1 00:00:03,000 --> 00:00:08,680 Welcome to this video on Build a Custom Math Toolkit Agent with LangChain. 2 00:00:08,680 --> 00:00:13,279 In this video, you'll learn how to create flexible agents using LangChain and LangGraph. 3 00:00:13,279 --> 00:00:18,959 You'll build a custom math toolkit with tools for addition, subtraction, multiplication, 4 00:00:18,959 --> 00:00:19,959 and division. 5 00:00:19,959 --> 00:00:25,559 You'll guide the agent with a custom prompt and run complex queries using the invoke method. 6 00:00:25,719 --> 00:00:30,600 Finally, you'll learn how to integrate tools and build agents that both calculate and retrieve 7 00:00:30,600 --> 00:00:32,720 information. 8 00:00:32,720 --> 00:00:37,279 So far, you've used Initialize Agent to set up agents with predefined strategies such 9 00:00:37,279 --> 00:00:39,560 as React or Structured Chat. 10 00:00:39,560 --> 00:00:44,959 Now, let's look at another way to build agents with even more flexibility, LangGraph, 11 00:00:44,959 --> 00:00:49,279 which is quickly becoming the preferred approach over Initialize_Agent for building 12 00:00:49,279 --> 00:00:53,040 robust, multi-step agent workflows. 13 00:00:53,040 --> 00:00:58,040 Using "create react agent", you can work with just one agent type, but you can fully customize 14 00:00:58,040 --> 00:01:02,479 the prompt template, giving you more control over the agent's behavior and reasoning 15 00:01:02,479 --> 00:01:04,239 style. 16 00:01:04,239 --> 00:01:08,080 Before creating the ReAct agent, you'll define a structured tool that handles numeric 17 00:01:08,080 --> 00:01:13,080 extraction and summation from input text. 18 00:01:13,080 --> 00:01:17,199 Just like LangChain, LangGraph helps you build intelligent agents. 19 00:01:17,199 --> 00:01:22,000 To create a React-style agent in LangGraph, use the "create react agent" function from the 20 00:01:22,000 --> 00:01:25,040 langgraph.prebuilt module. 21 00:01:25,040 --> 00:01:30,879 Create the agent using the "create react agent" and pass in the LLM and a list of tools, for 22 00:01:30,879 --> 00:01:34,639 example, sum numbers with complex output. 23 00:01:34,639 --> 00:01:40,319 Optionally, include a custom prompt to guide the agent's behavior. 24 00:01:40,319 --> 00:01:44,519 To interact with the agent, use the invoke method and pass a dictionary with a messages 25 00:01:44,519 --> 00:01:45,720 key. 26 00:01:45,720 --> 00:01:51,400 This key contains a list of role message pairs, such as human, and add the numbers "-10", 27 00:01:51,800 --> 00:01:56,400 "-20", "-30", which simulates a chat interaction. 28 00:01:56,400 --> 00:02:01,919 The agent processes the input, reasons through the problem, and uses tools as needed. 29 00:02:01,919 --> 00:02:07,019 It then appends its final response to the message list, extract the final answer. 30 00:02:07,019 --> 00:02:11,720 The full response object also includes other useful information, such as tool call traces 31 00:02:11,720 --> 00:02:17,960 and intermediate steps, which can help with debugging or tracing the agent's reasoning. 32 00:02:17,960 --> 00:02:23,839 In real-world use, a single tool often isn't enough to handle everything a user might ask. 33 00:02:23,839 --> 00:02:26,199 Some tasks might require different tools. 34 00:02:26,199 --> 00:02:30,479 For example, if you're building an agent for banking transactions, you might need one 35 00:02:30,479 --> 00:02:34,839 tool for deposits, another for withdrawals, and a third for transfers. 36 00:02:34,839 --> 00:02:39,580 To demonstrate this idea, you'll build a simple math toolkit. 37 00:02:39,580 --> 00:02:44,080 To begin with, the basic addition tool, while functional, might not be sufficient for all 38 00:02:44,080 --> 00:02:46,839 mathematical tasks a user might ask. 39 00:02:46,839 --> 00:02:52,720 Alongside the addition tool, you'll add a multiplication tool, a division tool, and 40 00:02:52,720 --> 00:02:55,080 a subtraction tool. 41 00:02:55,080 --> 00:02:57,639 You will now create a ReAct agent. 42 00:02:57,639 --> 00:03:01,080 Start by adding all the tools to a list called tools. 43 00:03:01,080 --> 00:03:06,520 Then create the ReAct agent using the "create react agent" function, passing in the LLM and the tools 44 00:03:06,520 --> 00:03:08,339 list as parameters. 45 00:03:08,339 --> 00:03:13,220 Also include an optional prompt or system message that helps guide the agent's behavior. 46 00:03:13,220 --> 00:03:18,779 In this case, define the agent as a helpful mathematical assistant. 47 00:03:18,779 --> 00:03:21,380 Here's a basic idea of how the agent works. 48 00:03:21,380 --> 00:03:26,100 The agent receives an input query, such as what is 7x3. 49 00:03:26,100 --> 00:03:30,100 The LLM analyzes the query and decides which tools to call. 50 00:03:30,100 --> 00:03:32,899 In this case, call the multiplication tool. 51 00:03:32,899 --> 00:03:36,300 The input parameters are passed to the multiplication tool. 52 00:03:36,300 --> 00:03:40,820 The tool returns the result, and the LLM checks that the output makes sense. 53 00:03:40,820 --> 00:03:46,100 The LLM then formats and returns the final answer to the user. 54 00:03:46,100 --> 00:03:51,179 Call the agent using invoke with a user message, multiply 2, 3, and 4. 55 00:03:51,179 --> 00:03:54,580 The agent selects the right tool and processes the request. 56 00:03:54,580 --> 00:03:57,139 The answer is received from response messages. 57 00:03:57,139 --> 00:03:59,820 The full response is shown. 58 00:03:59,820 --> 00:04:04,860 LangChain provides a rich ecosystem of pre-built tools that solve common tasks out of the box. 59 00:04:04,860 --> 00:04:06,339 Here are some examples. 60 00:04:06,339 --> 00:04:10,460 WikipediaQueryRun searches Wikipedia for factual information. 61 00:04:10,619 --> 00:04:14,539 GoogleSearchRun performs web searches using Google's API. 62 00:04:14,539 --> 00:04:20,700 PythonREPLTool runs Python code safely, useful for calculations or logic. 63 00:04:20,700 --> 00:04:26,179 OpenWeatherMapQueryRun fetches real-time weather data from the OpenWeatherMap API. 64 00:04:26,179 --> 00:04:29,820 YouTubeSearchTool searches for videos on YouTube. 65 00:04:29,820 --> 00:04:33,940 Now you'll create a custom Wikipedia search tool using the Tool Decorator and the LangChain 66 00:04:33,940 --> 00:04:36,380 Community's Wikipedia API wrapper. 67 00:04:36,380 --> 00:04:39,660 First, define clear input and output types. 68 00:04:39,660 --> 00:04:43,739 Include a detailed docstring describing the tool's purpose, extended parameters, and 69 00:04:43,739 --> 00:04:45,019 return value. 70 00:04:45,019 --> 00:04:49,500 The tool simply wraps the WikipediaQueryRun object, passing the input query to its 71 00:04:49,500 --> 00:04:51,500 run method. 72 00:04:51,500 --> 00:04:56,019 Using the invoke method, the tool searches Wikipedia for the query, what is tool calling, 73 00:04:56,019 --> 00:04:59,059 and returns a summary from Wikipedia as the output. 74 00:04:59,059 --> 00:05:03,700 Next, create a new agent that combines the math tools with the Search_Wikipedia 75 00:05:03,700 --> 00:05:04,700 tool. 76 00:05:04,700 --> 00:05:08,820 For this, create an updated tools list that includes the Wikipedia tool alongside the 77 00:05:08,980 --> 00:05:10,420 math tools. 78 00:05:10,420 --> 00:05:14,660 Then use "create react agent" to build a new agent that can both perform calculations and 79 00:05:14,660 --> 00:05:18,739 fetch information from Wikipedia. 80 00:05:18,739 --> 00:05:24,059 Now test the updated agent with a query that combines information retrieval and math operation. 81 00:05:24,059 --> 00:05:29,480 Provide the input, what is the population of Canada, multiply it by 0.75, to the agent 82 00:05:29,480 --> 00:05:30,980 using invoke. 83 00:05:30,980 --> 00:05:35,640 The agent first uses Search_Wikipedia to find Canada's population. 84 00:05:35,640 --> 00:05:39,600 Then it applies the multiplication tool to calculate 75% of that number. 85 00:05:39,600 --> 00:05:44,200 Finally, it returns a natural language response with a result. 86 00:05:44,200 --> 00:05:46,079 In this video, you learned to 87 00:05:46,079 --> 00:05:50,820 Build React-style agents using the Create_React_Agent method for 88 00:05:50,820 --> 00:05:53,519 greater customization and control. 89 00:05:53,519 --> 00:05:59,160 Construct a multi-tool math assistant by combining tools for addition, subtraction, multiplication, 90 00:05:59,160 --> 00:06:00,160 and division. 91 00:06:00,160 --> 00:06:04,440 Guide agent behavior using custom prompts and structure tool inputs. 92 00:06:04,440 --> 00:06:09,440 Orchestrate multiple tools within a single agent to handle real-world multi-step queries. 93 00:06:09,440 --> 00:06:13,399 Extend agent functionality by integrating external tools, such as Wikipedia Search for 94 00:06:13,399 --> 00:06:15,119 dynamic, hybrid responses.