基于MetaGPT的LLM Agent学习实战(一)

前言

我最近一直在做基于AI Agent 的个人项目， 因为工作加班较多，设计思考时间不足，这里借着Datawhale的开源学习课程《MetaGPT智能体理论与实战》课程，来完善自己的思路，抛砖引玉，和各位开发者一起学习！

一、介绍

今天是打卡的第一天，先说说主要的学习内容:

• 获取MetaGPT
  • 部署到本地环境
• 配置MetaGPT
  • 申请ChatGPT API Key
  • 基于ChatGPT API构建调用代码
• 运行MetaGPT案例代码进行测试

今天学习的内容较为简单，我会尽量以简洁的语言详细描述清楚这个流程，带着读者一起学习Agent开发；

二、配置MetaGPT运行环境

声明

• python版本为3.9+
• 为了方便学习，这里我使用jupyter notebook进行讲解；
• 所有代码我都会同步提交到Github和Gitee，如果各位读者觉得我写的不错，可以给我一个Star.

1. 查看Python版本

为了确保我们的Python环境正确，首先要检查Python的版本。可以使用以下命令来查看Python版本：

!python3 --version

如果上面的命令不起作用或者报错，可以尝试使用以下命令：

python --version

输出

Python 3.10.13

2. 安装MetaGPT

要安装MetaGPT，我们可以使用pip来获取它。以下是在终端中安装MetaGPT的命令：

pip install metagpt==0.6.6

如果你在国内环境，并且希望加速安装过程，可以使用清华源进行按照：

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple metagpt==0.6.6

也可以通过拉取官方仓库进行安装：

git clone https://github.com/geekan/MetaGPT.git
cd /your/path/to/MetaGPT
pip install -e .

这里有个重点，如果你的OpenAI API Key是直连且不限速版本，你只需要安装包即可，
如果你的API Key为免费API且有速率限制，我这里建议你直接clone MetaGPT的GitHub仓库，其可以在config2.yaml中自定义配置代理服务器和Key，我在运行MetaGPT的过程中遇到的最大问题就是API限速导致程序报错；所以一定要注意这一点；

作者因为使用的是中转的API Key，因此选择了方法3：

git clone https://github.com/geekan/MetaGPT.git
cd MetaGPT
pip install -e .

我们在config/config2.yaml中配置自己的api key和 base_url 以及选择的model:

llm:api_type: "openai"  # or azure / ollama / groq etc.model: "gpt-4-turbo"  # or gpt-3.5-turbobase_url: "https://api.openai.com/v1"  # or forward url / other llm urlapi_key: "YOUR_API_KEY"

3. 配置MetaGPT

为了配置MetaGPT，你需要调用ChatGPT API服务。你可以在这里查看具体配置方式。如果你没有科学环境，也可以通过去tb buy 一个 中转的 API Key来实现。我们主要介绍官方申请方法：

中转方案修改的部分我在代码中也已经标出

① 登录自己的账号

②创建API Key


③本地配置环境变量

import os
os.environ["OPENAI_API_KEY"] = "sk-..."  # 填入你自己的OpenAI API key
os.environ["OPENAI_API_MODEL"] = "gpt-3.5-turbo" # 选择你要使用的模型，例如：gpt-4, gpt-3.5-turbo
os.environ["OPENAI_API_BASE"] = "https://api.openai-forward.com/v1" # 调整API请求地址，设置访问中转代理服务器,如果是商家购买的，可以联系商家要代理服务器地址，这里并不是固定的

④验证配置是否成功：

from openai import OpenAI# client = OpenAI(api_key='sk-......'） # 官网直连版本
client = OpenAI(base_url="https://xxxx.com", # 这里填写你的中转服务器地址api_key='sk-......') # 这里填写你的中转apikey
completion = client.chat.completions.create(model="gpt-3.5-turbo",messages=[{"role": "system", "content": "你是一个WebGIS开发者，测绘地理和全栈开发精通."},{"role": "user", "content": "聊聊国内外WebGIS开发与AI结合的场景现在我们可以聊聊国内外WebGIS开发与AI LLM Agent结合的场景吧"}]
)
print(completion.choices[0].message.content)

运行结果如下:

🎉🎉🎉bingo！！运行成功，我们成功拿到了我们要的方案！

通过以上步骤，我们终于成功配置MetaGPT，并开始使用它进行各种任务了。

三. 使用MetaGPT

接下来，我们通过下面这个案例，我们用以验证环境配置是否成功，并初次体验多智能体框架中的指令 - 动作 - 角色 - 环境 - 团队的抽象概念。在这个示例中，我们创建了一个团队，其中包括产品经理、架构师、项目经理和工程师。然后，我们投资并运行一个项目，最后让团队运行五轮。

import asyncio
from metagpt.roles import (Architect,Engineer,ProductManager,ProjectManager,
)
from metagpt.team import Teamasync def startup(idea: str):company = Team()company.hire([ProductManager(),Architect(),ProjectManager(),Engineer(),])company.invest(investment=3.0)company.run_project(idea=idea)await company.run(n_round=5)await startup(idea="write a cli blackjack game")

这里我copy了其中几轮Agent的回答，可以看到我们的AI团队已经运行起来了；

[CONTENT]
{"Language": "en_us","Programming Language": "Python","Original Requirements": "write a cli blackjack game","Project Name": "cli_blackjack_game","Product Goals": ["Create an engaging and interactive gameplay experience","Ensure smooth and intuitive user interface for seamless gameplay","Implement various difficulty levels to cater to different player skills"],"User Stories": ["As a player, I want to be able to start a new game easily","As a player, I want to see my current score and progress during the game","As a player, I want to have options to hit, stand, or double down during my turn","As a player, I want to receive clear instructions on how to play the game","As a player, I want to feel the excitement and challenge of a real blackjack game"],"Competitive Analysis": ["Blackjack Game A: Basic interface, lacks interactive features","Blackjack Pro: Offers advanced gameplay options and strategy guides","Blackjack Master: Provides a realistic casino experience with multiplayer mode"],"Competitive Quadrant Chart": "quadrantChart\n    title \"Engagement and User Experience\"\n    x-axis \"Low Engagement\" --> \"High Engagement\"\n    y-axis \"Low User Experience\" --> \"High User Experience\"\n    quadrant-1 \"Enhance Features\"\n    quadrant-2 \"Improve User Experience\"\n    quadrant-3 \"Optimize Engagement\"\n    quadrant-4 \"Maximize User Satisfaction\"\n    \"Blackjack Game A\": [0.3, 0.4]\n    \"Blackjack Pro\": [0.6, 0.7]\n    \"Blackjack Master\": [0.8, 0.9]\n    \"Our CLI Blackjack Game\": [0.5, 0.6]","Requirement Analysis": "","Requirement Pool": [["P0","Implement basic game logic for blackjack"],["P1","Create a scoring system to track player progress"],["P2","Develop a user-friendly interface for easy navigation"],["P2","Incorporate different difficulty levels for player choice"],["P1","Include clear instructions on how to play the game"]],"UI Design draft": "The UI will include options for hitting, standing, and doubling down. It will display the player's current score and provide clear instructions for gameplay.","Anything
2024-05-12 17:36:48.720 | ERROR    | metagpt.utils.common:log_it:554 - Finished call to 'metagpt.actions.action_node.ActionNode._aask_v1' after 10.724(s), this was the 1st time calling it. exp: openai.types.completion_usage.CompletionUsage() argument after ** must be a mapping, not NoneTypeUNCLEAR": ""
}
[/CONTENT][CONTENT]
{"Language": "en_us","Programming Language": "Python","Original Requirements": "write a cli blackjack game","Project Name": "cli_blackjack_game","Product Goals": ["Create an engaging CLI experience for users","Ensure smooth gameplay and fair card dealing logic","Provide an enjoyable and interactive blackjack game"],"User Stories": ["As a player, I want to be able to place bets and receive cards","As a player, I want to have options like hit, stand, double down","As a player, I want to see my current balance and game outcome"],"Competitive Analysis": ["Blackjack Game A: Basic CLI interface, lacks interactive features","cli-blackjack.io: Offers various betting options and clear game instructions","blackjack-cli.com: Provides realistic card dealing but lacks betting flexibility"],"Competitive Quadrant Chart": "quadrantChart\n    title \"Engagement and User Experience\"\n    x-axis \"Low Engagement\" --> \"High Engagement\"\n    y-axis \"Low User Experience\" --> \"High User Experience\"\n    quadrant-1 \"Enhance Features\"\n    quadrant-2 \"Improve User Experience\"\n    quadrant-3 \"Optimize Engagement\"\n    quadrant-4 \"Maintain Quality\"\n    \"Blackjack Game A\": [0.3, 0.6]\n    \"cli-blackjack.io\": [0.45, 0.23]\n    \"blackjack-cli.com\": [0.57, 0.69]\n    \"Our CLI Blackjack Game\": [0.5, 0.6]","Requirement Analysis": "","Requirement Pool": [["P0","Implement card dealing and betting system"],["P1","Include game logic for hit, stand, and double down actions"],["P2","Display player balance and game outcomes"]],"UI Design draft": "Simple text-based interface with clear instructions and game status
2024-05-12 17:36:57.136 | ERROR    | metagpt.utils.common:log_it:554 - Finished call to 'metagpt.actions.action_node.ActionNode._aask_v1' after 19.140(s), this was the 2nd time calling it. exp: openai.types.completion_usage.CompletionUsage() argument after ** must be a mapping, not NoneTypeupdates.","Anything UNCLEAR": ""
}
[/CONTENT]

通过以上步骤，我们可以开始使用MetaGPT进行各种任务，并看到AI Agent的强大潜力！

四、总结

本文是这个打卡系列的第一篇文章，也是后续学习的基础，通过这篇文章，我们了解了MetaGPT开发的基础环境配置方法，在下一篇文章中，我们将深入理解AI Agent的理论，并通过代码来实现Agent的每个模块需求，希望我的文章对各位读者和开发者有所帮助！期待后续学习！！😀😀😀

参考文献

• MetaGPT 官方文档

项目地址

• Github地址

---

如果觉得我的文章对您有帮助，三连+关注便是对我创作的最大鼓励！或者一个star🌟也可以😂.