Introduction to LangGraph

Key Takeaways

Hi everyone. It's good to see you. I am Husefa and I'm going to be your instructor for this module. In this series, we are going to cover the foundations of LangGraph. We'll talk about what LangGraph is, what are the different components, how it is orchestrated, and how this framework allows you to create end-to-end LLM workflows. So, without further delay, let's get started. So, this entire course has been divided into three sections. The first section we are going to cover the basics of LangGraph. Here, we are going to understand what are the components of LangGraph, like some of the basics. We'll also understand or create some basic simple agent workflows to get you used to to this entire framework. And finally, we'll cover how you actually manage state and memory in order to improve your entire workflow. In the next section, we are going to cover the orchestrating or architecting of multi-agent workflows using LangGraph. And here, we are going to talk about the concept of branching and conditional flows, using which you can basically design how the workflow should flow based upon the output from the previous agents or the previous step that you have executed. Also, we're going to talk about the concept of tools and how can you integrate some external tools into your existing workflow. And finally, we are going to cover how you can build an autonomous research engine. This is basically our capstone project. It's going to try everything or tie up everything together that we have basically discussed in this entire module. So, you need to be used to the IDE of Jupiter because all the notebooks that we are going to practice are also created using Jupiter notebook itself. And finally, the basics of prompt engineering, tools, and vector databases are also required for this module because here we are going to only cover LangGraph, like the prerequisites of tools, prompts, vector DBs. It's expected that you have seen our previous modules as well that cover these topics in detail. Now, it's time to understand or take a bird's-eye view of what exactly LangGraph is. I mentioned this previously that LangGraph is a framework. What's the meaning of the word framework? Framework is nothing but a collection of all the classes, utilities, functions, whatever is required to help you, you know, develop these end-to-end large language model frameworks or applications in general. So, take a very simple example. Say you want to build a chatbot which can help users navigate the various queries that they might experiencing when navigating through your website. It could be something as simple as some queries related to their orders. It could be some information that they want like so they want your help in order to guide them towards which is the best product to purchase. The things like that. So, in this case, depending upon the user query, you need to redirect the user to the corresponding agent. So, if the user has asked a question around what exactly is the status of my order? So, in that case, you need to route your agent or the user to your, let's say, ordering agent. Second, if the user queries is about, let's say, new purchases. Right? So, in that case, you need to route this particular agent to another, let's say, sales agent or something like that who can basically talk to the user around the recommendations for sales, take in their requirements and things like that. So, in in this entire thing, there are different types of agents which have the capability to do their own kind of interaction. And you need to choose on you need to a framework that can, you know, automatically select and decide on its own where you need to route it. And to orchestrate this whole framework, you have LangGraph which has all the utilities which we'll talk about in the coming sessions. And use them to create this entire application from scratch. And that's basically the concept of LangGraph. It's a framework for building stateful LLM workflows as graphs. We'll talk about how what graphs are, what are the different things in this case in detail in a minute, but think of it as a simple like a structured flow of information. So, in this case, for example, the user query will come in. And then based upon what type of query it is, if it is basically let's say a query around orders, so it will route it to some other agent. If it is around sales or purchase, then it will route it to some other agent. So, this is basically kind of a graphical flow of information. Wherein based upon the type of query you are then deciding to, you know, initiate the interaction with the corresponding correct agents. And that's basically what this entire framework allows you to do. There are a lot of terms here, stateful and such. We'll talk about states and why states are required in a coming because that's when it is most obvious to use there. Then this framework of LangGraph internally uses another very popular Python package called LangChain, which is basically like a precursor to LangGraph. What LangChain allows you to do, LangChain allows you to basically create a chain of steps involving a large language model. Wherein you can basically define things like wrap a prompt, use that prompt to initiate the agent, take the agent and collect the response. So, this is basically like a chain of events which happens in order to get your responses. Now, LangGraph wraps it in its own framework, wherein it is using the utilities provided by LangChain to define an entire end-to-end workflow based upon graphical nodes and edges. Right? So, the precursor of LangGraph is LangChain, but LangGraph has much powerful way of orchestrating end-to-end workflows. And inside LangGraph, in order to create a graph, you have these nodes, which are these steps that you have. Like so, these are the nodes. And then you have edges, which basically essentially are the connections between nodes. So, based up from node one to node two, how do you move from one to two? Right? So, in this case, if your user query is talking about orders, then you initiate the ordering agent. If it is talking about purchases, then you you know, initiate the purchase agent. So, it's basically like edge which connects your root node to your child nodes or one parent node to its respective children. Right? So, here we are going to define these nodes and edges which connect different entities in that entire workflow. Now, this entire large language model or agentic workflow stack has different layers. And the question might come to your mind is that where exactly is LangGraph fitting into the stack? Now, obviously, you as a user are not going to train your own large language model. You might be using some third-party APIs such as OpenAI or Anthropic or Google, right? So, what is whatever you are using, you will have an OpenAI API or a large language model API from Anthropic, etc., which you will call internally. Right? That is one layer wherein you have an API layer which routes your calls to to any large language model and collects your Right? You send the request and the model responds back. So, you have this API that basically connects your application to the underlying language large language. You do not have to host it on your own machine. Then, you have tools which give ability to your agent to do multiple things simultaneously. So, tools are nothing but functions that allow your agent to perform tasks beyond its original capabilities. Let's say, for example, you have ChatGPT, which internally is basically supposed to interact with the agent. But now you want that same same ChatGPT to essentially allow users to book flights as well. So, you need to give it a tool which using which it can actually connect to a website from, let's say, uh Air India or any such website and any airlines' website, take the user information, feed that information to that website, and then book the flight on your behalf. So, without tools, an agent is limited to only performing tasks which are in its power. Right? So, whatever it has been trained to do. But if you want to do it some additional stuff like booking flights or, uh you know, booking hotels. Uh in this case, it could be as simple as, let's say, book me a vehicle or plan me a calendar. Right? So, it's anything any any sort of activity which is beyond the scope of this root, uh you know, basic bare-bone agent requires you to add tools which gives these agents abilities to perform tasks beyond their original capacity. And the last part is orchestration, which is basically where the LangGraph comes in. Because now, how do you let agent know when it needs to call which tool? Right? That is where the orchestration comes in, wherein you tie all of these pieces together. You have an API from large language model providers like OpenAI and Anthropic. Then you have websites which have tools available to you which you can call in order to give abilities to your agents. But then, how do you orchestrate everything? You use frameworks like LangGraph to help you orchestrate this entire sequence better. That's the whole idea of where LangGraph fits in in this entire stack. And there are multiple use cases. One of them I've already explained, right? Wherein you have a support assistant or a chatbot to help your users navigate better when they are actually coming to your websites. But then, you have something called a research co-pilot, wherein based upon the users' queries, you your agent will help the user, you know, navigate the entire ecosystem of research documents, right? And they will help the researcher to better do their research on whatever the initial, you know, citations are already available to look at. And then, you have like workflow agents which can design some complicated office-oriented or or any workflow that can, you know, help you automate your day-to-day tasks. It could be as simple as, let's say, whenever you get an email, based upon the contents of an email, it can update your calendar and add a to-do activity. Right? So, you can actually create a workflow that can help you do this. This is a very simple example I'm giving, but obviously you can create some more complicated workflows as well. Another very good example could be, let's say, if a user is requesting you to schedule a call, the agent will automatically navigate that, go to your calendar, and book a meeting, and basically send an invite back to the user. Right? So, all these workflows can be created very fast and you can actually use it to have multi-step stateful LLMs. I'll talk about states and everything later, but yeah, the beauty of LangGraph is it's a stateful workflow which allows you to introduce states which can persist information that is supposed to be used by downstream agents as well. We'll talk about that when the time comes. And that's all for this introductory session. I hope this was, you know, intriguing enough for you to basically follow along with us for the remaining of the videos as well. I'll see you in the next lecture. Thank you very much.