通过返回直接工具控制代理推理循环¶

所有工具都有一个 return_direct 选项——如果将其设置为 True，并且调用了关联工具（未调用其他任何工具），则会终止代理推理循环并直接返回工具输出。

这在以下场景非常有用：当您确定工具输出已足够理想时，可以加快响应速度；避免代理重写响应；以及终止推理循环。

本笔记本演示了一个场景：智能代理需要通过向用户收集信息来完成餐厅预订。

In [ ]:

%pip install llama-index-core llama-index-llms-anthropic

%pip install llama-index-core llama-index-llms-anthropic

In [ ]:

import os

os.environ["ANTHROPIC_API_KEY"] = "sk-..."

import os os.environ["ANTHROPIC_API_KEY"] = "sk-..."

工具配置¶

In [ ]:

from typing import Optional

from llama_index.core.tools import FunctionTool
from pydantic import BaseModel

# we will store booking under random IDs
bookings = {}


# we will represent and track the state of a booking as a Pydantic model
class Booking(BaseModel):
    name: Optional[str] = None
    email: Optional[str] = None
    phone: Optional[str] = None
    date: Optional[str] = None
    time: Optional[str] = None


def get_booking_state(user_id: str) -> str:
    """Get the current state of a booking for a given booking ID."""
    try:
        return str(bookings[user_id].dict())
    except:
        return f"Booking ID {user_id} not found"


def update_booking(user_id: str, property: str, value: str) -> str:
    """Update a property of a booking for a given booking ID. Only enter details that are explicitly provided."""
    booking = bookings[user_id]
    setattr(booking, property, value)
    return f"Booking ID {user_id} updated with {property} = {value}"


def create_booking(user_id: str) -> str:
    """Create a new booking and return the booking ID."""
    bookings[user_id] = Booking()
    return "Booking created, but not yet confirmed. Please provide your name, email, phone, date, and time."


def confirm_booking(user_id: str) -> str:
    """Confirm a booking for a given booking ID."""
    booking = bookings[user_id]

    if booking.name is None:
        raise ValueError("Please provide your name.")

    if booking.email is None:
        raise ValueError("Please provide your email.")

    if booking.phone is None:
        raise ValueError("Please provide your phone number.")

    if booking.date is None:
        raise ValueError("Please provide the date of your booking.")

    if booking.time is None:
        raise ValueError("Please provide the time of your booking.")

    return f"Booking ID {user_id} confirmed!"


# create tools for each function
get_booking_state_tool = FunctionTool.from_defaults(fn=get_booking_state)
update_booking_tool = FunctionTool.from_defaults(fn=update_booking)
create_booking_tool = FunctionTool.from_defaults(
    fn=create_booking, return_direct=True
)
confirm_booking_tool = FunctionTool.from_defaults(
    fn=confirm_booking, return_direct=True
)

from typing import Optional from llama_index.core.tools import FunctionTool from pydantic import BaseModel # we will store booking under random IDs bookings = {} # we will represent and track the state of a booking as a Pydantic model class Booking(BaseModel): name: Optional[str] = None email: Optional[str] = None phone: Optional[str] = None date: Optional[str] = None time: Optional[str] = None def get_booking_state(user_id: str) -> str: """Get the current state of a booking for a given booking ID.""" try: return str(bookings[user_id].dict()) except: return f"Booking ID {user_id} not found" def update_booking(user_id: str, property: str, value: str) -> str: """Update a property of a booking for a given booking ID. Only enter details that are explicitly provided.""" booking = bookings[user_id] setattr(booking, property, value) return f"Booking ID {user_id} updated with {property} = {value}" def create_booking(user_id: str) -> str: """Create a new booking and return the booking ID.""" bookings[user_id] = Booking() return "Booking created, but not yet confirmed. Please provide your name, email, phone, date, and time." def confirm_booking(user_id: str) -> str: """Confirm a booking for a given booking ID.""" booking = bookings[user_id] if booking.name is None: raise ValueError("Please provide your name.") if booking.email is None: raise ValueError("Please provide your email.") if booking.phone is None: raise ValueError("Please provide your phone number.") if booking.date is None: raise ValueError("Please provide the date of your booking.") if booking.time is None: raise ValueError("Please provide the time of your booking.") return f"Booking ID {user_id} confirmed!" # create tools for each function get_booking_state_tool = FunctionTool.from_defaults(fn=get_booking_state) update_booking_tool = FunctionTool.from_defaults(fn=update_booking) create_booking_tool = FunctionTool.from_defaults( fn=create_booking, return_direct=True ) confirm_booking_tool = FunctionTool.from_defaults( fn=confirm_booking, return_direct=True )

用户已进入！让我们协助完成预订¶

In [ ]:

from llama_index.llms.anthropic import Anthropic
from llama_index.core.llms import ChatMessage
from llama_index.core.agent.workflow import FunctionAgent
from llama_index.core.workflow import Context

llm = Anthropic(model="claude-3-sonnet-20240229", temperature=0.1)

user = "user123"
system_prompt = f"""You are now connected to the booking system and helping {user} with making a booking.
Only enter details that the user has explicitly provided.
Do not make up any details.
"""

agent = FunctionAgent(
    tools=[
        get_booking_state_tool,
        update_booking_tool,
        create_booking_tool,
        confirm_booking_tool,
    ],
    llm=llm,
    system_prompt=system_prompt,
)

# create a context for the agent to hold the state/history of a session
ctx = Context(agent)

from llama_index.llms.anthropic import Anthropic from llama_index.core.llms import ChatMessage from llama_index.core.agent.workflow import FunctionAgent from llama_index.core.workflow import Context llm = Anthropic(model="claude-3-sonnet-20240229", temperature=0.1) user = "user123" system_prompt = f"""You are now connected to the booking system and helping {user} with making a booking. Only enter details that the user has explicitly provided. Do not make up any details. """ agent = FunctionAgent( tools=[ get_booking_state_tool, update_booking_tool, create_booking_tool, confirm_booking_tool, ], llm=llm, system_prompt=system_prompt, ) # create a context for the agent to hold the state/history of a session ctx = Context(agent)

In [ ]:

from llama_index.core.agent.workflow import AgentStream, ToolCallResult

handler = agent.run(
    "Hello! I would like to make a booking, around 5pm?", ctx=ctx
)

async for ev in handler.stream_events():
    if isinstance(ev, AgentStream):
        print(f"{ev.delta}", end="", flush=True)
    elif isinstance(ev, ToolCallResult):
        print(
            f"\nCall {ev.tool_name} with {ev.tool_kwargs}\nReturned: {ev.tool_output}"
        )

response = await handler

from llama_index.core.agent.workflow import AgentStream, ToolCallResult handler = agent.run( "Hello! I would like to make a booking, around 5pm?", ctx=ctx ) async for ev in handler.stream_events(): if isinstance(ev, AgentStream): print(f"{ev.delta}", end="", flush=True) elif isinstance(ev, ToolCallResult): print( f"\nCall {ev.tool_name} with {ev.tool_kwargs}\nReturned: {ev.tool_output}" ) response = await handler

Okay, let's create a new booking for you.{"user_id": "user123"}
Call create_booking with {'user_id': 'user123'}
Returned: Booking created, but not yet confirmed. Please provide your name, email, phone, date, and time.

In [ ]:

print(str(response))

print(str(response))

Booking created, but not yet confirmed. Please provide your name, email, phone, date, and time.

完美，我们可以看到函数输出被直接返回，未经任何修改或最终的LLM调用！

In [ ]:

handler = agent.run(
    "Sure! My name is Logan, and my email is test@gmail.com?", ctx=ctx
)

async for ev in handler.stream_events():
    if isinstance(ev, AgentStream):
        print(f"{ev.delta}", end="", flush=True)
    elif isinstance(ev, ToolCallResult):
        print(
            f"\nCall {ev.tool_name} with {ev.tool_kwargs}\nReturned: {ev.tool_output}"
        )

response = await handler

handler = agent.run( "Sure! My name is Logan, and my email is test@gmail.com?", ctx=ctx ) async for ev in handler.stream_events(): if isinstance(ev, AgentStream): print(f"{ev.delta}", end="", flush=True) elif isinstance(ev, ToolCallResult): print( f"\nCall {ev.tool_name} with {ev.tool_kwargs}\nReturned: {ev.tool_output}" ) response = await handler

Got it, thanks for providing your name and email. I've updated the booking with that information.{"user_id": "user123", "property": "name", "value": "Logan"}{"user_id": "user123", "property": "email", "value": "test@gmail.com"}
Call update_booking with {'user_id': 'user123', 'property': 'name', 'value': 'Logan'}
Returned: Booking ID user123 updated with name = Logan

Call update_booking with {'user_id': 'user123', 'property': 'email', 'value': 'test@gmail.com'}
Returned: Booking ID user123 updated with email = test@gmail.com
Please also provide your phone number, preferred date, and time for the booking.

In [ ]:

print(str(response))

print(str(response))

Please also provide your phone number, preferred date, and time for the booking.

In [ ]:

handler = agent.run(
    "Right! My phone number is 1234567890, the date of the booking is April 5, at 5pm.",
    ctx=ctx,
)

async for ev in handler.stream_events():
    if isinstance(ev, AgentStream):
        print(f"{ev.delta}", end="", flush=True)
    elif isinstance(ev, ToolCallResult):
        print(
            f"\nCall {ev.tool_name} with {ev.tool_kwargs}\nReturned: {ev.tool_output}"
        )

response = await handler

handler = agent.run( "Right! My phone number is 1234567890, the date of the booking is April 5, at 5pm.", ctx=ctx, ) async for ev in handler.stream_events(): if isinstance(ev, AgentStream): print(f"{ev.delta}", end="", flush=True) elif isinstance(ev, ToolCallResult): print( f"\nCall {ev.tool_name} with {ev.tool_kwargs}\nReturned: {ev.tool_output}" ) response = await handler

Great, thank you for providing the additional details. I've updated the booking with your phone number, date, and time.{"user_id": "user123", "property": "phone", "value": "1234567890"}{"user_id": "user123", "property": "date", "value": "2023-04-05"}{"user_id": "user123", "property": "time", "value": "17:00"}
Call update_booking with {'user_id': 'user123', 'property': 'phone', 'value': '1234567890'}
Returned: Booking ID user123 updated with phone = 1234567890

Call update_booking with {'user_id': 'user123', 'property': 'date', 'value': '2023-04-05'}
Returned: Booking ID user123 updated with date = 2023-04-05

Call update_booking with {'user_id': 'user123', 'property': 'time', 'value': '17:00'}
Returned: Booking ID user123 updated with time = 17:00
Looks like I have all the necessary details. Let me confirm this booking for you.{"user_id": "user123"}
Call confirm_booking with {'user_id': 'user123'}
Returned: Booking ID user123 confirmed!

In [ ]:

print(str(response))

print(str(response))

Booking ID user123 confirmed!

In [ ]:

print(bookings["user123"])

print(bookings["user123"])

name='Logan' email='test@gmail.com' phone='1234567890' date='2023-04-05' time='17:00'