1 00:00:00,000 --> 00:00:08,260 Welcome to this video on Build LLM Agents with Tools. 2 00:00:08,260 --> 00:00:12,939 In this video, you will learn the first steps in manual tool calling. You'll begin by 3 00:00:12,939 --> 00:00:18,360 initializing the chat model for tool interactions. You'll then define and bind custom tools 4 00:00:18,360 --> 00:00:23,659 to the LLM and add more tools for expanded functionality. Finally, you will use mapping 5 00:00:23,659 --> 00:00:27,219 dictionaries for dynamic function calls. 6 00:00:27,219 --> 00:00:32,259 In this video, you will enable an LLM to manually call a tool. This is a key building block 7 00:00:32,259 --> 00:00:36,939 for transforming a basic model into a real, interactive agent capable of engaging with 8 00:00:36,939 --> 00:00:39,380 the real world. 9 00:00:39,380 --> 00:00:43,380 First begin by building an agent that can do simple arithmetic operations using tools 10 00:00:43,380 --> 00:00:48,619 for addition, subtraction, and multiplication. In this case, you'll take an LLM and give 11 00:00:48,619 --> 00:00:54,259 it access to a set of tools. For example, if the input query is, what is 3 plus 2? The 12 00:00:54,259 --> 00:00:58,740 LLM will first extract the parameters, which are 3 and 2, and determine which tool to 13 00:00:58,740 --> 00:01:04,059 call. Next, the parameters are fed into the selected tool, which then processes the input 14 00:01:04,059 --> 00:01:08,779 and generates the output. Finally, the result is returned to the LLM, which combines this 15 00:01:08,779 --> 00:01:14,819 information to produce the final response. The addition of 3 and 2 is 5. 16 00:01:14,819 --> 00:01:20,220 Let's initialize the chat model to set up the LLM for tool interactions. First, import 17 00:01:20,220 --> 00:01:26,779 initChatModel from LangChain's ChatModels module. This function handles all the setup 18 00:01:26,779 --> 00:01:33,339 needed to connect to a chat model. Next, load GPT 4.0 mini with model underscore provider 19 00:01:33,339 --> 00:01:39,099 equals openAI, creating an object that can send and receive messages. Store this object 20 00:01:39,099 --> 00:01:44,819 in a variable called LLM. From this point on, whenever you see LLM dot invoke, it means 21 00:01:44,819 --> 00:01:50,099 you're interacting with the same model instance. In the diagram, the LLM box represents this 22 00:01:50,099 --> 00:01:55,699 LLM object, the central hub for all tool calls and user queries. 23 00:01:55,699 --> 00:02:00,860 With the model now live, the next step is to define a custom tool that the LLM can call. 24 00:02:00,860 --> 00:02:06,339 First, enable the model to perform basic arithmetic by providing it with a simple addition tool. 25 00:02:06,339 --> 00:02:11,619 Second, import the atTool decorator. This tells LangChain that the function can be called 26 00:02:11,619 --> 00:02:15,740 by the model. Next, define a function named add that takes 27 00:02:15,740 --> 00:02:21,699 two integers a and b and returns their sum. The doc string add a and b is what the model 28 00:02:21,699 --> 00:02:27,619 uses to determine when to use this tool. At this point, the tool is defined and ready, 29 00:02:27,619 --> 00:02:30,500 but it hasn't been connected to the model yet. 30 00:02:30,500 --> 00:02:36,220 Next, connect the add function to the model so it recognizes it as a callable tool. To 31 00:02:36,220 --> 00:02:41,339 do this, first place the add function in a list of tools. Then, bind this list to the 32 00:02:41,339 --> 00:02:45,899 chat model, creating a new object called LLMWithTools. 33 00:02:45,899 --> 00:02:49,740 This wraps the original LLM in a way that makes it aware of every function in the tools 34 00:02:49,740 --> 00:02:54,860 list. From this point on, whenever the call is invoked, the model will recognize and use 35 00:02:54,860 --> 00:02:59,059 the add tool whenever it needs to compute a sum. 36 00:02:59,059 --> 00:03:04,979 Let's add more tools to the LLM. First, create a subtract tool for basic subtraction. Similarly, 37 00:03:04,979 --> 00:03:10,899 create a multiplication tool to handle multiplication. Later on, you will need to call a function 38 00:03:10,899 --> 00:03:15,699 dynamically using its name. To achieve this, first create a mapping dictionary that links 39 00:03:15,699 --> 00:03:19,660 tool names as strings to their corresponding tool functions. 40 00:03:19,660 --> 00:03:25,059 Next, define the input arguments as a dictionary, where the keys match the function's parameter 41 00:03:25,059 --> 00:03:30,699 names. For example, a equals 1 and b equals 2. Then, call the function by using the tool 42 00:03:30,699 --> 00:03:36,380 name as the key, for example add, to access the function from the mapping dictionary. 43 00:03:36,380 --> 00:03:41,100 Finally, use invoke(Input) to call the function which automatically matches the dictionary 44 00:03:41,100 --> 00:03:45,699 keys a and b to the function's parameters and executes the tool. 45 00:03:45,699 --> 00:03:51,500 Next, create a list of tools that includes the add, subtract, and multiply functions. 46 00:03:51,500 --> 00:03:56,100 You can then bind this list to the model, allowing the LLM to recognize and use these 47 00:03:56,100 --> 00:03:58,539 tools as needed. 48 00:03:58,539 --> 00:04:03,380 In this video, you learned how to transform LLMs into interactive agents by integrating 49 00:04:03,380 --> 00:04:07,500 custom tools, enabling more context-aware responses. 50 00:04:07,500 --> 00:04:13,860 Set up chat models to handle real-world inputs, allowing the LLM to process user queries accurately. 51 00:04:13,860 --> 00:04:17,579 Create and connect tools such as addition, subtraction, and multiplication functions 52 00:04:17,579 --> 00:04:20,079 for dynamic calculations. 53 00:04:20,079 --> 00:04:25,619 Use mapping dictionaries for flexible, name-based function execution, expanding the LLM's 54 00:04:25,619 --> 00:04:26,619 capabilities. 55 00:04:26,619 --> 00:04:32,260 Understand how LLMs identify the right tool and manage parameter inputs for precise outputs. 56 00:04:32,260 --> 00:04:37,619 You can manage chat history, preserving conversation context for more accurate, personalized responses.