WEBVTT 00:00.000 --> 00:05.000 Good night, talk, William. 00:05.000 --> 00:07.000 Great. 00:07.000 --> 00:09.000 Hey, everyone. 00:09.000 --> 00:13.000 Great. 00:13.000 --> 00:15.000 Thanks for joining. 00:15.000 --> 00:16.000 Good to see everyone. 00:16.000 --> 00:19.000 I'll be too much yelling at the bar last night. 00:19.000 --> 00:23.000 It's my voice is a little weak, but we'll try to give it a go. 00:24.000 --> 00:30.000 So this talk is building a gented graphical APIs with LLM tool use and knowledge graphs. 00:30.000 --> 00:35.000 You can find the slides, either that QR code or that short link there. 00:35.000 --> 00:36.000 My name's Will. 00:36.000 --> 00:38.000 I work for a company called HyperMode. 00:38.000 --> 00:41.000 We'll talk a little bit about what that is in a second. 00:41.000 --> 00:47.000 But before we dive in, there's a lot of terms in the title of this talk, right? 00:47.000 --> 00:49.000 It ends up being a little buzz wordy. 00:49.000 --> 00:52.000 I didn't really intend that when I wrote it. 00:52.000 --> 00:58.000 But let's maybe level set a little bit and talk some about the pieces here. 00:58.000 --> 01:00.000 So the first one is GraphQL. 01:00.000 --> 01:03.000 How many folks are using GraphQL or have used GraphQL? 01:03.000 --> 01:04.000 Okay. 01:04.000 --> 01:06.000 Maybe that's like half maybe. 01:06.000 --> 01:09.000 So GraphQL is an API query language. 01:09.000 --> 01:12.000 And so we have this strict type system. 01:12.000 --> 01:15.000 We define the types in our API, how they're connected. 01:15.000 --> 01:17.000 This is the graph piece. 01:17.000 --> 01:21.000 And then at query time, the client describes exactly the data they want to bring back. 01:21.000 --> 01:24.000 How to traverse through their application data. 01:24.000 --> 01:29.000 So GraphQL makes this observation that your application data is a graph. 01:29.000 --> 01:39.000 And you get back response that exactly matches the selection set, the pieces of the data graph that you said you wanted to be returned. 01:39.000 --> 01:42.000 I worked on open source GraphQL tooling. 01:42.000 --> 01:49.000 When I worked at a company called Neo4j, when GraphQL first was open source in like 2016. 01:49.000 --> 01:55.000 And that was like a database integration where we were trying to generate a schema from a database. 01:55.000 --> 01:58.000 Generate database queries from a GraphQL request. 01:58.000 --> 02:04.000 And as part of that work, I wrote this book published on by Manning, full-site GraphQL applications. 02:04.000 --> 02:10.000 This talk today is a little different take on how we're using GraphQL at Hyper mode and some of our tooling. 02:10.000 --> 02:13.000 Anyway, so that's the GraphQL piece. 02:13.000 --> 02:16.000 The next piece of this is Knowledge Graphs. 02:16.000 --> 02:22.000 Knowledge Graphs, I think, are having this like, resurgent interest now in the context of AI. 02:22.000 --> 02:33.000 It turns out that if you use a Knowledge Graph to keep track of your data and pass that context to these LOM AI models, 02:33.000 --> 02:36.000 then really improve the results. 02:36.000 --> 02:38.000 So we'll talk a bit about that. 02:38.000 --> 02:39.000 But what is a Knowledge Graph? 02:39.000 --> 02:43.000 There's lots of different definitions of Knowledge Graph out there. 02:43.000 --> 02:46.000 To me, a Knowledge Graph is just an instance of a property graph. 02:46.000 --> 02:54.000 So a property graph is this data model where we have nodes, nodes are the entities, the things. 02:54.000 --> 02:58.000 They have a label like the type, a way to group them. 02:58.000 --> 03:01.000 That's why we sometimes call this the labeled property graph model. 03:01.000 --> 03:07.000 And then we store arbitrary key value pair properties on nodes and relationships. 03:07.000 --> 03:11.000 So here's a very small Knowledge Graph. 03:11.000 --> 03:17.000 So I'm up there somewhere a person named Will employed by Hyper mode. 03:17.000 --> 03:20.000 I'm giving a talk that has a topic, LOMs. 03:20.000 --> 03:23.000 There's another person up there, David Allen. 03:23.000 --> 03:24.000 He have a talk yesterday. 03:24.000 --> 03:26.000 It also had the topic, LOMs. 03:26.000 --> 03:28.000 We both used to work for Neo4j. 03:28.000 --> 03:36.000 So I think one of the important pieces of Knowledge Graph is that there's some canonical representation of the thing. 03:36.000 --> 03:43.000 So for example, we have the topic LOM that's connected to a couple of talks at this conference. 03:43.000 --> 03:48.000 And when we have another talk that's about LOMs, we don't create a duplicate LOM node. 03:48.000 --> 03:51.000 We connect relationships to that one node. 03:51.000 --> 03:54.000 We have a canonical representation of the topic LOM. 03:54.000 --> 03:59.000 I think this is kind of the most important piece of this concept of a Knowledge Graph. 03:59.000 --> 04:05.000 When Google released their Knowledge Graphs, and I want to say like 2012 is part of the Google API. 04:05.000 --> 04:08.000 They talked about this concept of things, not strings. 04:08.000 --> 04:15.000 And I think that's a really good way to reason about Knowledge Graphs. 04:15.000 --> 04:18.000 Okay, so the next buzzword in my title is LOMs. 04:18.000 --> 04:21.000 How many folks are using LOMs to some degree? 04:21.000 --> 04:22.000 Yeah, a lot of folks. 04:22.000 --> 04:28.000 So these are models like Cloud or some of the OpenAI models, 04:28.000 --> 04:33.000 or the Open Source models like DeepSeek, which came out fairly recently, 04:33.000 --> 04:37.000 or LOMA, which is made available by meta. 04:37.000 --> 04:43.000 And the basic way to interact with these, we give the model some prompts and LOMs predicts text. 04:43.000 --> 04:47.000 So they're just predicting, given your prompt and everything I know, 04:47.000 --> 04:50.000 like what are the most likely next strings of text? 04:50.000 --> 04:53.000 And that's what we get back. 04:53.000 --> 04:58.000 But we can also use models in more sophisticated ways. 04:58.000 --> 05:04.000 And there's this concept of agentic workflows for working with LOMs. 05:04.000 --> 05:07.000 Again, there's lots of different definitions of like, what is an agent? 05:07.000 --> 05:14.000 But fundamentally, I think an agent or an agentic workflow is just when your LOM is making a decision 05:14.000 --> 05:17.000 about the control flow of your application. 05:17.000 --> 05:22.000 This is typically done with what's called tool use or function calling, 05:22.000 --> 05:29.000 where you're defining tools or functions that are available for the LOM to decide to call. 05:29.000 --> 05:35.000 And you describe the type of data that you can get back from calling that function or that tools. 05:35.000 --> 05:40.000 And then the data is passed back along with your original prompt, 05:40.000 --> 05:45.000 back to the LOM again, and you say, hey, this prompts with the new data, 05:45.000 --> 05:46.000 are we done yet? 05:46.000 --> 05:48.000 Or do you want to call another tool? 05:48.000 --> 05:51.000 And you sort of iterate on this until you get to the output. 05:51.000 --> 05:57.000 Sort of like the basic building block I think for this concept of LOM agents. 05:57.000 --> 06:00.000 So at HyperMode, this is a company I work for now. 06:00.000 --> 06:07.000 We build open source tooling to enable developers to build what we call model native apps. 06:07.000 --> 06:09.000 We'll talk about what that is in a minute. 06:09.000 --> 06:14.000 Our main open source projects are Modus, which is a serverless API framework. 06:14.000 --> 06:16.000 This is mostly we're going to talk about today. 06:16.000 --> 06:20.000 We also build and maintain D graph, which is a graph database. 06:20.000 --> 06:22.000 Anyone using D graph? 06:22.000 --> 06:24.000 Of course, a couple of folks. 06:24.000 --> 06:28.000 Badger, which is distributed key value store in Restredo, 06:28.000 --> 06:29.000 which is a cache library. 06:29.000 --> 06:31.000 These are all written in Go by the way. 06:31.000 --> 06:35.000 Go is a big piece of what we build with. 06:35.000 --> 06:40.000 And then all these projects are open source, Apache 2, build, deploy whatever you want. 06:40.000 --> 06:46.000 But the HyperMode hosting platform like aims to be the best cloud hosting experience for these things. 06:46.000 --> 06:49.000 That's how we're supporting those projects. 06:49.000 --> 06:58.000 And so today we're going to talk about Modus, which is our serverless API framework that HyperMode builds and maintains. 06:59.000 --> 07:02.000 We released this. I want to say in October. 07:02.000 --> 07:07.000 So like fairly, fairly new in the world. 07:07.000 --> 07:11.000 So I mentioned this term model native app, like what is that? 07:11.000 --> 07:15.000 Well, we think back to the three tier architecture. 07:15.000 --> 07:17.000 This is how I learned to build apps. 07:17.000 --> 07:19.000 This infrastructure still exists today. 07:19.000 --> 07:24.000 We have a client that talks to an application server that talks to a database. 07:24.000 --> 07:29.000 If we want to scale this up, we need to deploy another instance of our application server. 07:29.000 --> 07:32.000 Put a load balancer in front and the sort of thing. 07:32.000 --> 07:35.000 And like all those pieces still exist today. 07:35.000 --> 07:37.000 And apps still get built exactly this way. 07:37.000 --> 07:42.000 But I think it's the abstractions that developers think about that have changed. 07:42.000 --> 07:45.000 So instead of thinking about application servers, 07:45.000 --> 07:51.000 typically we're just thinking about functions and deploying this sort of serverless environment. 07:51.000 --> 07:58.000 And AI models, and we now want to incorporate us like a first class citizen into the applications we're building. 07:58.000 --> 08:05.000 And we're using different types of databases and APIs that sort of built from from scale for the beginning. 08:05.000 --> 08:12.000 So that's really what we mean by model native app, where the model is sort of a first class citizen. 08:12.000 --> 08:17.000 And we're using more modern frameworks in ways to build and deploy our apps. 08:17.000 --> 08:21.000 Developer experience is really important in this world. 08:21.000 --> 08:25.000 So let's take a look at the developer experience for modus. 08:25.000 --> 08:28.000 So we said this is a serverless framework. 08:28.000 --> 08:32.000 The abstraction that we think of for business logic is a function. 08:32.000 --> 08:34.000 So we write a function. 08:34.000 --> 08:42.000 Now modus uses web assembly to target as many languages we can for development currently. 08:42.000 --> 08:44.000 We support assembly script. 08:44.000 --> 08:46.000 Anyone using assembly script? 08:47.000 --> 08:53.000 A simply script is a type script like language for web assembly. 08:53.000 --> 08:55.000 We also support go. 08:55.000 --> 08:57.000 How many go folks are there? 08:57.000 --> 08:58.000 Okay, cool. 08:58.000 --> 08:59.000 A lot of folks. 08:59.000 --> 09:01.000 And we're adding more language as soon as the Python's coming soon. 09:01.000 --> 09:04.000 Other languages coming soon want to add as many as we can. 09:04.000 --> 09:06.000 So you write your function. 09:06.000 --> 09:11.000 And in the modus SDK, you have connections to data and APIs, 09:12.000 --> 09:14.000 abstractions for working with models. 09:14.000 --> 09:17.000 Your functions compiled to web assembly. 09:17.000 --> 09:24.000 We inspect the types that you've defined and the signature of your functions to generate a GraphQL API. 09:24.000 --> 09:27.000 And then when you query that in point from the client. 09:27.000 --> 09:30.000 Modus because we're using web assembly. 09:30.000 --> 09:38.000 Sort of using this execution plan that maps resolvers to functions that we've defined. 09:39.000 --> 09:44.000 Has this sandboxed invocation specific to that invocation of your function. 09:44.000 --> 09:50.000 So web assembly gives us secure sandbox memory space for each invocation, 09:50.000 --> 09:52.000 which is quite interesting. 09:52.000 --> 09:55.000 If we look at the components of modus for the development, 09:55.000 --> 10:06.000 a lot of this is driven by a CLI that exposes the way for skeleton and new app for working with the SDK 10:06.000 --> 10:11.000 and for compiling and running locally in web assembly, 10:11.000 --> 10:14.000 then for production. 10:14.000 --> 10:17.000 We're typically just looking at the run time. 10:17.000 --> 10:22.000 And again, we can deploy this to hyper mode, 10:22.000 --> 10:24.000 but the zombine source. 10:24.000 --> 10:30.000 We can run this anywhere that we can run web assembly modules. 10:30.000 --> 10:34.000 If we dive into the modus run time architecture, 10:34.000 --> 10:40.000 some interesting pieces here are perhaps pointing out that there are host services, 10:40.000 --> 10:44.000 so there are functions that run outside of the web assembly sandbox. 10:44.000 --> 10:51.000 So these are our connections to models, our database connections, things like this. 10:51.000 --> 10:55.000 We mentioned this concept of the web assembly sandbox, 10:55.000 --> 11:02.000 which gives us these nice memory safe secure sandboxes for each invocation. 11:03.000 --> 11:08.000 Our run time is built on top of the was zero, go web assembly run time, 11:08.000 --> 11:10.000 if anyone's using that. 11:10.000 --> 11:17.000 In the GraphQL piece, we leverage a lot of wonder graphs, go GraphQL tooling 11:17.000 --> 11:21.000 to generate the schema again based on the types that you've defined 11:21.000 --> 11:26.000 and the signature of your function. 11:26.000 --> 11:31.000 Cool, so let's take a look at what using this looks like. 11:32.000 --> 11:36.000 Is that big enough? 11:36.000 --> 11:37.000 No. 11:37.000 --> 11:38.000 Cool. 11:38.000 --> 11:42.000 So to get started with modus, we installed the modus CLI from MPM. 11:42.000 --> 11:46.000 I've already done this, so that shouldn't really do anything. 11:46.000 --> 11:50.000 And then we're going to say modus new. 11:50.000 --> 11:53.000 This is going to create a new modus project. 11:53.000 --> 11:56.000 Like I said, currently support go and assembly script. 11:56.000 --> 12:01.000 It's very similar to type script. 12:01.000 --> 12:10.000 So we will choose that, and this is going to skeleton out a new modus project for us. 12:10.000 --> 12:17.000 And pull down the assembly script SDK and just starting off with a very basic 12:17.000 --> 12:22.000 hello world project. 12:22.000 --> 12:29.000 There are a couple of interesting, well, first we're going to see what files we have. 12:29.000 --> 12:34.000 So a couple of interesting things to look at is modus.json. 12:34.000 --> 12:36.000 So this is our app manifest. 12:36.000 --> 12:41.000 This is where we configure the endpoints we want to generate. 12:41.000 --> 12:47.000 And then we would also add connections to any data APIs or databases here, 12:47.000 --> 12:52.000 or models that we would want to invoke in our app. 12:52.000 --> 12:59.000 And we can use local models with hosted by HyperMode in this development environment. 12:59.000 --> 13:03.000 So HyperMode can host models off of hugging phase. 13:03.000 --> 13:11.000 We also host models like DeepSeek and Lama, or we can connect to a third party model provider like 13:11.000 --> 13:17.000 Tropic, company I, who just declare that in this modus.json. 13:17.000 --> 13:23.000 And the other thing file is our index.ts, which just has a very basic function. 13:23.000 --> 13:27.000 Hello world, it takes an optional string and returns a string. 13:27.000 --> 13:30.000 So let's say modus dev. 13:30.000 --> 13:35.000 And is going to compile our project. 13:35.000 --> 13:39.000 Generator GraphQL schema and run that locally. 13:39.000 --> 13:44.000 So we'll switch back to, this is the modus API Explorer. 13:44.000 --> 13:49.000 If you think of this as kind of like graphical if you've used GraphQL before. 13:49.000 --> 13:52.000 And we can see here we have one query field. 13:52.000 --> 13:56.000 Say hello that maps to that function that we defined. 13:56.000 --> 13:58.000 If we run it, we get hello world. 13:58.000 --> 14:01.000 We can pass in our optional argument. 14:02.000 --> 14:06.000 So that's like a very basic hello world example. 14:06.000 --> 14:08.000 That's not super interesting though. 14:08.000 --> 14:12.000 Let's see what are some more interesting things we can do. 14:12.000 --> 14:17.000 Well, going back to this concept of the model native app. 14:17.000 --> 14:22.000 I think there are like some building blocks here that we really want to expose. 14:22.000 --> 14:26.000 In modus for building these model natives at native apps. 14:26.000 --> 14:31.000 And this is where models are first class citizen in the application that we're building. 14:31.000 --> 14:33.000 We're not just building a chatbot. 14:33.000 --> 14:40.000 Like we actually want to leverage our LM to make decisions about control flow in our application. 14:40.000 --> 14:44.000 And so there's kind of three like building blocks or three steps here. 14:44.000 --> 14:47.000 The first is tuning data and LMs together. 14:47.000 --> 14:55.000 So this is where like our internal enterprise data is fetched and passed in the prompt in the context of our LM. 14:55.000 --> 15:00.000 To make that output more relevant. 15:00.000 --> 15:10.000 This idea of function calling or tool use, which we talked about at the beginning, where we're defining functions for the LLM to sort of choose the control flow. 15:10.000 --> 15:15.000 Where to fetch data from what APIs to call to resolve our prompts. 15:15.000 --> 15:19.000 And this idea of knowledge graph rag. 15:19.000 --> 15:22.000 Rag is retrieval augmented generation. 15:22.000 --> 15:28.000 Where we're fetching some data and passing that in the context to our LLM. 15:28.000 --> 15:31.000 That's very similar to that first chaining data and LMs together. 15:31.000 --> 15:34.000 That's basically what they call naive rag. 15:34.000 --> 15:39.000 With knowledge graph rag, we can leverage the power of the knowledge graph, 15:39.000 --> 15:46.000 so the context of relationships in our data to again provide more relevant context for our LLM. 15:46.000 --> 15:52.000 So let's look at this first example, chaining data and LMs together. 15:52.000 --> 15:56.000 So we'll take a look at a basic demo app that I've built. 15:56.000 --> 16:04.000 Let's assume that we're running an online blogging website where you type in, 16:04.000 --> 16:09.000 like the text of your blog and then you hit a button and publish it on the internet. 16:09.000 --> 16:11.000 And something that might be useful. 16:11.000 --> 16:15.000 Some AI back feature that would be really nice to add to that would be, 16:15.000 --> 16:21.000 one, could we generate the SEO meta tags automatically based on the content? 16:21.000 --> 16:31.000 This is non-trivial because there's a certain way you want to construct the meta description tag to optimize for SEO and these sorts of things. 16:31.000 --> 16:37.000 And then also, could we generate suggested titles based not only on the content of the blog, 16:37.000 --> 16:40.000 but also in the style of the blog author. 16:40.000 --> 16:47.000 So that's going to be based on data that we have in the database that backs this online blog application. 16:47.000 --> 16:52.000 And so this is an approach called naive rag or retrieval augmented generation, 16:52.000 --> 16:57.000 where we're fetching some data from our internal database that's relevant for this specific query. 16:57.000 --> 17:05.000 The specific prompt that we're going to send to the LLM and passing that along to improve the results. 17:06.000 --> 17:09.000 So let's stop this guy. 17:11.000 --> 17:13.000 I think this is the right one. 17:13.000 --> 17:17.000 So this is the code called this moduspress. 17:17.000 --> 17:19.000 And there's a couple of functions here. 17:19.000 --> 17:22.000 One is generate SEO. 17:22.000 --> 17:25.000 So you can see the prompt we're passing to the LLM, 17:25.000 --> 17:30.000 your SEO expert creating each file meta tag for this content. 17:30.000 --> 17:34.000 Here's an example of how we want to return it. 17:34.000 --> 17:36.000 And then we paste in here the post content. 17:36.000 --> 17:40.000 And this content comes from a postgres database. 17:40.000 --> 17:49.000 Somewhere down here we have a query just a fetch author from a postgres database that we have running locally. 17:49.000 --> 17:58.000 That has a very small table here that just has one author and one bio. 17:58.000 --> 18:03.000 So let's run this modus dev. 18:03.000 --> 18:06.000 This is going to compile that. 18:06.000 --> 18:16.000 Oh, and we can also take a look at the modus JSON for this project. 18:16.000 --> 18:23.000 And we can see here that the model that we're using is meta's LM model hosted on hyper modes. 18:23.000 --> 18:27.000 This is an open source model provided by hugging face. 18:27.000 --> 18:34.000 I then also connecting to just a local postgres instance. 18:34.000 --> 18:39.000 Okay, so back to our API explorer. 18:39.000 --> 18:42.000 So now we have a couple of new endpoints. 18:42.000 --> 18:48.000 We have generate SEO and post content. 18:48.000 --> 18:50.000 So let's start with SEO. 18:50.000 --> 18:52.000 So I wrote a blog post earlier. 18:52.000 --> 18:59.000 Actually, Claude wrote a blog post for me about eating chocolate in Belgium. 18:59.000 --> 19:02.000 So we'll copy that. 19:02.000 --> 19:07.000 And so again imagine, this is the back end piece, the API piece. 19:07.000 --> 19:12.000 But imagine the front end of our app is here's some UI for writing a blog post. 19:12.000 --> 19:17.000 So writing this blog post about eating chocolate in Belgium. 19:17.000 --> 19:23.000 We want to automatically generate these HTML meta tags that are optimized for SEO. 19:23.000 --> 19:28.000 And again, there's a certain way that you write these where it's like verb first or something. 19:28.000 --> 19:32.000 I don't know how SEO works, but here's an example. 19:32.000 --> 19:35.000 Discover the art of Belgian chocolate making blah, blah, blah. 19:35.000 --> 19:39.000 And what's interesting is these LMs are non-deterministic. 19:39.000 --> 19:44.000 So every time we run this, we get a slightly different response. 19:44.000 --> 19:45.000 Which is kind of fun. 19:45.000 --> 19:48.000 Okay, so we're not like, that's not a rag example. 19:48.000 --> 19:51.000 We're not like passing any relevant information there. 19:51.000 --> 19:54.000 So let's look at the example for generating a title. 19:54.000 --> 19:58.000 And remember this one, we want to be in the style of the author. 19:58.000 --> 20:01.000 So that's me. 20:01.000 --> 20:03.000 Travel category. 20:03.000 --> 20:04.000 Let's say travel. 20:04.000 --> 20:07.000 And then here's the blog post again. 20:07.000 --> 20:10.000 And so what this is doing now is it's going to that post-gres database. 20:10.000 --> 20:15.000 Looking up the author, reading their bio and the prompt says, 20:15.000 --> 20:17.000 You know, in the style of the author based on this bio, 20:17.000 --> 20:22.000 If we had, you know, my full post history, we'd pass those along as well. 20:22.000 --> 20:29.000 And here we're getting selected suggested titles based on our blog content. 20:29.000 --> 20:34.000 And again, non-deterministic, so each time we get a slightly different response. 20:34.000 --> 20:37.000 Which is kind of fun. 20:37.000 --> 20:38.000 Cool. 20:38.000 --> 20:42.000 So that's sort of the basic naive rag approach. 20:42.000 --> 20:47.000 The next one here is function calling or tool use. 20:47.000 --> 20:51.000 And so this is how we want to implement like this basic, 20:51.000 --> 20:59.000 agentic workflow where the LLM is making some decision about the control flow of our app. 20:59.000 --> 21:02.000 And so we define some functions. 21:02.000 --> 21:04.000 You can look at the code for this. 21:04.000 --> 21:07.000 Let's stop this one. 21:07.000 --> 21:13.000 And here we go, zoom in a bit. 21:13.000 --> 21:19.000 So in this example, we have a very basic warehouse that has some products. 21:19.000 --> 21:26.000 So we have things like shirts, shoes, trousers, hats. 21:26.000 --> 21:28.000 And we have two functions. 21:28.000 --> 21:31.000 We have get product info. 21:31.000 --> 21:34.000 So you can pass a specific product that will say, 21:34.000 --> 21:37.000 here's how many shirts I have, here much they cost. 21:37.000 --> 21:42.000 And then we have get product types that will say we have shirts, trousers, shoes, hats. 21:42.000 --> 21:50.000 And we basically define these as tools that are available to the LLM. 21:50.000 --> 21:54.000 And in our LLM called based on the query that we have. 21:54.000 --> 21:59.000 And maybe just an interest of time, we'll think I have this one in the slides. 21:59.000 --> 22:03.000 Yeah, so we can pass a natural language query. 22:03.000 --> 22:06.000 This one says what's available to buy. 22:06.000 --> 22:09.000 And you can see here the response is the available products to buy. 22:09.000 --> 22:11.000 Our shoe hat trousers are in shirt. 22:11.000 --> 22:16.000 And we're also logging like the functions that the LLM chose to invoke. 22:16.000 --> 22:23.000 And in this case, it just invoked get product list to return what are the products that you have. 22:23.000 --> 22:27.000 But if we ask something we're complicated, this kind of runs off. 22:27.000 --> 22:32.000 How many complete outfits can I buy for $1,600, but I don't like hats. 22:32.000 --> 22:36.000 And so now you can see the different calls the LLM makes. 22:36.000 --> 22:39.000 Well, first it does get product info. 22:39.000 --> 22:41.000 shoe hat trousers shirts. 22:41.000 --> 22:44.000 And then now it's getting the price for each one of those. 22:44.000 --> 22:49.000 And it knows enough not to include a hat in my complete outfit. 22:49.000 --> 22:50.000 So I said I didn't like hats. 22:50.000 --> 22:53.000 So it knows a complete outfit is shoe trousers shirt. 22:53.000 --> 23:00.000 And I can buy two of those because they cost $800 all together. 23:00.000 --> 23:03.000 Very fancy clothing shop, I guess. 23:03.000 --> 23:06.000 Not somewhere I would be shopping. 23:06.000 --> 23:07.000 Okay, cool. 23:07.000 --> 23:11.000 So that's the basic idea of function calling and tool use. 23:11.000 --> 23:14.000 This next piece, knowledge graph rag. 23:14.000 --> 23:21.000 So this is where we want to leverage a knowledge graph to improve the retriever piece, 23:21.000 --> 23:26.000 to improve the data that we're fetching to pass to the context of the LLM. 23:26.000 --> 23:28.000 So in the first example, 23:28.000 --> 23:32.000 chaining data and LLM together, the naive rag approach, 23:32.000 --> 23:40.000 we're just doing a select star from author table where the name is my name. 23:40.000 --> 23:46.000 But we can do much more complicated, much more sophisticated retrievers 23:46.000 --> 23:48.000 to find relevant information. 23:48.000 --> 23:52.000 So for example, we could use embeddings. 23:52.000 --> 23:55.000 If we're dealing with lots of unstructured data, 23:55.000 --> 24:00.000 we can chunk up a bunch of text, calculate embeddings, do vector search. 24:00.000 --> 24:08.000 We can also traverse the knowledge graph to find more relevant pieces of data. 24:08.000 --> 24:11.000 And this is going to depend on what the domain is. 24:11.000 --> 24:16.000 But typically, if we're dealing with unstructured data and constructing a knowledge graph from that, 24:16.000 --> 24:20.000 we'll have our domain graph, which, in this case, 24:20.000 --> 24:22.000 we're going to build a movie recommendation app. 24:22.000 --> 24:26.000 So our domain graph, in this case, we have movies, 24:26.000 --> 24:30.000 good users that have rated a movie, so this is like a mashup of the movie lens, 24:30.000 --> 24:35.000 data sets, and IMDB movie lens was used for, 24:35.000 --> 24:41.000 I think, like, the Netflix prize for finding movie recommendations. 24:41.000 --> 24:45.000 But anyway, we have a domain graph, but then sometimes we also build, 24:45.000 --> 24:49.000 like, a lexical graph, where we've chunked up unstructured data, 24:49.000 --> 24:52.000 and that we're calculating embeddings based on that, 24:52.000 --> 24:56.000 and we can traverse from our lexical, into our domain graph, and sort of through there. 24:56.000 --> 24:58.000 There's this graphrag.com site. 24:58.000 --> 25:01.000 It has a lot of information on the different patterns, 25:01.000 --> 25:04.000 different retrievers that are used for graphrag. 25:04.000 --> 25:08.000 But if a couple minutes left, let's take a look at this example. 25:08.000 --> 25:13.000 So we're going to use embeddings, embeddings is basically just taking text, 25:13.000 --> 25:17.000 and converting that to a vector of numbers, 25:17.000 --> 25:20.000 and then we can do what's called vector search, 25:20.000 --> 25:27.000 vector similarity search, to find embedding values that are close together in vector space. 25:27.000 --> 25:31.000 So most databases support vector search now. 25:31.000 --> 25:35.000 Let's see if we have this one running, maybe. 25:35.000 --> 25:39.000 So we take a look at the code for this one. 25:39.000 --> 25:45.000 So we're connecting to an ear for J instance. 25:45.000 --> 25:50.000 We've already calculated these embeddings using the minilem model, 25:50.000 --> 25:55.000 which is, again, hosted by head of the code. 25:55.000 --> 25:59.000 So we're connecting to an ear for J instance. 25:59.000 --> 26:04.000 We've already calculated these embeddings using the minilem model, 26:04.000 --> 26:09.000 which is, again, hosted by head of the code, provided by hugging face. 26:09.000 --> 26:15.000 And then at query time, we're basically just doing vector search to find similar movies. 26:15.000 --> 26:18.000 So let's take a look at the app here. 26:18.000 --> 26:21.000 Is this running? 26:21.000 --> 26:25.000 So we search for a movie by title, size search for Cosmos, 26:25.000 --> 26:28.000 and we're looking that up in the database, 26:28.000 --> 26:33.000 and then doing vector search to find similar movies based on the embeddings. 26:33.000 --> 26:36.000 And the embedding of the, in this case, the plot. 26:36.000 --> 26:42.000 So that's the thing that we embedded in the database. 26:42.000 --> 26:50.000 So the code for this, if you're interested, is available in this modus recipes. 26:50.000 --> 26:51.000 GitHub repo. 26:51.000 --> 26:53.000 This one is specific for the movies kit. 26:53.000 --> 26:57.000 All the examples I showed today are in this modus recipes repo. 26:57.000 --> 27:02.000 This just shows different ways to use modus for building cool apps like that. 27:02.000 --> 27:06.000 And that is all I have, getting grab the slides here, 27:06.000 --> 27:08.000 and I think we are out of time as well. 27:08.000 --> 27:11.000 So thanks everyone, and we'll see you around. 27:11.000 --> 27:15.000 Thank you.