跳到主要内容

WTF Langchain极简入门: 08. 代理 (Agent)

最近在学习Langchain框架，顺手写一个“WTF Langchain极简入门”，供小白们使用（编程大佬可以另找教程）。本教程默认以下前提：

  • 使用Python版本的Langchain
  • LLM使用OpenAI的模型
  • Langchain目前还处于快速发展阶段，版本迭代频繁，为避免示例代码失效，本教程统一使用版本 0.0.235

根据Langchain的代码约定，Python版本 ">=3.8.1,<4.0"。

推特：@verysmallwoods

所有代码和教程开源在github: github.com/sugarforever/wtf-langchain

简介​

Agent 也就是 代理，它的核心思想是利用一个语言模型来选择一系列要执行的动作。LangChain 的链将一系列的动作硬编码在代码中。而在 Agent 中，语言模型被用作推理引擎，来确定应该执行哪些动作以及以何种顺序执行。

这就涉及到几个关键组件：

  • Agent 代理
  • Tool 工具
  • Toolkit 工具包
  • AgentExecutor 代理执行器

接下来我们做逐一介绍。注，该极简入门系列将略过工具包的介绍，这部分内容将包含在进阶系列中。

Agent​

Agent 由一个语言模型和一个提示词驱动，决定下一步要采取什么措施的类。提示词可以包括以下内容：

  • 代理的个性（用于使其以特定方式回应）
  • 代理的背景（用于为其提供更多关于所要执行任务类型的上下文信息）
  • 引导策略（用于激发更好的推理能力）

LangChain 提供了不同类型的代理：

  • Zero-shot ReAct

    利用 ReAct 框架根据工具的描述来决定使用哪个工具，可以使用多个工具，但需要为每个工具提供描述信息。工具的选择单纯依靠工具的描述信息。关于 ReAct 框架的更多信息，请参考 ReAct。

  • Structured Input ReAct

    相较于单一字符串作为输入的代理，该类型的代理可以通过工具的参数schema创建结构化的动作输入。

  • OpenAI Functions

    该类型的代理用来与OpenAI Function Call机制配合工作。

  • Conversational

    这类代理专为对话场景设计，使用具有对话性的提示词，利用 ReAct 框架选择工具，并利用记忆功能来保存对话历史。

  • Self ask with search

    这类代理利用工具查找问题的事实性答案。

  • ReAct document store

    利用 ReAct 框架与文档存储进行交互，使用时需要提供 Search 工具和 Lookup 工具，分别用于搜索文档和在最近找到的文档中查找术语。

  • Plan-and-execute agents

    代理规划要做的事情，然后执行子任务来达到目标。

这里我们多次提到 “工具”，也就是 Tool，接下来我们就介绍什么是 Tool。

Tool​

Tool 工具，是代理调用的功能，通常用来与外部世界交互，比如维基百科搜索，资料库访问等。LangChain 内置的工具列表，请参考 Tools。

Toolkit​

通常，在达成特定目标时，需要使用一组工具。LangChain 提供了 Toolkit 工具包的概念，将多个工具组合在一起。

AgentExecutor​

代理执行器是代理的运行时。程序运行中，由它来调用代理并执行其选择的动作。

组件实例​

Tool​

LangChain 提供了一系列工具，比如 Search 工具，AWS 工具，Wikipedia 工具等。这些工具都是 BaseTool 的子类。通过调用 run 函数，执行工具的功能。

我们以 LangChain 内置的工具 DuckDuckGoSearchRun 为例，来看看如何使用工具。

注，要使用DuckDuckGoSearchRun工具，需要安装以下python包:

pip install duckduckgo-search

  1. 通过工具类创建工具实例

    该类提供了通过 DuckDuckGo 搜索引擎搜索的功能。

    from langchain.tools import DuckDuckGoSearchRun
    search = DuckDuckGoSearchRun()
    search.run("Who is winner of FIFA worldcup 2018?")

    你应该期望如下输出：

    The 2018 FIFA World Cup was the 21st FIFA World Cup, ... Mario Zagallo (Brazil) and Franz Beckenbauer (Germany) have also achieved the feat.

    注，限于篇幅，这里对模型的回答文本在本讲中做了截取。

  2. 通过辅助函数 load_tools 加载

    LangChain 提供了函数 load_tools 基于工具名称加载工具。

    先来看看DuckDuckGoSearchRun类的定义：

    class DuckDuckGoSearchRun(BaseTool):
        """Tool that adds the capability to query the DuckDuckGo search API."""

        name = "duckduckgo_search"
        description = (
            "A wrapper around DuckDuckGo Search. "
            "Useful for when you need to answer questions about current events. "
            "Input should be a search query."
        )

    name 变量定义了工具的名称。这正是我们使用 load_tools 函数加载工具时所需要的。当然，目前比较棘手的是，load_tools 的实现对工具名称做了映射，因此并不是所有工具都如实使用工具类中定义的 name。比如，DuckDuckGoSearchRun 的名称是 duckduckgo_search，但是 load_tools 函数需要使用 ddg-search 来加载该工具。

    请参考源代码 load_tools.py 了解工具数据初始化的详情。

    用法

    from langchain.agents import load_tools

    tools = load_tools(['ddg-search'])
    search = tools[0]
    search.run("Who is winner of FIFA worldcup 2018?")

    你应该期望与方法1类似的输出。

    最后，分享一个辅助函数 get_all_tool_names，用于获取所有工具的名称。

    from langchain.agents import get_all_tool_names
    get_all_tool_names()

    当前 LangChain 版本 0.0.235 中，我们应该能看到如下列表：

       ['python_repl',
        'requests',
        'requests_get',
        'requests_post',
        'requests_patch',
        'requests_put',
        'requests_delete',
        'terminal',
        'sleep',
        'wolfram-alpha',
        'google-search',
        'google-search-results-json',
        'searx-search-results-json',
        'bing-search',
        'metaphor-search',
        'ddg-search',
        'google-serper',
        'google-serper-results-json',
        'serpapi',
        'twilio',
        'searx-search',
        'wikipedia',
        'arxiv',
        'pupmed',
        'human',
        'awslambda',
        'sceneXplain',
        'graphql',
        'openweathermap-api',
        'dataforseo-api-search',
        'dataforseo-api-search-json',
        'news-api',
        'tmdb-api',
        'podcast-api',
        'pal-math',
        'pal-colored-objects',
        'llm-math',
        'open-meteo-api']

Agent​

Agent 通常需要 Tool 配合工作，因此我们将 Agent 实例放在 Tool 之后。我们以 Zero-shot ReAct 类型的 Agent 为例，来看看如何使用。代码如下：

from langchain.agents import load_tools
from langchain.agents import initialize_agent
from langchain.agents import AgentType
from langchain.llms import OpenAI

import os
os.environ['OPENAI_API_KEY'] = "您的有效openai api key"

llm = OpenAI(temperature=0)
tools = load_tools(["ddg-search", "llm-math"], llm=llm)
agent = initialize_agent(tools, llm, agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True)
agent.run("What is the height difference between Eiffel Tower and Taiwan 101 Tower?")

代码解释：

  1. 设置环境变量 OPENAI_API_KEY 并实例化 OpenAI 语言模型，用于后续的推理。
  2. 通过load_tools加载 DuckDuckGo 搜索工具和 llm-math 工具。
  3. 通过 initialize_agent 函数初始化代理执行器，指定代理类型为 ZERO_SHOT_REACT_DESCRIPTION，并打开 verbose 模式，用于输出调试信息。
  4. 通过 run 函数运行代理。

参考 initialize_agent 的实现，我们会看到它返回的是 AgentExecutor 类型的实例。这也是代理执行器的常见用法。请前往源代码 initialize.py 了解更多初始化代理执行器的详情。

def initialize_agent(
    tools: Sequence[BaseTool],
    llm: BaseLanguageModel,
    agent: Optional[AgentType] = None,
    callback_manager: Optional[BaseCallbackManager] = None,
    agent_path: Optional[str] = None,
    agent_kwargs: Optional[dict] = None,
    *,
    tags: Optional[Sequence[str]] = None,
    **kwargs: Any,
) -> AgentExecutor:
    """Load an agent executor given tools and LLM.

你应该期望如下输出：

> Entering new AgentExecutor chain...
 I need to find out the heights of both towers.
Action: duckduckgo_search
Action Input: "Eiffel Tower height"
Observation: It was the first structure in the world to surpass both the 200-metre and 300-metre mark in height. Due to the addition of a broadcasting aerial at the top of the tower in 1957, it is now taller than the Chrysler Building by 5.2 metres (17 ft). Eiffel's concept of a 300-metre (984-foot) tower built almost entirely of open-lattice wrought iron aroused amazement, skepticism, and no little opposition on aesthetic grounds. When completed, the tower served as the entrance gateway to the exposition. Built for the 1889 World's Fair in Paris, the Eiffel Tower is a 1,000-foot tall wrought iron tower, considered an architectural wonder and one of the world's most recognizable structures. The Eiffel Tower reached the height of 330 meters (1,083 feet) on March 15, 2022 thanks to this successful and dizzying operation carried out by helicopter. Which was a first ! Browse our resources on the links between the Tower and radio. Discover how did radio save the Tower. All about the Tower's Scientific Uses Height of the Eiffel Tower. The main structure of the is Eiffel Tower is 300 meters tall, although the height including antennas is 324 meters. This height is roughly equivalent to that of an 81-storey building. Interestingly, the Eiffel Tower can shrink by 6 inches during cold temperatures. The tower has a square base that measures 125 meters ...
Thought: I need to find out the height of Taiwan 101 Tower.
Action: duckduckgo_search
Action Input: "Taiwan 101 Tower height"
Observation: Designed by C.Y. Lee & Partners, a local architectural firm, the skyscraper has 101 stories and reaches a height, including the spire, of 1,667 feet (508 metres). At the time of its official opening in October 2004, it was the world's tallest building, having surpassed the Petronas Twin Towers in Kuala Lumpur, Malaysia. Taipei 101 is the tallest building in Taiwan. The elevators of Taipei 101 that transport passengers from the 5th to the 89th floor in 37 seconds (attaining 60.6 km/h ... The height of 101 floors commemorates the renewal of time: the new century that arrived as the tower was built (100+1) and all the new years that follow (1 January = 1-01 ... Last updated: Sep 1, 2021 • 3 min read Taipei 101, located in Taiwan, is one of the tallest high-rise buildings in the world. Visitors can explore the building that includes a large shopping mall, world-class restaurants, and an outdoor observation deck. Learn From the Best Food Design & Style Arts & Entertainment Music Business Sports & Gaming With a height of 1667 feet or 508 meters, Taipei 101 has the distinction of being the third tallest building in the world, and tallest outside the Middle East. Its spire adds to its height, and the total makes it higher than the Kuala Lumpur Petronas Twin Towers. The Taipei 101 tower in Taipei, Taiwan, was the world's tallest building from 2004 until 2010 when it was beaten out by Dubai's impressive Burj Khalifa. Regardless, Taipei 101 is still considered the tallest green building in the world for its innovative and energy-saving design. Even the 2015-2016 New Year's Eve firework's show was nature ...
Thought: I now know the heights of both towers, so I can calculate the difference.
Action: Calculator
Action Input: 330 - 508
Observation: Answer: -178
Thought: I now know the final answer.
Final Answer: The height difference between Eiffel Tower and Taiwan 101 Tower is 178 meters.

> Finished chain.

'The height difference between Eiffel Tower and Taiwan 101 Tower is 178 meters.'

总结​

本节课程中，我们学习了什么是 Agent 代理，Tool 工具，以及 AgentExecutor 代理执行器，并学习了它们的基本用法。下一讲我们将学习 Callback 回调。

本节课程的完整示例代码，请参考 08_Agents.ipynb。

相关文档资料链接：​

  1. Python Langchain官方文档