Harrison Chase: recommendation

Transcript context

…Yeah. I think that's where I start with LangChain. It's one of those things that you probably don't, if you're building an application, you probably don't start with it. You probably start with like a zero-shot prompt. But I think that's a really powerful one that's probably just talked about less because you don't need it right off the bat. And for those of you who don't know, that basically selects from a bunch of examples the ones that are maybe most relevant to the input at hand. So you can do some nice kind of like in-context learning there. I think that's, we've had that for a while. I don't think enough people use that, basically. Output parsers also used to be kind of important, but then function calling. There's this interesting thing where like the space is just like progressing so rapidly that a lot of things that were really important have kind of diminished a bit, to be honest. Output parsers definitely used to be an understated and underappreciated part. And I think if you're working with non-OpenAI models, they still are, but a lot of people are working with OpenAI models. But even within there, there's different things you can do with kind of like the function calling ability. Sometimes you want to have the option of having the text or the application you're building, it could return either. Sometimes you know that it wants to return in a structured format, and so you just want to take that structured format. Other times you're extracting things that are maybe a key in that structured format, and so you want to like pluck that key. And so there's just like some like annoying kind of like parsing of that to do. Agents, memory, and retrieval, we haven't talked at all. Retrieval, there's like five different subcomponents. You could also probably talk about all of those in depth. You've got the document loaders, the text splitters, the embedding models, the vector stores. Embedding models and vector stores, we don't really have, or sorry, we don't build, we integrate with those. Text splitters, I think we have like 15 or so. Like I think there's an under kind of like appreciated amount of those. And then... Well, it's actually, honestly, it's overwhelming. Nobody knows what to choose.…

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