Hooks

了解如何使用工具钩子来修改工具的行为。

工具钩子

您可以使用工具钩子在调用工具之前或之后执行验证、日志记录或任何其他逻辑。

工具钩子是一个函数，它接受函数名称、函数调用和参数。可选地，您还可以访问 Agent 或 Team 对象。在工具钩子内部，您必须调用该函数调用并返回结果。

定义工具钩子时，使用精确的参数名称非常重要。可用的参数有 agent、team、function_name、function_call 和 arguments。

例如：

def logger_hook(
    function_name: str, function_call: Callable, arguments: Dict[str, Any]
):
    """Log the duration of the function call"""
    start_time = time.time()

    # Call the function
    result = function_call(**arguments)

    end_time = time.time()
    duration = end_time - start_time

    logger.info(f"Function {function_name} took {duration:.2f} seconds to execute")

    # Return the result
    return result

或者

def confirmation_hook(
    function_name: str, function_call: Callable, arguments: Dict[str, Any]
):
    """Confirm the function call"""
    if function_name != "get_top_hackernews_stories":
        raise ValueError("This tool is not allowed to be called")
    return function_call(**arguments)

您可以为 agents 和 teams 分配工具钩子。这些工具钩子将应用于 agent 或 team 所做的所有工具调用。

例如：

agent = Agent(
    model=OpenAIChat(id="gpt-4o-mini"),
    tools=[DuckDuckGoTools()],
    tool_hooks=[logger_hook],
)

您还可以在工具钩子中访问 Agent 或 Team 对象。

def grab_customer_profile_hook(
    agent: Agent, function_name: str, function_call: Callable, arguments: Dict[str, Any]
):
    cust_id = arguments.get("customer")
    if cust_id not in agent.session_state["customer_profiles"]:
        raise ValueError(f"Customer profile for {cust_id} not found")
    customer_profile = agent.session_state["customer_profiles"][cust_id]

    # Replace the customer with the customer_profile for the function call
    arguments["customer"] = json.dumps(customer_profile)
    # Call the function with the updated arguments
    result = function_call(**arguments)

    return result

多个工具钩子

您还可以一次性分配多个工具钩子。它们将按分配的顺序应用。

agent = Agent(
    model=OpenAIChat(id="gpt-4o-mini"),
    tools=[DuckDuckGoTools()],
    tool_hooks=[logger_hook, confirmation_hook],  # The logger_hook will run on the outer layer, and the confirmation_hook will run on the inner layer
)

您还可以为特定的自定义工具分配工具钩子。

@tool(tool_hooks=[logger_hook, confirmation_hook])
def get_top_hackernews_stories(num_stories: int) -> Iterator[str]:
    """Fetch top stories from Hacker News.

    Args:
        num_stories (int): Number of stories to retrieve
    """
    # Fetch top story IDs
    response = httpx.get("https://hacker-news.firebaseio.com/v0/topstories.json")
    story_ids = response.json()

    # Yield story details
    final_stories = []
    for story_id in story_ids[:num_stories]:
        story_response = httpx.get(
            f"https://hacker-news.firebaseio.com/v0/item/{story_id}.json"
        )
        story = story_response.json()
        if "text" in story:
            story.pop("text", None)
        final_stories.append(story)

    return json.dumps(final_stories)

agent = Agent(
    model=OpenAIChat(id="gpt-4o-mini"),
    tools=[get_top_hackernews_stories],
)

前置钩子和后置钩子

前置钩子和后置钩子允许您修改在工具调用之前和之后发生的事情。它们是工具钩子的替代方案。

设置 @tool 装饰器中的 pre_hook 以在工具调用之前运行一个函数。

设置 @tool 装饰器中的 post_hook 以在工具调用之后运行一个函数。

这是一个使用 pre_hook、post_hook 以及 Agent Context 的演示示例。

pre_and_post_hooks.py

import json
from typing import Iterator

import httpx
from agno.agent import Agent
from agno.tools import FunctionCall, tool


def pre_hook(fc: FunctionCall):
    print(f"Pre-hook: {fc.function.name}")
    print(f"Arguments: {fc.arguments}")
    print(f"Result: {fc.result}")


def post_hook(fc: FunctionCall):
    print(f"Post-hook: {fc.function.name}")
    print(f"Arguments: {fc.arguments}")
    print(f"Result: {fc.result}")


@tool(pre_hook=pre_hook, post_hook=post_hook)
def get_top_hackernews_stories(agent: Agent) -> Iterator[str]:
    num_stories = agent.context.get("num_stories", 5) if agent.context else 5

    # Fetch top story IDs
    response = httpx.get("https://hacker-news.firebaseio.com/v0/topstories.json")
    story_ids = response.json()

    # Yield story details
    for story_id in story_ids[:num_stories]:
        story_response = httpx.get(
            f"https://hacker-news.firebaseio.com/v0/item/{story_id}.json"
        )
        story = story_response.json()
        if "text" in story:
            story.pop("text", None)
        yield json.dumps(story)


agent = Agent(
    context={
        "num_stories": 2,
    },
    tools=[get_top_hackernews_stories],
    markdown=True,
    show_tool_calls=True,
)
agent.print_response("What are the top hackernews stories?", stream=True)