Outils personnalisés pour les agents

Développement d'applications LLM avec LangChain

Jonathan Bennion

AI Engineer & LangChain Contributor

Formats d'outils

from langchain_community.agent_toolkits.load_tools import load_tools

tools = load_tools(["llm-math"], llm=llm)

print(tools[0].name)

Calculator

print(tools[0].description)

Useful for when you need to answer questions about math.

Utilisé par le LLM/l’agent comme contexte pour déterminer quand l'appeler

Formats d'outils

print(tools[0].return_direct)

False

Définition d'une fonction personnalisée

def financial_report(company_name: str, revenue: int, expenses: int) -> str:
    """Generate a financial report for a company that calculates net income."""
    net_income = revenue - expenses

    report = f"Financial Report for {company_name}:
"
    report += f"Revenue: ${revenue}
"
    report += f"Expenses: ${expenses}
"
    report += f"Net Income: ${net_income}

"

    return report

Appel de la fonction

print(financial_report(company_name="LemonadeStand", revenue=100, expenses=50))

Financial Report for LemonadeStand:
Revenue: $100
Expenses: $50
Net Income: $50

Des fonctions aux outils

from langchain_core.tools import tool


@tool

def financial_report(company_name: str, revenue: int, expenses: int) -> str:
    """Generate a financial report for a company that calculates net income."""
    net_income = revenue - expenses

    report = f"Financial Report for {company_name}:
"
    report += f"Revenue: ${revenue}
"
    report += f"Expenses: ${expenses}
"
    report += f"Net Income: ${net_income}

"

    return report

Analyse de notre nouvel outil

print(financial_report.name)

print(financial_report.description)

print(financial_report.return_direct)

print(financial_report.args)

financial_report

Generate a financial report for a company that calculates net income.

False

{'company_name': {'title': 'Company Name', 'type': 'string'},
 'revenue': {'title': 'Revenue', 'type': 'integer'},
 'expenses': {'title': 'Expenses', 'type': 'integer'}}

Intégration de l'outil personnalisé

from langgraph.prebuilt import create_react_agent

llm = ChatOpenAI(model="gpt-4o-mini", api_key=openai_api_key, temperature=0)
agent = create_react_agent(llm, [financial_report])


messages = agent.invoke({"messages": [("human", "TechStack generated made $10 million
with $8 million of costs. Generate a financial report.")]})
print(messages)

Intégration de l'outil personnalisé

{'messages': [
    HumanMessage(content='TechStack generated made $10 million dollars with $8 million of...', ...),
    AIMessage(content='', ..., tool_calls=[{'name': 'financial_report',
                                            'args': {'company_name': 'TechStack',
                                                     'revenue': 10000000, 'expenses': 8000000}, ...),
    ToolMessage(content='Financial Report for TechStack:
Revenue: $10000000
Expenses...', ...),
    AIMessage(content='Here is the financial report for TechStack...', ...)
]}

Sorties de l'outil

print(messages['messages'][-1].content)

Here is the financial report for TechStack:
- Revenue: $10,000,000
- Expenses: $8,000,000
- Net Income: $2,000,000

Financial Report for TechStack:
Revenue: $10000000
Expenses: $8000000
Net Income: $2000000

Passons à la pratique !

Développement d'applications LLM avec LangChain