10. How to call a tool using ToolNode

In this tutorial, LangGraph's pre-built for calling tools pre-built of ToolNode Covers how to use it.

ToolNode Is LangChain Runnable, which takes the graph status with a list of messages as input and updates the status as a result of the tool call.

It is designed for immediate use with LangGraph's pre-built Agent, and has a suitable redo for the state. messages All keys included StateGraph Can work with

Copy

# Configuration file for managing API keys as environment variables
from dotenv import load_dotenv

# Load API key information
load_dotenv()

Copy

 True 

Copy

# Set up LangSmith tracking. https://smith.langchain.com
# !pip install -qU langchain-teddynote
from langchain_teddynote import logging

# Enter a project name.
logging.langsmith("CH17-LangGraph-Modules")

Copy

 Start tracking LangSmith. 
[Project name] 
CH17-LangGraph-Modules 

Tool definition

First, let's define the tools.

Copy

from langchain_core.messages import AIMessage
from langchain_core.tools import tool
from langchain_experimental.tools.python.tool import PythonAstREPLTool
from langchain_teddynote.tools import GoogleNews
from typing import List, Dict


# Create a tool
@tool
def search_news(query: str) -> List[Dict[str, str]]:
    """Search Google News by input keyword"""
    news_tool = GoogleNews()
    return news_tool.search_by_keyword(query, k=5)


@tool
def python_code_interpreter(code: str):
    """Call to execute python code."""
    return PythonAstREPLTool().invoke(code)

Next ToolNode Let's look at how to call the tool using.

Copy

from langgraph.prebuilt import ToolNode, tools_condition

# Create a tool list
tools = [search_news, python_code_interpreter]

# Initializing ToolNode
tool_node = ToolNode(tools)

ToolNode Calling manually

ToolNode works in graph state with message list.

• important : At this time, the last message in the list tool_calls Contains properties AIMessage Should be.

First, let's look at how to manually call the tool node.

Copy

# Create an AI message object containing a single tool call.ww
# www
message_with_single_tool_call = AIMessage(
    content="",
    tool_calls=[
        {
            "name": "search_news",  # tool name
            "args": {"query": "AI"},  # tool argument
            "id": "tool_call_id",  # Tool Call ID
            "type": "tool_call",  # Tool call type
        }
    ],
)

# Process messages and execute weather information requests via tool nodes
tool_node.invoke({"messages": [message_with_single_tool_call]})

Copy

 {'messages': [ToolMessage (content=' [{" url": "https://news.google.com/rss/articles/CBMiVkFV", "content": "中, meta AI \'Lama\' using military chatbot development -TAG1>, } "https://news.google.com/rss/articles/CBMifEFVX", "content": "Where's the University of Popularity?News>Natural ‘AI, Semiconductor ’, Humanities ‘Management ’ -Trends" }, {"url": "https://news.google.com/rss/articles/CBMiW0FVX", "content": "Natural to Occupational Difficulty ‘AI.  search_news', tool_call_id='tool_call_id')]} 

Generally AIMessage There is no need to create manually, it is automatically generated by all LangChain chat models that support tool calls.

Also AIMessage of tool_calls If you pass multiple tool calls to the parameter ToolNode You can use to make parallel tool calls.

Copy

# Create and initialize an AI message object that includes multiple tool calls.
message_with_multiple_tool_calls = AIMessage(
    content="",
    tool_calls=[
        {
            "name": "search_news",
            "args": {"query": "AI"},
            "id": "tool_call_id",
            "type": "tool_call",
        },
        {
            "name": "python_code_interpreter",
            "args": {"code": "print(1+2+3+4)"},
            "id": "tool_call_id",
            "type": "tool_call",
        },
    ],
)

# Execute multiple tool calls by passing the generated message to the tool node.ww
tool_node.invoke({"messages": [message_with_multiple_tool_calls]})

Copy

 {'messages': [ToolMessage (content=' [{" url": "https://news.google.com/rss/articles/CBMiVkFV", "content": "中, meta AI \'Lama\' using military chatbot development -TAG1>, } "https://news.google.com/rss/articles/CBMifEFVX", "content": "Where's the University of Popularity?·Natural ‘AI· Semiconductor ’, Humanities ‘Management ’ -Trends" }, {"url": "https://news.google.com/rss/articles/CBMiW0FV"}]', name='search_news', tool_call_id='tool_call  10\n', name='python_code_interpreter', tool_call_id='tool_call_id')]} 

for use with llm

In order to use a chat model with a tool call function, you must first ensure that the model recognizes the available tools.

This is ChatOpenAI In the model .bind_tools This is done by calling the method.

Copy

from langchain_openai import ChatOpenAI

# LLM model initialization and tool binding
model_with_tools = ChatOpenAI(model="gpt-4o-mini", temperature=0).bind_tools(tools)

Copy

# Tool call confirmation
model_with_tools.invoke("처음 5개의 소수를 출력하는 python code 를 작성해줘").tool_calls

Copy

 [{'name':'python_code_interpreter','args': {'code':'def first_n_primes(n):\n primes = []\n num = 2 # Starting from the first prime numbernn if is_prime:\n primes.append(num)\n num += 1\n return primes\n\n# Get the first 5 prime numbers\nfirst_n_primes(5)'},'id':'call_I4g0TPHwqHNQJt5 

As you can see, the AI message generated by the chat model already tool_calls Because it is stuffed, ToolNode You can pass it directly to.

Copy

# Processing messages via tool nodes and generating tool-based responses in the LLM model
tool_node.invoke(
    {
        "messages": [
            model_with_tools.invoke(
                "Write a python code that prints the first 5 prime numbers."
            )
        ]
    }
)

Copy

 {'messages': [ToolMessage (content='[2, 3, 5, 7, 11]', name='python_code_interpreter', tool_call_id='call_STkrvwhghY2TvFHE4ih 

Use with Agent

Next, within the LangGraph graph ToolNode Let's look at how to use.

Let's set up a graph implementation for Agent. this Agent Receive the query as input and repeatedly call the tools until you get enough information to solve the query.

With the tools you just defined ToolNode And you will use the OpenAI model.

Copy

# Importing types for LangGraph workflow state and message processing
from langgraph.graph import StateGraph, MessagesState, START, END


# Process messages and generate responses using the LLM model, returning responses that include tool calls.
def call_model(state: MessagesState):
    messages = state["messages"]
    response = model_with_tools.invoke(messages)
    return {"messages": [response]}


# Initialize a message state-based workflow graph
workflow = StateGraph(MessagesState)

# Define agent and tool nodes and add them to the workflow graph.
workflow.add_node("agent", call_model)
workflow.add_node("tools", tool_node)

# Connecting from a workflow start point to an agent node
workflow.add_edge(START, "agent")

# Set conditional branch from agent node, link to tool node or end pointwwwwwwwww
workflow.add_conditional_edges("agent", tools_condition)

# Circular connection from tool node to agent node
workflow.add_edge("tools", "agent")

# Connecting from agent node to endpoint
workflow.add_edge("agent", END)


# Compile defined workflow graph and generate executable application
app = workflow.compile()

Copy

from langchain_teddynote.graphs import visualize_graph

visualize_graph(app)

I'll run and check the results.

Copy

# Run and check the results
for chunk in app.stream(
    {"messages": [("human", "처음 5개의 소수를 출력하는 python code 를 작성해줘")]},
    stream_mode="values",
):
    # Print last message
    chunk["messages"][-1].pretty_print()

Copy

 ================================ Human Message ================================= 
Create a python code that outputs the first 5 decimals 
================================== Ai Message ================================== 
Tool Calls: 
  python_code_interpreter (call_dIBd3Y7ibDapRosEbMAunfdl) 
 Call ID: call_dIBd3Y7ibDapRosEbMAunfdl 
  Args: 
    code: default first_n_primes(n): 
    primes = [] 
    num = 2 # Starting from the first prime number 
    while len (primes) < n: 
        is_prime = True 
        for i in range (2, int (num**0.5) + 1): 
            if num% i == 0: 
                is_prime = False 
                break 
        if is_prime: 
            primes.append(num) 
        num += 1 
    return primes 

# Get the first 5 prime numbers 
first_n_primes(5) 
================================= Tool Message ================================= 
Name: python_code_interpreter 

[2, 3, 5, 7, 11] 
================================== Ai Message ================================== 

The first five decimals are: [2, 3, 5, 7, 11] 

Copy

# Perform a search query
for chunk in app.stream(
    {"messages": [("human", "search google news about AI")]},
    stream_mode="values",
):
    chunk["messages"][-1].pretty_print()

Copy

 ================================ Human Message ================================= 
search google news about AI 
================================== Ai Message ================================== 
Tool Calls: 
  search_news (call_prnhiLquU2lQmS6kmmP1UCJh) 
 Call ID: call_prnhiLquU2lQmS6kmmP1UCJh 
  Args: 
    query: AI 
================================= Tool Message ================================= 
Name: search_news 

News [{"url": "https://news.google.com/rss/articles/CBMiVkFVX3", "content": "中, meta AI'Rama' used to develop a military chatbot -ZDnet Korea"}, {"url": "https://news.google.com/rss/articles/CBMiW0FV", "content": "Big Tech Big4,'AI "Ariving War'. Channel closed by repulsion in one week -AI time" }, {"url": "https://news.google.com/rss/articles/CBMiekFVX3", "content": "AI has gone to fate... NVIDIA pushes Intel and transfers Daiji-Trends" }] 
================================== Ai Message ================================== 
Here are some recent news articles about AI: 

One. **[Use meta AI'Lama' to develop military chatbots-ZDnet Korea] (https://news.google.com/rss/articles/CBMiVkFVX)**  
   Meta is developing a military chatbot using its AI model'Llama'. 

2. **[Big Tech Big4,'AI No wonder War'... This year's investment 288 trillion won prospects-Unified News] (https://news.google.com/rss/articles/CBMiW0FVX)**  
   The big four tech companies are projected to invest 288 trillion won in AI this year. 

3. **[Natural sequence ‘AI·semiconductor ’ bullish... Humanities ‘Department of Business Administration ’ - KBS News] (https://news.google.com/rss/articles/CBMiW0FVX)**  
   In the job market, there is a strong demand for AI and semiconductor skills, while humanities graduates are leaning towards business studies. 

4. **[Polish station fired by people and introduced AI Human, channel closed by backlash in a week-AITimes] (https://news.google.com/rss/articles/CBMiakFVX3)**  
   A Polish broadcaster that replaced human workers with AI has closed its channel after a week due to backlash. 

5. **[AI has gone fateful... NVIDIA pushes Intel and moves in Dai-Trends] (https://news.google.com/rss/articles/CBMiekFVX3l)**  
   AI has played a crucial role in NVIDIA surpassing Intel to be included in the Dow Jones index. 

Feel free to click on the links for more details! 

Copy

# Performing queries without requiring tool calls
for chunk in app.stream(
    {"messages": [("human", "Hello? Nice to meet you.")]},
    stream_mode="values",
):
    chunk["messages"][-1].pretty_print()

Copy

 ================================ Human Message ================================= 
Hello? Nice to meet you 
================================== Ai Message ================================== 
Hello! Nice to meet you. How can I help you? 

ToolNode It can also handle errors that occur during tool execution.

handle_tool_errors=True You can enable/disable this function by setting (it is enabled by default)