by now, a lot of people are used to the nature of large language models as generators of text. You know, they take a text prompt and they generate a response. So, you give them a question, for example, and they give you an answer. Or you ask them to write you an article about something, and they produce that article. You've come across maybe the concept of RAG, which is this ability of a large language model to go and get information from an external data source. For example, like searching the web or searching a specific dataset or company database. For example, tool use just takes this to the next level. And instead of just saying that we can only retrieve information from this external source, why not have it do other kinds of things? Why not use it as a tool? And here we'll look at an example. So, there are multiple tools, so web search is only one of a lot of things, but we'll look at an example here of a data analyst agent. And so, this is this fascinating idea about an LLM-backed agent. So now it's useful.

to not think of this as just communicating with the language model itself; rather, uh, you're talking with a software system that is better called an agent that is powered by a language model and that is able to use these external tools like web search, like a Python interpreter, like a calculator, like other sorts of APIs that are possible out there. And we will look at, if you go to Lang chain, there are a large number of APIs or tools that you can use. So, let's look at this example here of this data analyst agent, and we'll look at the code right after, but conceptually, it's this piece of software or LLM-backed agent that takes a user request, and then it doesn't directly just go and answer the request because these agents are built to solve more complex tasks. They start with outputting a plan and then start executing steps of that plan and reflecting at each time, and then finally, they go out and produce the result. So, let's look

at the example here of this agent, and so we'll go through the code in a bit, but here we can give this command to this agent and say, Create a plot of the number of full-time employees at the three tech companies with the highest market cap in the United States in 2024. Right, so this requires a number of steps: searching and then plotting. But how can you plot when you're a text generation model? Well, this text generation model is able to write code. That is one capability that Command R Plus brings to the table. But then it also can, if we give it this tool that is a Python interpreter, it's able to create graphing code and then give you that plot. And this is the final output of the model, so the data sources where it found the information, and then this is the plot that it sort of plotted. So this is just the input and the output, but we'll go through the successive steps of how that is generated. And so here you can see the

final plot this notebook uh starts out with uh installing Lang chain um and then coh here um together and then we go through the setup. So you need to put in your API key here. Uh if you don't have one, you can register and sort of sign up and copy it and paste it here. Now we can define our tools. We said we had these two tools, so let's start with the first one. So in LLY chain, there's this web search tool uh powered by Tav this API. So you need to go on here and get the API key and plug it here for for for tavali um. And then this is the name of the tool and then a description of the tool. And this is our second tool, the Python interpreter um. Also, it's imported from Lang chain experimental utilities uh. It's called the name of it is uh. We'll name it Python reppel and this is a description of what it does. It executes Python code and and returns the result. So now we've defined our two

tools here we start to um Define and declare our agent and sort of how it works so let's say we run this and so this is a react agent. This is the method of uh. This is one method of agents that are backed by llms. This is not to be confused with uh the JavaScript uh framework. Um after we've declared our agent and and set it up. We can now give it the request that we have and here we can start to sort of to see it um thinking and outputting the the results and so first it says. I will search for the three tech companies with the highest market cap in the US uh in 2024. Then I will search for the number of full-time employees and plot the data using python. So this is the plan of the model that it did in the first SP. So after it generated this description of its plan. Now it's starting to execute the plan um and saying okay now let me invoke this tool named internet search uh and give it this query top three tech companies.

highest market cap, and then this is one of the results that it's got, and then the model is reflecting on the result that it's got. So, this is the, let's say, the web page, and this is some of its content, um, and it says, I found that the three tech companies with the highest market cap are Microsoft, Apple, and Nvidia. Now, I will search for the number of full-time employees. So, this is what's called multihop, uh, tool use or multihop, uh, rag, where it's not necessarily, uh, all done in one step of searching or one step of, uh, tool use. So, the model sort of realizes that it needs more information, and so it will call the web, uh, Search tool again, and now it says, Okay, internet search Microsoft full-time employees, and then another search of Apple, and then another search of Nvidia. And so, now the model is, or the agent, uh, is doing four successive searches to find the information that you sort of, um, ask for. And then here, after it retrieves all of that information, it says, Okay, I have the

information now. This is what it's like. Now I will write and execute python code to plot this data. So now it throws. It says, Okay, this is code I'm going to send to the python interpreter code. And then this is the code that it wants to use to sort of create the plot. And this is the final output that includes that data. And this is the plot that was generated by Matplotlib. So you can try different kinds of tools here. You can go to LangChain, and there are tons and tons of different tools that you can set your model up with and have it sort of experiment with and build more and more complex systems this way. Notice here that if you, for example, ask the model a simple question like, How are you? It does not have to search for it. It says, Okay, I'll just directly answer. I don't need to search for this. And so this is the answer. I'm doing well, thank you for asking. So here you get

to see that now the llm is really able to reason on. Okay, how do I tackle different kinds of problems? And this is an absolutely fascinating way of thinking about the kinds of new applications that you can build with language models. Um, I hope you find this useful. Um, try it out. Let us know what you think.

Building LLM Agents with Tool Use

Building LLM Agents with Tool Use