subtle ways. And so like I think we added a playground where people could just like choose different options. We have like, and a lot of the ideas are really similar. You split on different characters, depending on kind of like the type of text that you have marked down, you might want to split on differently than HTML. And so we added a playground where you can kind of like choose between those. I don't know if those are like underappreciated though, because I think a lot of people talk about text splitting as being a hard part, and it is a really important part of creating these retrieval applications. But I think we have a lot of really cool retrieval algorithms as well. So like self query is maybe one of my favorite things in LangChain, which is basically this idea of when you have a user question, the typical kind of like thing to do is you embed that question and then find the document that's most similar to that question. But oftentimes questions have things that just, you don't really want to look up semantically, they have some other meaning. So like in the example that I use, the example in the docs is like movies about aliens in the year 1980. 1980, I guess there's some semantic meaning for that, but it's a very particular thing that you care about. And so what the self query retriever does is it splits out the metadata filter and most vector stores support like a metadata filter. So it splits out this metadata filter, and then it splits out the semantic bit. And that's actually like kind of tricky to do because there's a lot of different filters that you can have like greater than, less than, equal to, you can have and things if you have multiple filters. So we have like a pretty complicated like prompt that does all that. That might be one of my favorite things in LangChain, period. Like I think that's, yeah, I think that's really cool. How do you think about speed of development versus support of existing things? So we mentioned retrieval, like you got, or, you know, text splitting, you got like different options for all of them. As you get building LangChain, how do you decide which ones are not going to keep supporting, you know, which ones are going to leave behind? I think right now, as you said, the space moves so quickly that like you don't even know who's using what. What's that like for you?…

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