AutoGen多代理对话项目示例和工作流程分析

本文涉及的产品
实时数仓Hologres,5000CU*H 100GB 3个月
智能开放搜索 OpenSearch行业算法版,1GB 20LCU 1个月
实时计算 Flink 版,1000CU*H 3个月
简介: 在这篇文章中,我将介绍AutoGen的多个代理的运行。这些代理将能够相互对话,协作评估股票价格,并使用AmCharts生成图表。

我们创建对话的目的是要求代理分析特定公司的股票价格,并制作股票价格图表。

为了实现这一目标,我们创建一下代理并协同工作:

金融分析师:分析师的任务是获取股票价格数据,进行分析,然后将数据传递给UI设计人员以创建图表。它还负责执行UI设计器的代码来生成和显示图表,并且可以在必要时请求丢失的数据。

软件工程师:软件工程师的主要角色是检索由财务分析师指定的所需天数的股票价格信息的函数。

UI设计师:UI设计师的主要职责是使用Amcharts股票图表库创建股票图表。这包括生成完整的代码,无缝地集成金融分析师提供的股票价格数据,并准备立即执行的代码。

AutoGen

AutoGen是一个框架,它通过使用可以彼此对话的多个代理来共同处理各种任务,从而促进大型语言模型(LLM)应用程序的创建。

AssistantAgent是专门设计用来作为人工智能助手来解决LLM任务的。

UserProxyAgent在每个交互回合中主要寻求人工输入作为其默认响应时进行调用。它还具有执行代码和调用函数的能力。如果不提供人工用户输入,该代理将在识别传入消息中的可执行代码块并自动启动代码执行。可以通过将code_execution_config参数配置为False来禁用代码执行。可以理解为它是代理的代理,可以执行代码并向其他代理提供反馈。

代码样板

首先,我们需要安装AutoGen库。

 !pip install pyautogen~=0.1.0

创建配置文件OAI_CONFIG_LIST。

 [
     {
         "model": "gpt-4-32k",
         "api_key": "",
         "api_base": "",
         "api_type": "azure",
         "api_version": "2023-07-01-preview"
     }
 ]

加载配置文件

 config_list = autogen.config_list_from_json(
     "OAI_CONFIG_LIST",
     filter_dict={
         "model": ["gpt-4", "gpt-4-0314", "gpt4", "gpt-4-32k", "gpt-4-32k-0314", "gpt-4-32k-v0314"],
     },
 )

创建is_termination_msg函数来检查是否有终止消息来结束对话。

 def is_termination_msg(data):
     has_content = "content" in data and data["content"] is not None
     return has_content and "TERMINATE" in data["content"]

创建函数fetch_prices_for_symbol和fetch_prices检索指定时间范围内的历史价格数据,输入参数symbol和days。

 def fetch_prices_for_symbol(symbol: str, days: int) -> pd.DataFrame:
     # Define the date range for historical data
     end_date = pd.Timestamp.now()
     start_date = end_date - pd.DateOffset(days=days)

     # Fetch historical data using yfinance
     data = yf.download(symbol, start=start_date, end=end_date)

     # Reset the index and select relevant columns
     data.reset_index(inplace=True)
     data = data[["Date", "Open", "High", "Low", "Close", "Volume"]]
     data.rename(columns={"Date": "date", "Open": "open", "High": "high", "Low": "low", "Close": "close", "Volume": "volume"}, inplace=True)

     return data

 def fetch_prices(days: int, symbol: str) -> str:
     symbol_df = fetch_prices_for_symbol(symbol, days)

     symbol_txt = symbol_df.to_string(index=None)    
     return f"""

 {symbol_txt}
     """.strip()

通过llm_config定义LLM推理配置。这里还指定了函数fetch_prices和所需的输入参数days和symbol。

 llm_config = {
     "config_list": config_list,
     "use_cache": False,
     "temperature": 0,
     "request_timeout": 120,
     "functions": [
         {
             "name": "fetch_prices",
             "description": "Fetch daily stock prices",
             "parameters": {
                 "type": "object",
                 "properties": {
                     "days": {
                         "type": "integer",
                         "description": "days"
                     },
                     "symbol": {
                         "type": "string",
                         "description": "symbol"
                     }
                 },
                 "required": ["days", "symbol"]
             }
         }
     ]
 }

代理配置

各个代理角色的配置主要通过Prompt来进行指定:

1、金融分析师

 analyst_system_message = f"""

 Financial Analyst: As a Senior Financial Analyst responsible for a stock price indicator.

 Follow the plan:
 1. Obtain stock price data from the engineer.
 2. Analyze the stock price data, provide comment, and pass it to the UI designer for chart creation.
 3. Upon receiving the code from the UI designer along with the complete dataset, if it's not provided, kindly request the UI designer to provide the missing data.
 4. Execute the code received from the UI designer on your local machine to generate and display the chart.

 Upon successful completion of the plan, add "TERMINATE" to conclude the message.

 """

 analyst = AssistantAgent(
     name = "analyst",
     system_message = analyst_system_message,
     llm_config=llm_config,
     is_termination_msg=is_termination_msg,
     code_execution_config=False
 )

2、软件工程师

 engineer_system_message = f"""

 Engineer. You are a Senior Software Engineer that executes the fetch_prices functions as requested by the Financial Analyst.

 """

 engineer = AssistantAgent(
     name="engineer",
     system_message=engineer_system_message,
     llm_config=llm_config,
     function_map={"fetch_prices": fetch_prices},
     code_execution_config=False
 )

3、UI设计师

 uidesigner_system_message = f"""

 UI Designer: You are a Senior UI/UX designer with a specialization in crafting charts using the Amcharts Stock Chart library (referenced at https://www.amcharts.com/demos/stock-chart/), 
 your assignment is to develop a stock chart using the stock price data supplied by the Senior Financial Analyst.

 Your responsibility involves:
 1. Generating comprehensive code, including the actual COMPLETE data, for the Senior Analyst to successfully run the chart.
 2. Ensuring that the code is prepared for immediate execution, WITHOUT ANY PLACEHOLDERS TEXT or missing data.
 3. Guaranteeing the seamless integration of the received stock price data into the code.
 4. Structuring the code for execution and utilizing the Amcharts library to present the stock chart.

 """

 uidesigner = AssistantAgent(
     name = "uidesigner",
     system_message=uidesigner_system_message,
     code_execution_config=False,  # set to True or image name like "python:3" to use docker
     llm_config=llm_config
 )

4、UserProxyAgent,最后我们在定义一个人工的管理员,这样可以在需要进行人工干预时进行处理

 user_proxy = UserProxyAgent(
     name="admin",
     system_message="Human Admin: Let's engage with the analyst to have a discussion about the stock price report. Please ensure that the final report is shared with me for approval before it's finalized.",
     code_execution_config=False,  # set to True or image name like "python:3" to use docker
     human_input_mode="NEVER",
     is_termination_msg=is_termination_msg
 )

最后我们把这些角色加入到讨论群组中,并且指定讨论的轮数:

 groupchat = autogen.GroupChat(
     agents=[user_proxy, analyst, uidesigner, engineer], messages=[], max_round=20
 )
 manager=autogen.GroupChatManager(groupchat=groupchat, llm_config=llm_config)

最后就是调用

 message = f"""
 Analyze stock price for GRAB for the last 30 days and create a chart.
 """

 user_proxy.initiate_chat(manager, clear_history=True, message=message)

日志解析

我们来看看并分析他们的对话,这样可以了解到工作的具体细节

对话由人工发起,也就是我们最后调用代码的prompt

 admin (to chat_manager):
 Analyze stock price for GRAB for the last 30 days and create a chart.

聊天管理器向分析师发送了一条消息,要求检索过去30天的Grab股票数据,并向聊天管理器提供输入参数。随后,聊天管理器将此信息传达给工程师。

 analyst (to chat_manager):

 ***** Suggested function Call: fetch_prices *****
 Arguments: 

 {
   "days": 30,
   "symbol": "GRAB"
 }

在收到来自聊天管理器的指令后,工程师继续执行该函数,获取输出,然后通知聊天管理器。

 >>>>>>>> EXECUTING FUNCTION fetch_prices...
 [*********************100%%**********************]  1 of 1 completed
 engineer (to chat_manager):

 ***** Response from calling function "fetch_prices" *****
 date  open  high   low  close    volume
 2023-09-20  3.50 3.545 3.410   3.46  13639700
 2023-09-21  3.45 3.460 3.380   3.46  13781400
 2023-09-22  3.48 3.510 3.425   3.43   8708200
 2023-09-25  3.43 3.470 3.385   3.43  10990000
 2023-09-26  3.38 3.450 3.360   3.40  11458800
 2023-09-27  3.40 3.460 3.360   3.38  20921100
 2023-09-28  3.38 3.490 3.360   3.47  14482300
 2023-09-29  3.52 3.590 3.520   3.54  15580200
 2023-10-02  3.49 3.570 3.460   3.48  19845600
 2023-10-03  3.48 3.480 3.330   3.35  27370300
 2023-10-04  3.32 3.410 3.230   3.40  18224300
 2023-10-05  3.37 3.500 3.370   3.48  16904000
 2023-10-06  3.45 3.520 3.445   3.50   9386900
 2023-10-09  3.47 3.480 3.380   3.44  13512600
 2023-10-10  3.51 3.570 3.500   3.53  11492700
 2023-10-11  3.51 3.550 3.480   3.49  10962200
 2023-10-12  3.54 3.560 3.410   3.47  13216900
 2023-10-13  3.44 3.460 3.360   3.42  25474100
 2023-10-16  3.40 3.460 3.370   3.42  22883500
 2023-10-17  3.39 3.460 3.380   3.43  11997800
 2023-10-18  3.40 3.450 3.320   3.33  10501900
 2023-10-19  3.29 3.330 3.260   3.30  14663405
 *********************************************************

按照分析师的计划,聊天管理器将股票价格数据提供给UI设计师,UI设计师随后使用AmCharts为图表创建代码,并随后向聊天管理器报告。

 uidesigner (to chat_manager):

 Here is the complete code for the stock chart using the Amcharts library. The code includes the actual data for the last 30 days of the GRAB stock price.

 ```html
 <!DOCTYPE html>
 <html>
 <head>
     <title>AmCharts Stock Chart</title>
     <script src="https://www.amcharts.com/lib/3/amcharts.js"></script>
     <script src="https://www.amcharts.com/lib/3/serial.js"></script>
     <script src="https://www.amcharts.com/lib/3/amstock.js"></script>
     <script src="https://www.amcharts.com/lib/3/themes/light.js"></script>
 </head>
 <body>
     <div id="chartdiv" style="width: 100%; height: 500px;"></div>
     <script>
         var chartData = [
             { "date": new Date(2023, 8, 20), "open": 3.50, "high": 3.545, "low": 3.410, "close": 3.46, "volume": 13639700 },
             { "date": new Date(2023, 8, 21), "open": 3.45, "high": 3.460, "low": 3.380, "close": 3.46, "volume": 13781400 },
             { "date": new Date(2023, 8, 22), "open": 3.48, "high": 3.510, "low": 3.425, "close": 3.43, "volume": 8708200 },
             { "date": new Date(2023, 8, 25), "open": 3.43, "high": 3.470, "low": 3.385, "close": 3.43, "volume": 10990000 },
             { "date": new Date(2023, 8, 26), "open": 3.38, "high": 3.450, "low": 3.360, "close": 3.40, "volume": 11458800 },
             { "date": new Date(2023, 8, 27), "open": 3.40, "high": 3.460, "low": 3.360, "close": 3.38, "volume": 20921100 },
             { "date": new Date(2023, 8, 28), "open": 3.38, "high": 3.490, "low": 3.360, "close": 3.47, "volume": 14482300 },
             { "date": new Date(2023, 8, 29), "open": 3.52, "high": 3.590, "low": 3.520, "close": 3.54, "volume": 15580200 },
             { "date": new Date(2023, 9, 2), "open": 3.49, "high": 3.570, "low": 3.460, "close": 3.48, "volume": 19845600 },
             { "date": new Date(2023, 9, 3), "open": 3.48, "high": 3.480, "low": 3.330, "close": 3.35, "volume": 27370300 },
             { "date": new Date(2023, 9, 4), "open": 3.32, "high": 3.410, "low": 3.230, "close": 3.40, "volume": 18224300 },
             { "date": new Date(2023, 9, 5), "open": 3.37, "high": 3.500, "low": 3.370, "close": 3.48, "volume": 16904000 },
             { "date": new Date(2023, 9, 6), "open": 3.45, "high": 3.520, "low": 3.445, "close": 3.50, "volume": 9386900 },
             { "date": new Date(2023, 9, 9), "open": 3.47, "high": 3.480, "low": 3.380, "close": 3.44, "volume": 13512600 },
             { "date": new Date(2023, 9, 10), "open": 3.51, "high": 3.570, "low": 3.500, "close": 3.53, "volume": 11492700 },
             { "date": new Date(2023, 9, 11), "open": 3.51, "high": 3.550, "low": 3.480, "close": 3.49, "volume": 10962200 },
             { "date": new Date(2023, 9, 12), "open": 3.54, "high": 3.560, "low": 3.410, "close": 3.47, "volume": 13216900 },
             { "date": new Date(2023, 9, 13), "open": 3.44, "high": 3.460, "low": 3.360, "close": 3.42, "volume": 25474100 },
             { "date": new Date(2023, 9, 16), "open": 3.40, "high": 3.460, "low": 3.370, "close": 3.42, "volume": 22883500 },
             { "date": new Date(2023, 9, 17), "open": 3.39, "high": 3.460, "low": 3.380, "close": 3.43, "volume": 11997800 },
             { "date": new Date(2023, 9, 18), "open": 3.40, "high": 3.450, "low": 3.320, "close": 3.33, "volume": 10501900 },
             { "date": new Date(2023, 9, 19), "open": 3.29, "high": 3.330, "low": 3.260, "close": 3.30, "volume": 14663405 }
         ];

         var chart = AmCharts.makeChart("chartdiv", {
             "type": "stock",
             "theme": "light",
             "dataSets": [{
                 "fieldMappings": [{
                     "fromField": "open",
                     "toField": "open"
                 }, {
                     "fromField": "high",
                     "toField": "high"
                 }, {
                     "fromField": "low",
                     "toField": "low"
                 }, {
                     "fromField": "close",
                     "toField": "close"
                 }, {
                     "fromField": "volume",
                     "toField": "volume"
                 }],
                 "dataProvider": chartData,
                 "categoryField": "date"
             }],
             "panels": [{
                 "showCategoryAxis": false,
                 "title": "Value",
                 "percentHeight": 70,
                 "stockGraphs": [{
                     "id": "g1",
                     "valueField": "close",
                     "comparable": true,
                     "compareField": "close",
                     "balloonText": "[[title]]:<b>[[value]]</b>",
                     "compareGraphBalloonText": "[[title]]:<b>[[value]]</b>"
                 }],
                 "stockLegend": {
                     "periodValueTextComparing": "[[percents.value.close]]%",
                     "periodValueTextRegular": "[[value.close]]"
                 }
             }, {
                 "title": "Volume",
                 "percentHeight": 30,
                 "stockGraphs": [{
                     "valueField": "volume",
                     "type": "column",
                     "showBalloon": false,
                     "fillAlphas": 1
                 }],
                 "stockLegend": {
                     "periodValueTextRegular": "[[value.close]]"
                 }
             }],
             "chartScrollbarSettings": {
                 "graph": "g1"
             },
             "chartCursorSettings": {
                 "valueBalloonsEnabled": true,
                 "fullWidth": true,
                 "cursorAlpha": 0.1,
                 "valueLineBalloonEnabled": true,
                 "valueLineEnabled": true,
                 "valueLineAlpha": 0.5
             },
             "periodSelector": {
                 "position": "left",
                 "periods": [{
                     "period": "MM",
                     "selected": true,
                     "count": 1,
                     "label": "1 month"
                 }, {
                     "period": "YYYY",
                     "count": 1,
                     "label": "1 year"
                 }, {
                     "period": "YTD",
                     "label": "YTD"
                 }, {
                     "period": "MAX",
                     "label": "MAX"
                 }]
             },
             "dataSetSelector": {
                 "position": "left"
             },
             "export": {
                 "enabled": true
             }
         });
     </script>
 </body>
 </html>

This code will generate a stock chart for the GRAB stock price for the last 30 days. The chart includes open, high, low, close prices, and volume. The chart also includes a period selector and a data set selector.


因为这个过程不需要任何的人工干预,所以管理员admin没有任何的输入

admin (to chat_manager):



最后分析师提供对股票价格的分析,并与Admin共享反馈,以及生成的图表代码,发送TERMINATE消息以结束对话。

analyst (to chat_manager):

The stock price data for GRAB over the last 30 days shows a slight decrease in the closing price from 3.46 to 3.30. The highest closing price in this period was 3.54, and the lowest was 3.30. The volume of stocks traded also varied significantly, with the highest volume being 27,370,300 and the lowest being 8,708,200.

The chart created by the UI designer will provide a visual representation of these trends, making it easier to understand the stock's performance over the last month.

TERMINATE
```

上面就是AutoGen工作的一个完整的流程

总结

我们用AutoGen创建了多个代理。这些代理协作分析股票价格并使用AmCharts生成图表。软件工程师根据分析师的要求检索股票价格信息。设计师拿到股票价格数据、创建图表。而分析师则可以进行分析。AutoGen的AssistantAgents,促进了代理交互和代码执行,以对话的方式执行复杂的任务。

https://avoid.overfit.cn/post/2a2396c634a04041b1c0ffe6cf047ed1

目录
相关文章
|
人工智能 数据挖掘 API
AutoGen完整教程和加载本地LLM示例
Autogen是一个卓越的人工智能系统,它可以创建多个人工智能代理,这些代理能够协作完成任务,包括自动生成代码,并有效地执行任务。
1390 0
|
SQL 存储 物联网
基于 LLM 的知识图谱另类实践
大语言模型时代,我们有了 few-shot 和 zero-shot 的能力。借助这些 LLM 能力,如何更便捷地实现知识图谱的知识抽取,用知识图谱来解决相关问题。
853 1
基于 LLM 的知识图谱另类实践
|
7月前
|
人工智能 JavaScript Serverless
从零开始开发 MCP Server
本文介绍如何使用Serverless Devs CLI工具从零开发并一键部署MCP Server到阿里云函数计算(FC)。首先通过初始化MCP Server项目,完成本地代码编写,利用Node.js实现一个简单的Hello World工具。接着对代码进行打包,并通过Serverless Devs工具将项目部署至云端。部署完成后,提供三种客户端接入方式:官方Client、其他本地Client及在FC上部署的Client。最后可通过内置大模型的inspector测试部署效果。Serverless Devs简化了开发流程,提升了MCP Server的构建效率。
1457 119
|
4月前
|
消息中间件 人工智能 分布式计算
多智能体系统设计:协作、竞争与涌现行为
作为一名长期专注于分布式系统和人工智能领域的技术博主,我深深被多智能体系统(Multi-Agent Systems, MAS)的复杂性和优雅性所吸引。在过去几年的研究和实践中,我见证了多智能体系统从理论概念逐步走向实际应用的转变过程。多智能体系统不仅仅是简单的分布式计算模型,它更像是一个微观社会,其中每个智能体都具有自主性、反应性和社会性。这些智能体通过复杂的交互模式,展现出了令人惊叹的集体智能现象。从最初的简单协作模式,到复杂的竞争博弈,再到最终涌现出的群体智慧,多智能体系统为我们提供了一个全新的视角来理解和设计复杂系统。在本文中,我将从架构设计原则出发,深入探讨通信协议的设计要点,分析冲突
714 0
多智能体系统设计:协作、竞争与涌现行为
|
8月前
|
人工智能 JavaScript Java
从零开始教你打造一个MCP客户端
Anthropic开源了一套MCP协议,它为连接AI系统与数据源提供了一个通用的、开放的标准,用单一协议取代了碎片化的集成方式。本文教你从零打造一个MCP客户端。
6369 5
|
8月前
|
人工智能 自然语言处理 算法
HippoRAG 2:开源RAG框架革新知识检索,多跳推理+持续学习全搞定
HippoRAG 2 是俄亥俄州立大学推出的检索增强生成框架,通过个性化PageRank算法和知识图谱技术,显著提升了RAG系统在复杂问答任务中的表现。
1137 2
HippoRAG 2:开源RAG框架革新知识检索,多跳推理+持续学习全搞定
|
11月前
|
SQL API
金融行业 · 大模型挑战赛 |用大模型理解金融市场
2024金融行业大模型挑战赛即将开启,旨在推动大型语言模型在金融领域的应用。比赛提供金融多轮问答数据集,参赛者需使用GLM-4模型API,通过SQL、API等技术解决金融问题,涵盖数据查询、统计分析及复杂问题处理。赛事分初赛、复赛和决赛,总奖金20万元。报名时间为2024年12月2日至2025年2月6日。
916 16
金融行业 · 大模型挑战赛 |用大模型理解金融市场
|
人工智能 自然语言处理 PyTorch
从千问Agent看AI Agent——我们很强,但还有很长的路要走
本项目主要通过通义千问作为基础大模型,通义Agent浏览器助手实现网页和PDF材料,以帮助您快速了解多个页面的内容,总结您浏览过的内容,并减少繁琐的文字工作。实现数据分析与可视化、处理文件等的代码解释器功能。
从千问Agent看AI Agent——我们很强,但还有很长的路要走
|
11月前
|
人工智能 API 数据库
Qwen-Agent功能调用实践探索
本文详细解析了Qwen-Agent的核心功能——功能调用,涵盖其定义、工作流程、重要性和实际应用,通过实例展示了如何在Qwen-Agent中利用此功能与外部工具和API互动,扩展AI应用范围。