构建智能代理

前提条件

提示

本指南假设您已熟悉以下概念：

• 聊天模型
• 工具
• 智能代理

语言模型本身无法执行操作 - 它们只能输出文本。LangChain的一个重要用例是创建智能代理。代理是使用LLM作为推理引擎的系统，用于确定要执行的操作及其输入。执行操作后，结果将反馈回LLM，以决定是否需要更多操作或完成任务。

本教程将构建一个可以与搜索引擎交互的代理。您可以向该代理提问，观察其调用搜索工具，并与其进行多轮对话。

端到端代理示例

以下代码展示了一个完整的智能代理，使用LLM决定调用的工具。它配备了通用搜索工具，并具有对话记忆功能 - 可以作为多轮聊天机器人使用。

# 导入相关模块
from langchain_anthropic import ChatAnthropic
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_core.messages import HumanMessage
from langgraph.checkpoint.memory import MemorySaver
from langgraph.prebuilt import create_react_agent

# 创建代理
memory = MemorySaver()
model = ChatAnthropic(model_name="claude-3-sonnet-20240229")
search = TavilySearchResults(max_results=2)
tools = [search]
agent_executor = create_react_agent(model, tools, checkpointer=memory)

# 使用代理
config = {"configurable": {"thread_id": "abc123"}}
for chunk in agent_executor.stream(
    {"messages": [HumanMessage(content="hi im bob! and i live in sf")]}, config
):
    print(chunk)
    print("----")
for chunk in agent_executor.stream(
    {"messages": [HumanMessage(content="whats the weather where I live?")]}, config
):
    print(chunk)
    print("----")

输出示例

{'agent': {'messages': [AIMessage(content="Hello Bob! Since you didn't ask a specific question, I don't need to use any tools to respond. It's nice to meet you. San Francisco is a wonderful city with lots to see and do. I hope you're enjoying living there. Please let me know if you have any other questions!", response_metadata={'id': 'msg_01Mmfzfs9m4XMgVzsCZYMWqH', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 271, 'output_tokens': 65}}, id='run-44c57f9c-a637-4888-b7d9-6d985031ae48-0', usage_metadata={'input_tokens': 271, 'output_tokens': 65, 'total_tokens': 336})]}}
----
{'agent': {'messages': [AIMessage(content=[{'text': 'To get current weather information for your location in San Francisco, let me invoke the search tool:', 'type': 'text'}, {'id': 'toolu_01BGEyQaSz3pTq8RwUUHSRoo', 'input': {'query': 'san francisco weather'}, 'name': 'tavily_search_results_json', 'type': 'tool_use'}], response_metadata={'id': 'msg_013AVSVsRLKYZjduLpJBY4us', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'tool_use', 'stop_sequence': None, 'usage': {'input_tokens': 347, 'output_tokens': 80}}, id='run-de7923b6-5ee2-4ebe-bd95-5aed4933d0e3-0', tool_calls=[{'name': 'tavily_search_results_json', 'args': {'query': 'san francisco weather'}, 'id': 'toolu_01BGEyQaSz3pTq8RwUUHSRoo'}], usage_metadata={'input_tokens': 347, 'output_tokens': 80, 'total_tokens': 427})]}}
----
{'tools': {'messages': [ToolMessage(content='[{"url": "https://www.weatherapi.com/", "content": "{\'location\': {\'name\': \'San Francisco\', \'region\': \'California\', \'country\': \'United States of America\', \'lat\': 37.78, \'lon\': -122.42, \'tz_id\': \'America/Los_Angeles\', \'localtime_epoch\': 1717238643, \'localtime\': \'2024-06-01 3:44\'}, \'current\': {\'last_updated_epoch\': 1717237800, \'last_updated\': \'2024-06-01 03:30\', \'temp_c\': 12.0, \'temp_f\': 53.6, \'is_day\': 0, \'condition\': {\'text\': \'Mist\', \'icon\': \'//cdn.weatherapi.com/weather/64x64/night/143.png\', \'code\': 1030}, \'wind_mph\': 5.6, \'wind_kph\': 9.0, \'wind_degree\': 310, \'wind_dir\': \'NW\', \'pressure_mb\': 1013.0, \'pressure_in\': 29.92, \'precip_mm\': 0.0, \'precip_in\': 0.0, \'humidity\': 88, \'cloud\': 100, \'feelslike_c\': 10.5, \'feelslike_f\': 50.8, \'windchill_c\': 9.3, \'windchill_f\': 48.7, \'heatindex_c\': 11.1, \'heatindex_f\': 51.9, \'dewpoint_c\': 8.8, \'dewpoint_f\': 47.8, \'vis_km\': 6.4, \'vis_miles\': 3.0, \'uv\': 1.0, \'gust_mph\': 12.5, \'gust_kph\': 20.1}}"}, {"url": "https://www.timeanddate.com/weather/usa/san-francisco/historic", "content": "Past Weather in San Francisco, California, USA \\u2014 Yesterday and Last 2 Weeks. Time/General. Weather. Time Zone. DST Changes. Sun & Moon. Weather Today Weather Hourly 14 Day Forecast Yesterday/Past Weather Climate (Averages) Currently: 68 \\u00b0F. Passing clouds."}]', name='tavily_search_results_json', tool_call_id='toolu_01BGEyQaSz3pTq8RwUUHSRoo')]}}
----
{'agent': {'messages': [AIMessage(content='Based on the search results, the current weather in San Francisco is:\n\nTemperature: 53.6°F (12°C)\nConditions: Misty\nWind: 5.6 mph (9 kph) from the Northwest\nHumidity: 88%\nCloud Cover: 100% \n\nThe results provide detailed information like wind chill, heat index, visibility and more. It looks like a typical cool, foggy morning in San Francisco. Let me know if you need any other details about the weather where you live!', response_metadata={'id': 'msg_019WGLbaojuNdbCnqac7zaGW', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 1035, 'output_tokens': 120}}, id='run-1bb68bf3-b212-4ef4-8a31-10c830421c78-0', usage_metadata={'input_tokens': 1035, 'output_tokens': 120, 'total_tokens': 1155})]}}
----

环境准备

Jupyter Notebook

本指南（以及文档中的大多数其他指南）使用Jupyter Notebook，建议读者也使用该工具。Jupyter Notebook是学习LLM系统的理想交互式环境，因为LLM应用常出现意外输出或API问题，观察这些案例有助于深入理解构建过程。

安装依赖

%pip install -U langchain-community langgraph langchain-anthropic tavily-python langgraph-checkpoint-sqlite

Tavily API密钥

使用Tavily搜索引擎需要设置API密钥：

export TAVILY_API_KEY="..."

或者，如果在笔记本中，您可以通过以下方式进行设置：

import getpass
import os

os.environ["TAVILY_API_KEY"] = getpass.getpass()

定义工具

我们将使用Tavily搜索引擎作为主要工具：

from langchain_community.tools.tavily_search import TavilySearchResults

search = TavilySearchResults(max_results=2)
search_results = search.invoke("what is the weather in SF")
print(search_results)
# If we want, we can create other tools.
# Once we have all the tools we want, we can put them in a list that we will reference later.
tools = [search]

API参考文档：TavilySearchResults

[{'url': 'https://www.weatherapi.com/',
  'content': "{'location': {'name': 'San Francisco', 'region': 'California', 'country': 'United States of America', 'lat': 37.78, 'lon': -122.42, 'tz_id': 'America/Los_Angeles', 'localtime_epoch': 1717238703, 'localtime': '2024-06-01 3:45'}, 'current': {'last_updated_epoch': 1717237800, 'last_updated': '2024-06-01 03:30', 'temp_c': 12.0, 'temp_f': 53.6, 'is_day': 0, 'condition': {'text': 'Mist', 'icon': '//cdn.weatherapi.com/weather/64x64/night/143.png', 'code': 1030}, 'wind_mph': 5.6, 'wind_kph': 9.0, 'wind_degree': 310, 'wind_dir': 'NW', 'pressure_mb': 1013.0, 'pressure_in': 29.92, 'precip_mm': 0.0, 'precip_in': 0.0, 'humidity': 88, 'cloud': 100, 'feelslike_c': 10.5, 'feelslike_f': 50.8, 'windchill_c': 9.3, 'windchill_f': 48.7, 'heatindex_c': 11.1, 'heatindex_f': 51.9, 'dewpoint_c': 8.8, 'dewpoint_f': 47.8, 'vis_km': 6.4, 'vis_miles': 3.0, 'uv': 1.0, 'gust_mph': 12.5, 'gust_kph': 20.1}}"},
 {'url': 'https://www.wunderground.com/hourly/us/ca/san-francisco/date/2024-01-06',
  'content': 'Current Weather for Popular Cities . San Francisco, CA 58 ° F Partly Cloudy; Manhattan, NY warning 51 ° F Cloudy; Schiller Park, IL (60176) warning 51 ° F Fair; Boston, MA warning 41 ° F ...'}]

使用语言模型

接下来，我们来学习如何通过 to call tools 来使用语言模型。LangChain 支持许多不同的语言模型。

pip install -qU langchain-openai

from langchain_openai import ChatOpenAI

api_key = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
base_url = "https://api.siliconflow.cn/v1/"

model = ChatOpenAI(
    base_url=base_url,
    api_key=api_key,
    model="Pro/deepseek-ai/DeepSeek-V3",
)

您可以通过传入消息列表来调用语言模型。默认情况下，响应是content字符串

from langchain_core.messages import HumanMessage

response = model.invoke([HumanMessage(content="hi!")])
response.content

API 参考文档：HumanMessage

'Hi there!'

我们现在可以看到使这个模型能够进行工具调用是什么感觉。为了实现这一点，我们使用 .bind_tools 为语言模型提供这些工具的知识

model_with_tools = model.bind_tools(tools)

现在，我们可以调用模型。让我们先用 normal 消息调用它，看看它如何响应。我们可以同时查看 content 字段和 tool_calls 字段。

response = model_with_tools.invoke([HumanMessage(content="Hi!")])

print(f"ContentString: {response.content}")
print(f"ToolCalls: {response.tool_calls}")

ContentString: Hello!
ToolCalls: []

我们可以看到现在没有文本内容，但有一个 tool 调用！它希望我们调用 Tavily 搜索工具。

这还不是在调用那个工具 - 它只是告诉我们这样做。为了实际调用它，我们需要创建我们的代理。

创建代理

现在我们已经定义了工具和 LLM，我们可以创建代理。我们将使用 LangGraph 来构建代理。目前我们使用的是高级接口来构造代理，但 LangGraph 的好处是，这个高级接口由一个低级、高度可控的 API 作为后盾，以防你想修改代理逻辑。

现在，我们可以使用 LLM 和工具初始化代理。

请注意，我们传入的是model，而不是 model_with_tools。那是因为 create_react_agent 会在幕后为我们打电话.bind_tools。

from langgraph.prebuilt import create_react_agent

agent_executor = create_react_agent(model, tools)

运行代理

现在，我们可以在几个查询上运行代理！请注意，目前，这些都是无状态查询（它不会记住以前的交互）。请注意，代理将在交互结束时返回最终状态（包括任何输入，我们稍后将了解如何仅获取输出）。

首先，让我们看看当不需要调用工具时它如何响应：

response = agent_executor.invoke({"messages": [HumanMessage(content="hi!")]})

response["messages"]

[HumanMessage(content='hi!', id='a820fcc5-9b87-457a-9af0-f21768143ee3'),
 AIMessage(content='Hello!', response_metadata={'id': 'msg_01VbC493X1VEDyusgttiEr1z', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 264, 'output_tokens': 5}}, id='run-0e0ddae8-a85b-4bd6-947c-c36c857a4698-0', usage_metadata={'input_tokens': 264, 'output_tokens': 5, 'total_tokens': 269})]

现在让我们在一个应该调用该工具的示例上尝试一下

response = agent_executor.invoke(
    {"messages": [HumanMessage(content="whats the weather in sf?")]}
)
response["messages"]

[HumanMessage(content='whats the weather in sf?', id='1d6c96bb-4ddb-415c-a579-a07d5264de0d'),
 AIMessage(content=[{'id': 'toolu_01Y5EK4bw2LqsQXeaUv8iueF', 'input': {'query': 'weather in san francisco'}, 'name': 'tavily_search_results_json', 'type': 'tool_use'}], response_metadata={'id': 'msg_0132wQUcEduJ8UKVVVqwJzM4', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'tool_use', 'stop_sequence': None, 'usage': {'input_tokens': 269, 'output_tokens': 61}}, id='run-26d5e5e8-d4fd-46d2-a197-87b95b10e823-0', tool_calls=[{'name': 'tavily_search_results_json', 'args': {'query': 'weather in san francisco'}, 'id': 'toolu_01Y5EK4bw2LqsQXeaUv8iueF'}], usage_metadata={'input_tokens': 269, 'output_tokens': 61, 'total_tokens': 330}),
 ToolMessage(content='[{"url": "https://www.weatherapi.com/", "content": "{\'location\': {\'name\': \'San Francisco\', \'region\': \'California\', \'country\': \'United States of America\', \'lat\': 37.78, \'lon\': -122.42, \'tz_id\': \'America/Los_Angeles\', \'localtime_epoch\': 1717238703, \'localtime\': \'2024-06-01 3:45\'}, \'current\': {\'last_updated_epoch\': 1717237800, \'last_updated\': \'2024-06-01 03:30\', \'temp_c\': 12.0, \'temp_f\': 53.6, \'is_day\': 0, \'condition\': {\'text\': \'Mist\', \'icon\': \'//cdn.weatherapi.com/weather/64x64/night/143.png\', \'code\': 1030}, \'wind_mph\': 5.6, \'wind_kph\': 9.0, \'wind_degree\': 310, \'wind_dir\': \'NW\', \'pressure_mb\': 1013.0, \'pressure_in\': 29.92, \'precip_mm\': 0.0, \'precip_in\': 0.0, \'humidity\': 88, \'cloud\': 100, \'feelslike_c\': 10.5, \'feelslike_f\': 50.8, \'windchill_c\': 9.3, \'windchill_f\': 48.7, \'heatindex_c\': 11.1, \'heatindex_f\': 51.9, \'dewpoint_c\': 8.8, \'dewpoint_f\': 47.8, \'vis_km\': 6.4, \'vis_miles\': 3.0, \'uv\': 1.0, \'gust_mph\': 12.5, \'gust_kph\': 20.1}}"}, {"url": "https://www.timeanddate.com/weather/usa/san-francisco/hourly", "content": "Sun & Moon. Weather Today Weather Hourly 14 Day Forecast Yesterday/Past Weather Climate (Averages) Currently: 59 \\u00b0F. Passing clouds. (Weather station: San Francisco International Airport, USA). See more current weather."}]', name='tavily_search_results_json', id='37aa1fd9-b232-4a02-bd22-bc5b9b44a22c', tool_call_id='toolu_01Y5EK4bw2LqsQXeaUv8iueF'),
 AIMessage(content='Based on the search results, here is a summary of the current weather in San Francisco:\n\nThe weather in San Francisco is currently misty with a temperature of around 53°F (12°C). There is complete cloud cover and moderate winds from the northwest around 5-9 mph (9-14 km/h). Humidity is high at 88%. Visibility is around 3 miles (6.4 km). \n\nThe results provide an hourly forecast as well as current conditions from a couple different weather sources. Let me know if you need any additional details about the San Francisco weather!', response_metadata={'id': 'msg_01BRX9mrT19nBDdHYtR7wJ92', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 920, 'output_tokens': 132}}, id='run-d0325583-3ddc-4432-b2b2-d023eb97660f-0', usage_metadata={'input_tokens': 920, 'output_tokens': 132, 'total_tokens': 1052})]

流式处理消息

除了流回消息之外，流回令牌也很有用。我们可以使用 .astream_events 方法来实现这一点。

提示

这种.astream_events方法仅适用于 Python 3.11 或更高版本。

async for event in agent_executor.astream_events(
    {"messages": [HumanMessage(content="whats the weather in sf?")]}, version="v1"
):
    kind = event["event"]
    if kind == "on_chain_start":
        if (
            event["name"] == "Agent"
        ):  # Was assigned when creating the agent with `.with_config({"run_name": "Agent"})`
            print(
                f"Starting agent: {event['name']} with input: {event['data'].get('input')}"
            )
    elif kind == "on_chain_end":
        if (
            event["name"] == "Agent"
        ):  # Was assigned when creating the agent with `.with_config({"run_name": "Agent"})`
            print()
            print("--")
            print(
                f"Done agent: {event['name']} with output: {event['data'].get('output')['output']}"
            )
    if kind == "on_chat_model_stream":
        content = event["data"]["chunk"].content
        if content:
            # Empty content in the context of OpenAI means
            # that the model is asking for a tool to be invoked.
            # So we only print non-empty content
            print(content, end="|")
    elif kind == "on_tool_start":
        print("--")
        print(
            f"Starting tool: {event['name']} with inputs: {event['data'].get('input')}"
        )
    elif kind == "on_tool_end":
        print(f"Done tool: {event['name']}")
        print(f"Tool output was: {event['data'].get('output')}")
        print("--")

--
Starting tool: tavily_search_results_json with inputs: {'query': 'current weather in San Francisco'}
Done tool: tavily_search_results_json
Tool output was: [{'url': 'https://www.weatherapi.com/', 'content': "{'location': {'name': 'San Francisco', 'region': 'California', 'country': 'United States of America', 'lat': 37.78, 'lon': -122.42, 'tz_id': 'America/Los_Angeles', 'localtime_epoch': 1714427052, 'localtime': '2024-04-29 14:44'}, 'current': {'last_updated_epoch': 1714426200, 'last_updated': '2024-04-29 14:30', 'temp_c': 17.8, 'temp_f': 64.0, 'is_day': 1, 'condition': {'text': 'Sunny', 'icon': '//cdn.weatherapi.com/weather/64x64/day/113.png', 'code': 1000}, 'wind_mph': 23.0, 'wind_kph': 37.1, 'wind_degree': 290, 'wind_dir': 'WNW', 'pressure_mb': 1019.0, 'pressure_in': 30.09, 'precip_mm': 0.0, 'precip_in': 0.0, 'humidity': 50, 'cloud': 0, 'feelslike_c': 17.8, 'feelslike_f': 64.0, 'vis_km': 16.0, 'vis_miles': 9.0, 'uv': 5.0, 'gust_mph': 27.5, 'gust_kph': 44.3}}"}, {'url': 'https://www.weathertab.com/en/c/e/04/united-states/california/san-francisco/', 'content': 'San Francisco Weather Forecast for Apr 2024 - Risk of Rain Graph. Rain Risk Graph: Monthly Overview. Bar heights indicate rain risk percentages. Yellow bars mark low-risk days, while black and grey bars signal higher risks. Grey-yellow bars act as buffers, advising to keep at least one day clear from the riskier grey and black days, guiding ...'}]
--
The| current| weather| in| San| Francisco|,| California|,| USA| is| sunny| with| a| temperature| of| |17|.|8|°C| (|64|.|0|°F|).| The| wind| is| blowing| from| the| W|NW| at| a| speed| of| |37|.|1| k|ph| (|23|.|0| mph|).| The| humidity| level| is| at| |50|%.| [|Source|](|https|://|www|.weather|api|.com|/)|

添加内存

如前所述，此代理是无状态的。这意味着它不会记住以前的交互。为了给它内存，我们需要传入一个 checkpointer。当传入一个 checkpointer 时，我们还必须在调用 agent 时传入一个 thread_id（这样它就知道要从哪个线程 / conversation 恢复）

如前所述，此代理是无状态的。这意味着它不会记住以前的交互。为了给它内存，我们需要传入一个 checkpointer。当传入一个 checkpointer 时，我们还必须在调用 agent 时传入一个 thread_id（这样它就知道要从哪个线程 / conversation 恢复）

agent_executor = create_react_agent(model, tools, checkpointer=memory)

config = {"configurable": {"thread_id": "abc123"}}

for chunk in agent_executor.stream(
    {"messages": [HumanMessage(content="hi im bob!")]}, config
):
    print(chunk)
    print("----")

{'agent': {'messages': [AIMessage(content="Hello Bob! It's nice to meet you again.", response_metadata={'id': 'msg_013C1z2ZySagEFwmU1EsysR2', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 1162, 'output_tokens': 14}}, id='run-f878acfd-d195-44e8-9166-e2796317e3f8-0', usage_metadata={'input_tokens': 1162, 'output_tokens': 14, 'total_tokens': 1176})]}}
----

for chunk in agent_executor.stream(
    {"messages": [HumanMessage(content="whats my name?")]}, config
):
    print(chunk)
    print("----")

{'agent': {'messages': [AIMessage(content='You mentioned your name is Bob when you introduced yourself earlier. So your name is Bob.', response_metadata={'id': 'msg_01WNwnRNGwGDRw6vRdivt6i1', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 1184, 'output_tokens': 21}}, id='run-f5c0b957-8878-405a-9d4b-a7cd38efe81f-0', usage_metadata={'input_tokens': 1184, 'output_tokens': 21, 'total_tokens': 1205})]}}
----

如果我想开始新的对话，我所要做的就是更改使用的thread_id

config = {"configurable": {"thread_id": "xyz123"}}
for chunk in agent_executor.stream(
    {"messages": [HumanMessage(content="whats my name?")]}, config
):
    print(chunk)
    print("----")

{'agent': {'messages': [AIMessage(content="I'm afraid I don't actually know your name. As an AI assistant without personal information about you, I don't have a specific name associated with our conversation.", response_metadata={'id': 'msg_01NoaXNNYZKSoBncPcLkdcbo', 'model': 'claude-3-sonnet-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 267, 'output_tokens': 36}}, id='run-c9f7df3d-525a-4d8f-bbcf-a5b4a5d2e4b0-0', usage_metadata={'input_tokens': 267, 'output_tokens': 36, 'total_tokens': 303})]}}
----

到此结束！在本快速入门中，我们介绍了如何创建简单的代理。然后，我们展示了如何流回响应 - 不仅是中间步骤，还有令牌！我们还添加了内存，以便您可以与他们进行对话。代理是一个复杂的话题，有很多东西要学！

有关 Agent 的更多信息，请查看 LangGraph 文档。它有自己的一组概念、教程和作指南。