1 00:00:04,000 --> 00:00:08,500 Welcome to this video on Build Interactive LLM Agents. 2 00:00:08,500 --> 00:00:13,000 In this video, you'll learn how to extract tool names and arguments for precise function 3 00:00:13,000 --> 00:00:14,000 calls. 4 00:00:14,000 --> 00:00:18,059 You'll learn how to build agent classes that manage the entire tool-calling process 5 00:00:18,059 --> 00:00:21,780 for smarter, context-aware responses. 6 00:00:21,780 --> 00:00:26,600 Now that you've built a foundation in setting up LLMs with basic tool integrations, it's 7 00:00:26,600 --> 00:00:28,920 time to take things a step further. 8 00:00:28,920 --> 00:00:32,360 Begin by setting up the actual question you want the model to answer. 9 00:00:32,360 --> 00:00:34,159 "What is 3 plus 2?" 10 00:00:34,159 --> 00:00:38,200 First, insert the user's question into the chat history so the model can see it as a 11 00:00:38,200 --> 00:00:40,659 part of the conversation context. 12 00:00:40,659 --> 00:00:44,520 Convert the plain text question into a HumanMessage, which is a symbol wrapper that tells 13 00:00:44,520 --> 00:00:47,659 LangChain, This text came from the user." 14 00:00:47,659 --> 00:00:52,259 It carries the user's input and marks it appropriately for the model during the conversation. 15 00:00:52,259 --> 00:00:55,840 To do this, import HumanMessage from LangChain core messages. 16 00:00:55,840 --> 00:00:58,419 Next, build the chat history. 17 00:00:58,419 --> 00:01:03,099 Take the query string, wrap it in a HumanMessage, and place it into a list called chat 18 00:01:03,099 --> 00:01:04,379 history. 19 00:01:04,379 --> 00:01:10,059 This list will hold all messages exchanged during the conversation, including user inputs, 20 00:01:10,059 --> 00:01:15,540 tool outputs, and model responses, ensuring that the model has complete context at every 21 00:01:15,540 --> 00:01:16,540 step. 22 00:01:16,540 --> 00:01:18,660 It's time to run the tool-enabled model. 23 00:01:18,660 --> 00:01:23,980 First, pass the complete chat history, which contains the user's question, into LLM with 24 00:01:23,980 --> 00:01:24,980 tools. 25 00:01:24,980 --> 00:01:30,260 At this point, the model reviews the conversation context, identifies the available tools, select 26 00:01:30,260 --> 00:01:34,580 the appropriate one, such as add function, and extract the parameters. 27 00:01:34,580 --> 00:01:36,180 Notice the model's output. 28 00:01:36,180 --> 00:01:41,379 Instead of a plain text response, an AIMessage is received containing a tool calls array. 29 00:01:41,379 --> 00:01:44,779 Focus only on the part of the message related to tool usage. 30 00:01:44,779 --> 00:01:48,940 This is the model's way of saying, "I want to call this function with these arguments." 31 00:01:48,940 --> 00:01:53,739 Next, extract the details of the tool call and manually execute the addition ourselves. 32 00:01:53,900 --> 00:01:58,819 Finally, append the AIMessage, including the tool call instruction, into chat_history 33 00:01:58,819 --> 00:02:04,660 so the next model invocation has access to both the user's original query and the 34 00:02:04,660 --> 00:02:06,800 model's previous output. 35 00:02:06,800 --> 00:02:09,339 Response1 is an AIMessage object. 36 00:02:09,339 --> 00:02:11,619 Let's take a quick look at its key details. 37 00:02:11,619 --> 00:02:13,740 Here's the tool call parameters. 38 00:02:13,740 --> 00:02:16,399 "Add" is the name of the tool the model wants to call. 39 00:02:16,399 --> 00:02:21,460 The function arguments are a JSON string that specifies the inputs to pass into the tool. 40 00:02:21,460 --> 00:02:24,779 For example, "A equals 3 and B equals 2." 41 00:02:24,779 --> 00:02:27,300 Here is a unique identifier for this tool call. 42 00:02:27,300 --> 00:02:30,500 It's used to link the response back to the request. 43 00:02:30,500 --> 00:02:32,860 This is if more than one tool is called. 44 00:02:32,860 --> 00:02:37,699 The type field specifies that this is a tool call, not a text or other output type. 45 00:02:37,699 --> 00:02:42,139 Now parse the tool parameters, execute the tool manually, and then send the result back 46 00:02:42,139 --> 00:02:44,660 to the LLM as part of the conversation. 47 00:02:44,660 --> 00:02:50,419 First, parse the add tool with arguments "A equals 3 and B equals 2" and feed them into 48 00:02:50,419 --> 00:02:51,899 the function. 49 00:02:51,899 --> 00:02:55,360 Then extract the model's tool-call instructions into a variable. 50 00:02:55,360 --> 00:03:00,539 You will get a list showing exactly which tools the LLM selected and with what arguments, 51 00:03:00,539 --> 00:03:03,380 ready to be inspected and manually invoked. 52 00:03:03,380 --> 00:03:07,619 Next you will extract the tool's name from the first entry by running this command. 53 00:03:07,619 --> 00:03:12,539 You will then pull out "AddTool," telling exactly which function to call next. 54 00:03:12,539 --> 00:03:18,300 Extract the parameters to pass the function using the args key, for example "A3 and B2." 55 00:03:18,300 --> 00:03:20,660 Finally extract the tool call "ID." 56 00:03:20,660 --> 00:03:25,259 This ID links the tool's result back to the model's original request, so the LLM 57 00:03:25,259 --> 00:03:28,380 knows which tool call the response belongs to. 58 00:03:28,380 --> 00:03:32,339 This is especially important when multiple tools are called at once. 59 00:03:32,339 --> 00:03:35,619 Use the tool map created earlier to call the function you need. 60 00:03:35,619 --> 00:03:39,820 The key is the tool name determined by the LLM stored in tool_1_name. 61 00:03:39,820 --> 00:03:45,380 In addition, use the tool parameters provided by the LLM stored in tool_1_args. 62 00:03:45,380 --> 00:03:50,139 The tool map returns the output generated via the tool, in this case "2 plus 3 equals 63 00:03:50,139 --> 00:03:51,139 5." 64 00:03:51,139 --> 00:03:55,699 Finally, wrap the tool's output in a Tool Message by creating a Tool Message object. 65 00:03:55,699 --> 00:04:00,699 Finally, update the chat history by appending the new tool message to the chat history list. 66 00:04:00,699 --> 00:04:05,860 The history now contains the original HumanMessage, the AIMessage from the model, and 67 00:04:05,860 --> 00:04:08,380 the new ToolMessage with the tool's result. 68 00:04:08,380 --> 00:04:14,059 Begin by passing the updated chat history to the LLM with tools using the invoke method. 69 00:04:14,059 --> 00:04:19,239 The LLM will generate a final response using the tool's output, but format the answer 70 00:04:19,239 --> 00:04:21,619 naturally as a part of the conversation. 71 00:04:21,619 --> 00:04:25,540 You can create an agent class that encapsulates everything you did above. 72 00:04:25,540 --> 00:04:32,220 This class, ToolCallingAgent, binds tools to the LLM, manages chat history, extracts 73 00:04:32,220 --> 00:04:35,619 tool names and arguments, and handles tool calls. 74 00:04:35,619 --> 00:04:40,720 This structure captures the entire agentic process, from receiving a user query to returning 75 00:04:40,720 --> 00:04:42,459 the final response. 76 00:04:42,459 --> 00:04:47,459 Update the ToolCallingAgent object to handle the full tool-calling process automatically. 77 00:04:47,459 --> 00:04:51,839 The LLM parses the input text and determines the appropriate tool to call. 78 00:04:51,839 --> 00:04:54,380 The user input doesn't need to be perfect. 79 00:04:54,380 --> 00:04:58,000 Even imprecise queries such as "1 minus 2" can be understood. 80 00:04:58,000 --> 00:05:02,059 The agent identifies and executes the correct tool based on intent. 81 00:05:02,059 --> 00:05:07,220 In this video, you learned to structure user interactions for real-time, context-aware 82 00:05:07,220 --> 00:05:11,100 conversations by adding queries to the chat history. 83 00:05:11,260 --> 00:05:16,260 Extract tool names and arguments to precisely match user intent with the right tool calls. 84 00:05:16,260 --> 00:05:21,320 Parse complex tool instructions, including handling multiple tool calls and linking responses 85 00:05:21,320 --> 00:05:23,459 for seamless execution. 86 00:05:23,459 --> 00:05:28,100 Build and refine agent classes to automate the entire tool calling process, enabling 87 00:05:28,100 --> 00:05:32,519 LLMs to function as responsive, multi-step agents. 88 00:05:32,519 --> 00:05:36,940 Understand how these components work together to transform LLMs from passive responders 89 00:05:36,940 --> 00:05:41,179 into intelligent agents capable of making decisions and interacting with real-world data.