Semantic Kernel | Multi-Agent system using Semantic Kernel

前言

Semantic Kernel 允許載入許多 Plugin，當啟用Function calling(ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions)後，
會讓 LLM 將解析用戶的問題，看看是否需要呼叫Plugin來完成用戶的問題。
例如要 onboard 一位新員工時，需要 1.建立員工帳號(MIS), 2.購買筆電(MIS), 3.申請識別證(HR), 4.寄送歡迎信件(HR) …等工作，
如果在一個 Agent 的狀況下，我們需要將這些工作全都轉成 Function 來讓 LLM 排出要完成的計劃(組合各個 Function)。
或是建立多個 Agent (ChatCompletionAgent)來協同完成用戶的問題。
之前探索 Semantic Kernel 在客服 ChatBot 中的應用(我的客服 Bot 除了會說，還會做)中已說明一個 Agent 的方式，
接下來，就來實作多個 Agent 一起來完成企業中的問題。

實作

1.加入套件(套件使用 1.15.1-alpha)

Microsoft.SemanticKernel.Agents.Abstractions
Microsoft.SemanticKernel.Agents.Core
Microsoft.SemanticKernel.Agents.OpenAI
Microsoft.SemanticKernel.Connectors.OpenAI

2.建立 MIS, HR 及 MIS 的 Plugin，如下，

public class KMPlugin
{
    [KernelFunction]
    [Description("Search Guidelines by Question")]
    [return: Description("Guidelines")]
    public async Task<string> Search(
        [Description("Question")] string question
    )
    {
        Console.WriteLine("Call KMPlugin=>Search");
        var result = "";
        if (question.ToLower().Contains("password"))
        {
            result = @"
#### 1. Reset user password , which is organised by MIS.
#### 2. Send Reset Password , which is organised by MIS.
";
        }
        else
        {
            result = @"
#### 1. set up their office 365 account and software access, which is organised by MIS.
#### 2. Order laptop for the employee, which is organised by MIS.
#### 3. Send an email welcoming the employee to the company, which is organised by HR.
#### 4. Order a badge for the employee so they can get access to the office, which is organised by HR.

";
        }
        return await Task.FromResult(result);
    }
}

public class HRPlugin
{
    [KernelFunction]
    [Description("申請Badge")]
    [return: Description("申請結果")]
    public async Task<string> IssueBadge(
        [Description("Email")] string email
    )
    {
        Console.WriteLine("Call HRPlugin=>IssueBadge");
        var result = $"The badge has been provisioned for: {email}";
        return await Task.FromResult(result);
    }

    [KernelFunction]
    [Description("Send Welcome Email")]
    [return: Description("send result")]
    public async Task<string> SendWelcomeEmail(
        [Description("Email")] string email
    )
    {
        Console.WriteLine("Call HRPlugin=>SendWelcomeEmail");
        var result = $"""
The welcome email has been send successfully to:{email}
""";
        return await Task.FromResult(result);
    }
}


public class MISPlugin
{
    [KernelFunction]
    [Description("Setup Office 365 account")]
    [return: Description("setup result")]
    public async Task<string> CreateOffice365Account(
        [Description("Email")] string email
    )
    {
        Console.WriteLine("Call MISPlugin=>CreateOffice365Account");
        var result = $"""
The office 365 account has beend created successfully.
The credentials are send to the employee's email address at : {email}
""";
        return await Task.FromResult(result);
    }


    [KernelFunction]
    [Description("Order Laptop")]
    [return: Description("order result")]
    public async Task<string> OrderLaptop(
        [Description("Email")] string email
    )
    {
        Console.WriteLine("Call MISPlugin=>OrderLaptop");
        var result = $"""
The laptio has been order for : {email}
""";
        return await Task.FromResult(result);
    }

    [KernelFunction]
    [Description("Reset User Password")]
    [return: Description("reset password result")]
    public async Task<string> ResetPassword(
        [Description("Email")] string email
    )
    {
        Console.WriteLine("Call MISPlugin=>ResetPassword");
        var result = $"""
Reset Password for : {email}
""";
        return await Task.FromResult(result);
    }


    [KernelFunction]
    [Description("Send Reset Password Link Email")]
    [return: Description("send email result")]
    public async Task<string> SendResetPasswordLinkEmail(
        [Description("Email")] string email
    )
    {
        Console.WriteLine("Call MISPlugin=>SendResetPasswordLinkEmail");
        var result = $"""
Send Reset Password Link Email for : {email}
""";
        return await Task.FromResult(result);
    }
}

• 註: 主要會依 KMPlugin 的 Search Function 來找出要執行的 Guidelines 來讓 LLM 幫我們編排工作
• 註: Guidelines 中有各 Agent 要做的 Tasks

3.建立 Agetns (MIS, HR, AgentsManager, Manager)
AgentsManager: 依用戶的問題呼叫負責任的 Agent 去找出要執行的 Guidelines ，解析 Guidelines 指定給負責任的 Agent 執行
MIS, HR: 透過 KernelBuilder 來載入各別的 Plugin(BuildHRKernel,BuildMISKernel)，並負責執行 AgentsManager 指派的工作
Manager: 負責判斷所有工作是否完成，如果完成就回覆approve，做為結束的依據
最後建立AgentGroupChat，並加入上述 Agents, 並設定結束的策略為Manager發出approve就結束

//使用 AOAI
var endpoint = "https://你的aoai.openai.azure.com/";
var deploymentName = "你的aoai LLM deploymentName";
var apikey = "你的aoai apikey";

string agentsManager = """
你是 Agents Manager ，你的任務是將用戶的問題呼叫 Search Function 取得要執行的內容。
依上述的內容，並依照順序，指定給所屬的Agent去執行。
不要重覆執行 Search ，如果沒有內容，就回結束
""";


string hr = """
你是 HR，你的目標是處理 Agents Manager 指派的工作。
""";

string mis = """
你是 MIS，你的目標是處理 Agents Manager 指派的工作。
""";

string manager = """
You are Manager which will review agents output
and make sure all tasks are completed.
Once all requirements are completed,
you can approve the request by just responding "approve"
""";

ChatCompletionAgent agentsManagerAgent =
            new()
            {
                Instructions = agentsManager,
                Name = "AgentsManager",
                ExecutionSettings = new OpenAIPromptExecutionSettings() { ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions },
                Kernel = BuildKMKernel()
            };

ChatCompletionAgent hrAgent =
            new()
            {
                Instructions = hr,
                Name = "HR",
                Kernel = BuildHRKernel(),
                ExecutionSettings = new OpenAIPromptExecutionSettings() { ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions },
            };


ChatCompletionAgent misAgent =
            new()
            {
                Instructions = hr,
                Name = "MIS",
                Kernel = BuildMISKernel(),
                ExecutionSettings = new OpenAIPromptExecutionSettings() { ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions },
            };


ChatCompletionAgent managerAgent =
            new()
            {
                Instructions = manager,
                Name = "Manager",
                Kernel = GetChatCompletionBuilder().Build()
            };


AgentGroupChat groupChat =
        new(agentsManagerAgent, misAgent, hrAgent, managerAgent)
        {
            ExecutionSettings = new()
            {
                TerminationStrategy = new ApprovalTerminationStrategy
                {
                    Agents = [managerAgent],
                    MaximumIterations = 6,
                }
            }

        };

IKernelBuilder GetChatCompletionBuilder()
{
    IKernelBuilder builder = Kernel.CreateBuilder();
    builder.AddAzureOpenAIChatCompletion(deploymentName, endpoint, apikey);
    return builder;
}

Kernel BuildKMKernel()
{
    var builder = GetChatCompletionBuilder();
    builder.Plugins.AddFromType<KMPlugin>();
    return builder.Build();
}

Kernel BuildHRKernel()
{
    var builder = GetChatCompletionBuilder();
    builder.Plugins.AddFromType<HRPlugin>();
    return builder.Build();
}

Kernel BuildMISKernel()
{
    var builder = GetChatCompletionBuilder();
    builder.Plugins.AddFromType<MISPlugin>();
    return builder.Build();
}

sealed class ApprovalTerminationStrategy : TerminationStrategy
{
    protected override Task<bool> ShouldAgentTerminateAsync(Agent agent, IReadOnlyList<ChatMessageContent> history, CancellationToken cancellationToken)
        => Task.FromResult(history[history.Count - 1].Content?.Contains("approve", StringComparison.OrdinalIgnoreCase) ?? false);
}

4.將用戶問題發送到 AgentGroupChat 中，並執行InvokeAsync來取得結果，如下，

Console.WriteLine("Please input question:");
string input = Console.ReadLine();

groupChat.AddChatMessage(new ChatMessageContent(AuthorRole.User, input));
Console.WriteLine($"# {AuthorRole.User}: '{input}'");
await foreach (var content in groupChat.InvokeAsync())
{
    Console.WriteLine($"# {content.Role} - {content.AuthorName ?? "*"}: '{content.Content}'");
}

輸入i want to onboard a new employee. ther email is rm@gss.com.tw，如果如下，

Please input question:
i want to onboard a new employee. ther email is rm@gss.com.tw
# user: 'i want to onboard a new employee. ther email is rm@gss.com.tw'
Call KMPlugin=>Search
# assistant - AgentsManager: 'Here are the steps to onboard the new employee with email rm@gss.com.tw:

1. Set up their Office 365 account and software access (organized by MIS).
2. Order a laptop for the employee (organized by MIS).
3. Send an email welcoming the employee to the company (organized by HR).
4. Order a badge for the employee so they can get access to the office (organized by HR).

Depending on your company's workflow, you might need to coordinate with both MIS and HR departments for these tasks.'
Call MISPlugin=>CreateOffice365Account
Call MISPlugin=>OrderLaptop
# assistant - MIS: 'The Office 365 account has been created successfully, and the credentials have been sent to the email address rm@gss.com.tw. Additionally, a laptop has been ordered for the new employee.

Next steps:
1. Send a welcome email to the new employee.
2. Order a badge for the employee so they can access the office.'
Call HRPlugin=>SendWelcomeEmail
Call HRPlugin=>IssueBadge
# assistant - HR: 'The welcome email has been successfully sent to rm@gss.com.tw, and the badge has been provisioned for the new employee.

The onboarding process for the new employee is now completed!'
# assistant - Manager: 'approve'

Search 來取出 Guidelines 1.建帳號, 2.訂筆電, 3.寄 Welcom email, 4.申請識別證
知道要執行的 Agents ，執行的順序與加入 AgentGroupChat 順序有關

輸入i forgot password. ther email is rm@gss.com.tw，如果如下，

Please input question:
i forgot password. ther email is rm@gss.com.tw
# user: 'i forgot password. ther email is rm@gss.com.tw'
Call KMPlugin=>Search
# assistant - AgentsManager: 'To reset the password for the email address rm@gss.com.tw, follow these steps:

1. Contact your MIS (Management Information Systems) department to reset the user password.
2. After the password is reset, the MIS department will send the new password or password reset instructions to you.

Please reach out to your MIS department for further assistance.'
Call MISPlugin=>ResetPassword
Call MISPlugin=>SendResetPasswordLinkEmail
# assistant - MIS: 'The password for rm@gss.com.tw has been reset, and a reset password link email has been sent to the same address. Please check your email for further instructions.'
# assistant - HR: 'The password for rm@gss.com.tw has been reset, and the MIS department has sent a reset password link email to the same address. Please check your email for further instructions.'
# assistant - Manager: 'approve'

Guidelines 要執行 1.重設密碼, 2.寄送重設密碼連結的 Email

參考資源

Step by Step guide to develop AI Multi-Agent system using Microsoft Semantic Kernel and GPT-4o
Business-in-a-box: Applying AutoGen and multi-agent systems to an enterprise context
探索 Semantic Kernel 在客服 ChatBot 中的應用(我的客服 Bot 除了會說，還會做)

一條大路十二彎，這頭不彎那頭彎: 喻同一件事，同一個道理，有各種不同的看法，和不同的發展。