02. Tool Binding (Binding Tools)

In order for the LLM model to be able to call the tool, it is necessary to pass the tool schema to the model when making a chat request.

LangChain Chat Model supporting tool calling function .bind_tools() Implement a method to receive a list of LangChain tool objects, Pydantic classes or JSON schemas and bind to the chat model in the expected format by provider.

Subsequent calls to the bound Chat Model include a tool schema in every call to the model API.

Copy

# Configuration file for managing API KEY as environment variable
from dotenv import load_dotenv

# Load API KEY information
load_dotenv()

Copy

 True 

Copy

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

# Enter a project name.
logging.langsmith("CH15-Bind-Tools")

Copy

 Start tracking LangSmith. 
[Project name] 
CH15-Bind-Tools 

Tool definition to bind to LLM

Defines the tool for experimentation.

• get_word_length : A function that returns the length of a word

• add_function : A function that adds two numbers

• naver_news_crawl : A function that crawls a Naver news article to return the text content

Reference

• When defining tools @tool Define tools using decorators.

• docstring is recommended to write in English whenever possible.

Copy

import re
import requests
from bs4 import BeautifulSoup
from langchain.agents import tool


# Defines a tool.

@tool
def get_word_length(word: str) -> int:
    """Returns the length of a word."""
    return len(word)


@tool
def add_function(a: float, b: float) -> float:
    """Adds two numbers together."""
    return a + b


@tool
def naver_news_crawl(news_url: str) -> str:
    """Crawls a 네이버 (naver.com) news article and returns the body content."""
    # Sending an HTTP GET request
    response = requests.get(news_url)

    # Check if the request was successful
    if response.status_code == 200:
        # Parsing HTML using BeautifulSoup
        soup = BeautifulSoup(response.text, "html.parser")

        # Extract the information you want
        title = soup.find("h2", id="title_area").get_text()
        content = soup.find("div", id="contents").get_text()
        cleaned_title = re.sub(r"\n{2,}", "\n", title)
        cleaned_content = re.sub(r"\n{2,}", "\n", content)
    else:
        print(f"HTTP Request failed. Response code: {response.status_code}")

    return f"{cleaned_title}\n{cleaned_content}"


tools = [get_word_length, add_function, naver_news_crawl]

Binding tools to LLM with bind_tools()

llm model bind_tools() Use to bind the tool.

Copy

from langchain_openai import ChatOpenAI

# Create a model
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

# Tool binding
llm_with_tools = llm.bind_tools(tools)

Check the results.

result tool_calls Stored in. therefore, .tool_calls You can check the tool call result by checking.

Reference

• name Means the name of the silver tool.

• args means the factor passed to the tool.

Copy

# Execution results
llm_with_tools.invoke("What is the length of the word 'teddynote'?").tool_calls

Copy

 [{'name':'get_word_length','args': {'word':'teddynote'},'id':'call_ReeUlCuSd35Marox3yaxMgcH','type':'tool_call'}] 

Next llm_with_tools Wow JsonOutputToolsParser By connecting tool_calls Parsing to confirm the result.

Copy

from langchain_core.output_parsers.openai_tools import JsonOutputToolsParser

# Tool binding + tool parser
chain = llm_with_tools | JsonOutputToolsParser(tools=tools)

# Execution results
tool_call_results = chain.invoke("What is the length of the word 'teddynote'?")

The results of the run are as follows.

Reference

• type : The name of the tool

• args : Factors passed to the tool

Copy

print(tool_call_results, end="\n\n==========\n\n")
# Results of the first tool call
single_result = tool_call_results[0]
# Tool name
print(single_result["type"])
# Tool argument
print(single_result["args"])

Copy

 [{'args': {'word':'teddynote'},'type':'get_word_length'}] 

========== 

get_word_length 
{'word':'teddynote'} 

Find and run tools that match tool names.

Copy

tool_call_results[0]["type"], tools[0].name

Copy

 ('get_word_length','get_word_length') 

execute_tool_calls The function finds the tool and passes args to run the tool.

In other words, type Meaning the name of the silver tool args means the factor passed to the tool.

Copy

def execute_tool_calls(tool_call_results):
    """
    Function that executes the result of a tool call

    :param tool_call_results: Tool call result list
    :param tools: List of available tools
    """
    # Iterate over the list of tool call results.
    for tool_call_result in tool_call_results:
        # Extracts the tool name and arguments.
        tool_name = tool_call_result["type"]
        tool_args = tool_call_result["args"]

        # 도구 이름과 일치하는 도구를 찾아 실행합니다.
        # next() 함수를 사용하여 일치하는 첫 번째 도구를 찾습니다.
        matching_tool = next((tool for tool in tools if tool.name == tool_name), None)

        if matching_tool:
            # If you find a matching tool, run it.
            result = matching_tool.invoke(tool_args)
            # Prints the execution results.
            print(f"[실행도구] {tool_name}\n[실행결과] {result}")
        else:
            # If no matching tool is found, a warning message is printed.
            print(f"경고: {tool_name}에 해당하는 도구를 찾을 수 없습니다.")


# Execute tool call
# Execute the function by passing the previously obtained tool_call_results as an argument.
execute_tool_calls(tool_call_results)

Copy

 [Executives] get_word_length 
[Execution Results] 9 

bind_tools + Parser + Execution

This time, a series of processes are run at once.

• llm_with_tools : Model with tools bound

• JsonOutputToolsParser : Parser parsing tool call results

• execute_tool_calls : Function running tool call result

Flow cleanup One. Binding tools to the model 2. Parsing tool call results 3. Run tool call results

Copy

from langchain_core.output_parsers.openai_tools import JsonOutputToolsParser

# bind_tools + Parser + Execution
chain = llm_with_tools | JsonOutputToolsParser(tools=tools) | execute_tool_calls

Copy

# Execution results
chain.invoke("What is the length of the word 'teddynote'?")

Copy

 [Executives] get_word_length 
[Execution Results] 9 

Copy

# Execution results
chain.invoke("114.5 + 121.2")
print(114.5 + 121.2)

Copy

 [Executives] add_function 
[Execution Results] 235.7 
235.7 

Copy

# Execution results
chain.invoke(
    "Crawl the news article content: https://n.news.naver.com/mnews/hotissue/article/092/0002347672?type=series&cid=2000065"
)

Copy

 [Executives] naver_news_crawl 
[Execution] [Unemployed] Pay attention to the US CPI in September... S&P500 performance announcement in the third quarter 

            ▲10 days (local time) US September Consumer Index (CPI) scheduled to be announced. Pay attention to the price indicators, with the employment indicators coming out in good condition. In September, 250,000 non-agricultural employment in the United States increased the market estimate by 140,000. This is the largest in the sixth round, higher than the average increase of 200,300 people in the past 12 months. The unemployment rate in September fell to 4.1% for the second consecutive month, with average hourly wages increasing 4% year-on-year to the highest level since May.In the ▲market, the September headline CPI was 2.6% compared to the previous year in August, slowing for the fifth consecutive month and expecting further declines to 2.3% increase in September. Expected to slow 0.2% in August compared to the previous month. The source CPI observed levels similar to 3.2% in August.On the 11th of ▲, the United States September Manufacturing Index (PPI) was announced. The rebound trend has been broken until August from June to August. August increased by 1.7%. 
(Photo=Image Today) ▲ November US Federal Reserve (FOMC) public market (FOMC) expects a 0.50%p cut to crack down, FOMC minutes released on the 9th. At the last September meeting, it seems that you will be able to see the discussion of the big cut (0.50%p) shortening background and inflation outlook.▲US Standard and Poor (S&P) 500 companies begin their third quarter earnings announcement. The average margin rate estimate is 4.6% compared to the same period last year, with a plus or increase of five consecutive seasons expected to slow. From the 11th, the performance of large banks such as JPMorgance and Wellsfago was announced.▲FTSE Russell announces regular market classification results on the 8th. Korea is interested in incorporating into the World Debt (WGBI) this year after the designation of the observational target in 2022, up the market accessibility rating (L1). For stocks, we are interested in designating the country of observation for reasons such as short selling limits for Korean stocks that have been incorporated into the developed country index since 2009. Even if specified, there is a review period and is not immediately excluded. 
Reporter Son Hee-yeon ([e-mail]) 
Reporter Profile 
Reporter Seon Hee-yeon 
subscription 
Subscription 
Subscriber 0 
cheer 
0 
Covers the digital grafted financial and food and beverage industry.  
Do you apply for a loan without knowing it?... "Pre-blocking available through the invang" 
The third financial accident for our bank this year... "55 billion loans with false positives" 
Subscribe to Gidinet Korea Subscribe to many reporters! 
close 
Copyright ⓒ ZDNet Korea. All rights reserved. Prohibited unauthorized transfer and redistribution. 

This article was classified by the media as an IT section. 
Article section classification guide 
The section information in the article follows the classification of the media. Media outlets can categorize individual articles into two or more sections. 
close 
subscription 
Subscribe to the media editing news right now on the main! 
Subscription 
Check out the news of the media editing right in Maine right now! 
Gidinet Korea'Homepage' 
Clicking on the QR code will make it look great. 
Take a QR. 
Gidinet Korea'Homepage' 
close 
Subscribe to Naver Channel 
            Featured News Selected by the media and can be moved to the media page (outlink). 
KB Finance supports 1 million won per month for women self-employed in Gwangju 
Apartment only? Now Villa Zundam-do Galatia 
Kakaobank suspends financial transactions at dawn 6 
The third financial accident for our bank this year... "55 billion loans with false positives" 
I recommend this article 
Article recommendations can participate up to 50 meetings within 24 hours. 
close 
Shoot 
0 
exciting 
0 
Sympathy 
0 
Analytical month 
0 
Follow-up 
0 

What article would you like to show everyone?beta 
Please press the button to recommend this article.  Articles that have been highly recommended during the aggregation period are used as referral elements in the Naver Auto Tombstone area. 
Close layer 

series 
Miso-briefing 
Subscription 
subscription 
Subscription 
Thank you for subscribing 
Subscription serials can be checked on the settings page 
close 
View more 
US CPI attention in September... 3rd quarter S&P500 performance announcement 
Middle Eastern tension heightened... Crude oil price surged 
Pay attention to the announcement of the September employment card... The vice presidential debate is held. 
Zidinet Korea Press Home Shortcut 

View articles after reporter subscription 
Keep looking without subscription 
View articles after subscription 
Keep looking without subscription 

replaced by bind_tools > Agent & AgentExecutor

bind_tools() Provides a schema (tool) that can be used for models.

AgentExecutor It actually creates an execution loop for llm calls, routing, running, model re-calling, etc. with the right tools.

Reference

• Agent Wow AgentExecutor The following chapters are covered in detail.

Copy

from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_openai import ChatOpenAI

# Create Agent Prompt
prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are very powerful assistant, but don't know current events",
        ),
        ("user", "{input}"),
        MessagesPlaceholder(variable_name="agent_scratchpad"),
    ]
)

# Create a model
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

Copy

from langchain.agents import create_tool_calling_agent
from langchain.agents import AgentExecutor

# Use previously defined tools
tools = [get_word_length, add_function, naver_news_crawl]

# Create Agent
agent = create_tool_calling_agent(llm, tools, prompt)

# Create AgentExecutor
agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    verbose=True,
    handle_parsing_errors=True,
)

Copy

# Running Agent
result = agent_executor.invoke({"input": "How many letters in the word `teddynote`?"})

# Check the results
print(result["output"])

Copy

Entering new AgentExecutor chain... 

Invoking: `get_word_length` with `{'word':'teddynote'}` 


The word "teddynote" has 9 letters. 

Finished chain. 
The word "teddynote" has 9 letters. 

Copy

# Running Agent
result = agent_executor.invoke({"input": "114.5 + 121.2 the calculation result is?"})

# Check the results
print(result["output"])

Copy

Entering new AgentExecutor chain... 

Invoking: `add_function` with `{'a': 114.5,'b': 121.2}` 


The calculation result for 235.7114.5 + 121.2 is 235.7. 

Finished chain. 
The calculation result for 114.5 + 121.2 is 235.7. 

It doesn't end with a single run, but the model goes through the process of checking their results and calling themselves again.

Copy

# Running Agent
result = agent_executor.invoke(
    {"input": "114.5 + 121.2 + 34.2 + 110.1 의 계산 결과는?"}
)

# Check the results
print(result["output"])
print("==========\n")
print(114.5 + 121.2 + 34.2 + 110.1)

Copy

Entering new AgentExecutor chain... 

Invoking: `add_function` with `{'a': 114.5,'b': 121.2}` 


235.7 
Invoking: `add_function` with `{'a': 34.2,'b': 110.1}` 


144.3 
Invoking: `add_function` with `{'a': 235.7,'b': 144.3}` 


380.0114.5 + 121.2 + 34.2 + The calculation result of 110.1 is 380.0. 

Finished chain. 
The calculation result for 114.5 + 121.2 + 34.2 + 110.1 is 380.0. 
========== 

380.0 

This time, we will crawl the news results and make a request to summarize.

Copy

result = agent_executor.invoke(
    {
        "input": "summarize the news article: https://n.news.naver.com/mnews/hotissue/article/092/0002347672?type=series&cid=2000065"
    }
)
print(result["output"])

Copy

Entering new AgentExecutor chain... 

Invoking: `naver_news_crawl` with `{'news_url':'https://n.news.naver.com/mnews/hotissue/article/092/0002347672?type=series&cid=2000065'}` 


[Unemployed] September US CPI attention... 3rd quarter S&P500 performance announcement 

            ▲10 days (local time) US September Consumer Index (CPI) scheduled to be announced. Pay attention to the price indicators, with the employment indicators coming out in good condition. In September, 250,000 non-agricultural employment in the United States increased the market estimate by 140,000. This is the largest in the sixth round, higher than the average increase of 200,300 people in the past 12 months. The unemployment rate in September fell to 4.1% for the second consecutive month, with average hourly wages increasing 4% year-on-year to the highest level since May.In the ▲market, the September headline CPI was 2.6% compared to the previous year in August, slowing for the fifth consecutive month and expecting further declines to 2.3% increase in September. Expected to slow 0.2% in August compared to the previous month. The source CPI observed levels similar to 3.2% in August.On the 11th of ▲, the United States September Manufacturing Index (PPI) was announced. The rebound trend has been broken until August from June to August. August increased by 1.7%. 
(Photo=Image Today) ▲ November US Federal Reserve (FOMC) public market (FOMC) expects a 0.50%p cut to crack down, FOMC minutes released on the 9th. At the last September meeting, it seems that you will be able to see the discussion of the big cut (0.50%p) shortening background and inflation outlook.▲US Standard and Poor (S&P) 500 companies begin their third quarter earnings announcement. The average margin rate estimate is 4.6% compared to the same period last year, with a plus or increase of five consecutive seasons expected to slow. From the 11th, the performance of large banks such as JPMorgance and Wellsfago was announced.▲FTSE Russell announces regular market classification results on the 8th. Korea is interested in incorporating into the World Debt (WGBI) this year after the designation of the observational target in 2022, up the market accessibility rating (L1). For stocks, we are interested in designating the country of observation for reasons such as short selling limits for Korean stocks that have been incorporated into the developed country index since 2009. Even if specified, there is a review period and is not immediately excluded. 
Reporter Son Hee-yeon ([e-mail]) 
Reporter Profile 
Reporter Seon Hee-yeon 
subscription 
Subscription 
Subscriber 0 
cheer 
0 
Covers the digital grafted financial and food and beverage industry.  
Do you apply for a loan without knowing it?... "Pre-blocking available through the invang" 
The third financial accident for our bank this year... "55 billion loans with false positives" 
Subscribe to Gidinet Korea Subscribe to many reporters! 
close 
Copyright ⓒ ZDNet Korea. All rights reserved. Prohibited unauthorized transfer and redistribution. 

This article was classified by the media as an IT section. 
Article section classification guide 
The section information in the article follows the classification of the media. Media outlets can categorize individual articles into two or more sections. 
close 
subscription 
Subscribe to the media editing news right now on the main! 
Subscription 
Check out the news of the media editing right in Maine right now! 
Gidinet Korea'Homepage' 
Clicking on the QR code will make it look great. 
Take a QR. 
Gidinet Korea'Homepage' 
close 
Subscribe to Naver Channel 
            Featured News Selected by the media and can be moved to the media page (outlink). 
KB Finance supports 1 million won per month for women self-employed in Gwangju 
Apartment only? Now Villa Zundam-do Galatia 
Kakaobank suspends financial transactions at dawn 6 
The third financial accident for our bank this year... "55 billion loans with false positives" 
I recommend this article 
Article recommendations can participate up to 50 meetings within 24 hours. 
close 
Shoot 
0 
exciting 
0 
Sympathy 
0 
Analytical month 
0 
Follow-up 
0 

What article would you like to show everyone?beta 
Please press the button to recommend this article.  Articles that have been highly recommended during the aggregation period are used as referral elements in the Naver Auto Tombstone area. 
Close layer 

series 
Miso-briefing 
Subscription 
subscription 
Subscription 
Thank you for subscribing 
Subscription serials can be checked on the settings page 
close 
View more 
US CPI attention in September... 3rd quarter S&P500 performance announcement 
Middle Eastern tension heightened... Crude oil price surged 
Pay attention to the announcement of the September employment card... The vice presidential debate is held. 
Zidinet Korea Press Home Shortcut 

View articles after reporter subscription 
Keep looking without subscription 
View articles after subscription 
Keep looking without subscription 
Article summary: 

On September 10 (local time), the United States' Consumer Index (CPI) announcement is scheduled, and interest in price indicators is focused as employment indicators are good. Non-agricultural employment in September increased by 254,000, significantly above market expectations, and the unemployment rate fell 4.1% for the second consecutive month. Average hourly wages increased 4% year over year. 

The headline CPI in September was 2.6% compared to August, slowing for five consecutive months, and is expected to increase by 2.3% in September. On the 11th, a manufacturing price index (PPI) is also scheduled. In addition, the announcement of the third quarter earnings for the US S&P 500 company begins, and the average profit growth rate is expected to be 4.6% compared to the same period last year. 

FTSE Russell is raising interest in the possibility of raising Korea's market accessibility rating. 

Finished chain. 
Article summary: 

On September 10 (local time), the United States' Consumer Index (CPI) announcement is scheduled, and interest in price indicators is focused as employment indicators are good. Non-agricultural employment in September increased by 254,000, significantly above market expectations, and the unemployment rate fell 4.1% for the second consecutive month. Average hourly wages increased 4% year over year. 

The headline CPI in September was 2.6% compared to August, slowing for five consecutive months, and is expected to increase by 2.3% in September. On the 11th, a manufacturing price index (PPI) is also scheduled. In addition, the announcement of the third quarter earnings for the US S&P 500 company begins, and the average profit growth rate is expected to be 4.6% compared to the same period last year. 

FTSE Russell is raising interest in the possibility of raising Korea's market accessibility rating.