Hi, my name is Wale Kadus. I'm the head of engineering here at AnyScale, and I'm really excited to talk to you today about the combination of LangChain and Ray. What we'll be talking about today is showing you why LangChain is such an

awesome library and why it's so popular within the large language model community. And then I'll show you a little bit about how Ray complements LangChain by allowing you to build indexes more quickly, serving LangChain on the web directly, and allowing you to put models and the serving in the same

place to keep latency nice and low. So let's imagine now that I'm I'm trying to build a search application, and this is part one of three parts that we're going to have. Part one is this part where I'm just going to build like

a little search engine. In part two, we'll look at enhancing that to be to use large language models to summarize those findings, but really just even doing these basic things with the search engine shows what's possible. So imagine that I'm trying to make docs.ray.io more

accessible. Well, the first thing that I can do is I can use the power of large language models to create a really nice semantic search engine that isn't just finding words but is actually finding meaning. And we're going to talk through how we do this uh in a moment,

but it's really exciting just how easy this is. I remember when you wanted to build a search engine, you used to need lots of machines and lots of organizing, and you still wouldn't get great results. It's just amazing to me how how LangChain makes this easy, so

the process that we're going to go through is essentially that we're going to start with um our docs.ray.io website, just scraped it, we're going to unload the files into memory and then each page we're going to divide into little chunks and it turns

out that this is going to generate about 48 000 sentences. So each page you can imagine breaking it up into chunks. It's not exactly sentences, they might be slightly longer, slightly shorter. Then we're going to use this sentence Transformer that Hugging Face makes

available to project every sentence into a 768 long Vector. This process of taking a sentence and converting it to a long Vector is called embedding, and so we're going to take those uh that row and then we're going to store them in an index, and so once you have the index

that allows us to then find the information. So I'm now going to go to our code and walk you through it. It shouldn't take very long. The first thing we're going to do is just look at how easy LangChain makes this, so we're going to go here and we're

going to specify where the index is going to go. This is all that I need to specify with LangChain to pull all the docs into memory, and then I define a very simple text splitter here that takes all of those pages and cuts them into little

pieces. And so sure enough we load these documents and we create the documents from them, making sure to keep our metadata. So all we're doing here is we're generating a chunks at the chunks using the metadata and then the final stage is for each of these individual

little chunks, we're going to embed them in the space, put them in FAISS, which is a good vector store, and then we're going to save it locally. Now, we used FAISS, but LangChain actually supports a huge variety of different systems. All right, so so this works really great.

The only issue is it takes a few minutes. Now, obviously, here docs.ray.io is not very large; it's about 130 megabytes, but having something this slow at operating scale could be an issue, and this is where the power of Ray and LangChain

together becomes really impressive. So, what I'm going to do here is I'm going to switch over to my Jupyter lab, I'm going to clear the screen, and we're going to start a process here, and while it's running, I'll come back and discuss

what what that process is doing. So, what we're going to do here is we're going to the first part of the pipeline is exactly the same, but then what we're going to do is split those that effort into, um, eight different shards. We're going to do

those embedding processes, which can be quite GPU intensive, using eight GPUs at once, and then finally, we're going to combine them into a single merged database that we're going to push to the FAISS index. And as you can see, it's running very nicely here; it's still loading the

documents. So, it's loaded the documents into memory, and now we've constructed those eight different shards, and now each one of those eight shards is processing the data at once. So, now you can see that you know there are 48,000 documents or so, each one of the shards is is processing

the data and you can see that they are they're loading the embeddings from disk, but it's doing it once on each GPU. And we can start to see that here if we go look at the cluster, so as you can see those GPUs are looking

really hot and they're now chunking away trying to do this data. This is a machine with eight GPUs and so you can see all of those are working really hard and now very quickly you can see it finishes the work and that process is now complete. Some of

the charts are finished. Some take slightly longer, but from taking five or six minutes it's now uh completely solved. So that's an example of how you can use Ray to speed things up. All right, let's go back. So now that we have our index, how can we serve it and

this is really again one of those instances where Ray can really help you to do something without it being much effort. So what we have here is that the webcast comes in and it's usually just a sentence that we'll see. We transform that one sentence

into a vector just like before using the power of embeddings and then we're going to try to find a set of four most relevant sentences or four sentences that help the user understand and return a response. Now let me just have a look at the source code for that, so what we

can do here is just show how easy this is. As you can see here all that we're doing is we add an annotation for serve deployment in the beginning. We load the database and we load our embeddings and then when the clutch comes in we simply do a self.search result here. This

gets the maximum marginal relevance and then just does some nice thing uh I think uh nice processing all we need to do to start the deployment and go through the code that we put together is just have a command that looks like this it names the file that we're going to

use and has the the vector store and it just names that deployment that we've created sure enough you know it's starting up now it might take a few minutes because it's packaging up everything to make it ready for serving and now the Moment of Truth we've

written a very small program here called query.py that simply takes a request at the command line and sends it to the web page that we've just loaded so let's see how that works I'm going to Simply ask the the engine does Ray support Reserve support

patching and sure enough when we hit that it hits a number of web pages that show you exactly where you need the information about racer support for matching thank you very much folks I hope you see now why I'm so excited about the combination of Lang chain and Ray

um if you want to find out more please come to anyscale.com and there's a lot of information there that you can find to do this as well as a blog post that has this information with links to actual code and look forward to talking to you

Open Source LLM Search Engine with LangChain on Ray

Welcome to the Anyscale YouTube channel! Join us as we explore Ray, an open-source framework designed to accelerate the development and deployment of machine learning applications. From beginner tutorials to advanced insights, our videos cater to all skill levels. Stay up-to-date with the latest advancements and hear from experts through interviews, panel discussions, and keynotes.