AI 大模型应用开发初探：基于 Semantic Kernel 构建简易 Agent

我的需求其实包含两个步骤：第一步是更新工单的 Quantity，第二步是查询更新后的工单信息。而这两个步骤我们假设其实都是需要去调用 MES WorkOrderService API 才能获得的，这就需要我们给 LLM 加入 Function Calling 的能力，当然 LLM 自己得知道如何规划执行的步骤，哪个步骤先执行，哪个后执行。

示例代码结构

我们将项目分为几个关键部分：配置管理、API 处理、业务服务模拟、以及主程序逻辑。

1. 配置管理 (Shared)

首先定义 OpenAI 兼容服务的配置类，支持多提供商切换。

public class OpenAiConfiguration
{
    public string Provider { get; set; }
    public string ModelId { get; set; }
    public string EndPoint { get; set; }
    public string ApiKey { get; set; }

    public OpenAiConfiguration(string modelId, string endPoint, string apiKey)
    {
        // 默认使用 OpenAI 兼容协议
        Provider = ConfigConstants.LLMProviders.OpenAI;
        ModelId = modelId;
        EndPoint = endPoint;
        ApiKey = apiKey;
    }

    public OpenAiConfiguration(string provider, string modelId, string endPoint, string apiKey)
    {
        Provider = provider;
        ModelId = modelId;
        EndPoint = endPoint;
        ApiKey = apiKey;
    }
}

为了支持不同的 LLM 提供商（如智谱 AI、通义千问等），我们需要自定义 HttpClientHandler 来重写请求路径。

public class CustomLLMApiHandler : HttpClientHandler
{
    private readonly string _openAiProvider;
    private readonly string _openAiBaseAddress;

    public CustomLLMApiHandler(string openAiProvider, string openAiBaseAddress)
    {
        _openAiProvider = openAiProvider;
        _openAiBaseAddress = openAiBaseAddress;
    }

    protected override async Task<HttpResponseMessage> SendAsync(
        HttpRequestMessage request, CancellationToken cancellationToken)
    {
        UriBuilder uriBuilder;
        Uri uri = new Uri(_openAiBaseAddress);
        switch (request.RequestUri?.LocalPath)
        {
            case "/v1/chat/completions":
                switch (_openAiProvider)
                {
                    case ConfigConstants.LLMProviders.ZhiPuAI:
                        uriBuilder = new UriBuilder(request.RequestUri)
                        {
                            Scheme = "https",
                            Host = uri.Host,
                            Path = ConfigConstants.LLMApiPaths.ZhiPuAIChatCompletions,
                        };
                        request.RequestUri = uriBuilder.Uri;
                        break;
                    default: // Default: OpenAI-Compatible API Providers
                        uriBuilder = new UriBuilder(request.RequestUri)
                        {
                            Scheme = "https",
                            Host = uri.Host,
                            Path = ConfigConstants.LLMApiPaths.OpenAIChatCompletions,
                        };
                        request.RequestUri = uriBuilder.Uri;
                        break;
                }
                break;
        }

        HttpResponseMessage response = await base.SendAsync(request, cancellationToken);
        return response;
    }
}

2. 业务服务模拟 (WorkOrderService)

在实际场景中，这里应该替换为真实的 HTTP 客户端调用企业 ERP 或 MES 系统的 API。为了演示方便，我们直接在内存中模拟数据。

public class WorkOrderService
{
    // 模拟数据库中的工单列表
    private static List<WorkOrder> workOrders = new List<WorkOrder>
    {
        new WorkOrder { WorkOrderName = "9050100", ProductName = "A5E900100", ProductVersion = "001 / AB", Quantity = 100, Status = "Ready" },
        new WorkOrder { WorkOrderName = "9050101", ProductName = "A5E900101", ProductVersion = "001 / AB", Quantity = 200, Status = "Ready" },
        new WorkOrder { WorkOrderName = "9050102", ProductName = "A5E900102", ProductVersion = "001 / AB", Quantity = 300, Status = "InProcess" },
        new WorkOrder { WorkOrderName = "9050103", ProductName = "A5E900103", ProductVersion = "001 / AB", Quantity = 400, Status = "InProcess" },
        new WorkOrder { WorkOrderName = "9050104", ProductName = "A5E900104", ProductVersion = "001 / AB", Quantity = 500, Status = "Completed" }
    };

    // 查询工单详情
    public WorkOrder GetWorkOrderInfo(string orderName)
    {
        return workOrders.Find(o => o.WorkOrderName == orderName);
    }

    // 更新工单状态
    public string UpdateWorkOrderStatus(string orderName, string newStatus)
    {
        var workOrder = this.GetWorkOrderInfo(orderName);
        if (workOrder == null)
            return "Operate Failed : The work order is not existing!";
        workOrder.Status = newStatus;
        return "Operate Succeed!";
    }

    // 减少工单数量
    public string ReduceWorkOrderQuantity(string orderName, int newQuantity)
    {
        var workOrder = this.GetWorkOrderInfo(orderName);
        if (workOrder == null)
            return "Operate Failed : The work order is not existing!";

        // 业务校验逻辑
        if (workOrder.Status == "Completed")
            return "Operate Failed : The work order is completed, can not be reduced!";
        if (newQuantity <= 1 || newQuantity >= workOrder.Quantity)
            return "Operate Failed : The new quantity is invalid!";
        
        workOrder.Quantity = newQuantity;
        return "Operate Succeed!";
    }
}

// 简单的数据模型
public class WorkOrder
{
    public string WorkOrderName { get; set; }
    public string ProductName { get; set; }
    public string ProductVersion { get; set; }
    public int Quantity { get; set; }
    public string Status { get; set; }
}

3. 配置文件 (appsettings.json)

建议使用环境变量或密钥管理服务（如 Azure Key Vault、HashiCorp Vault）来管理敏感信息，避免硬编码在配置文件中。

{
  "LLM_API_PROVIDER": "ZhiPuAI",
  "LLM_API_MODEL": "glm-4",
  "LLM_API_BASE_URL": "https://open.bigmodel.cn",
  "LLM_API_KEY": "your_api_key_here"
}

4. 初始化 Kernel 与插件

在应用程序启动时，加载配置并初始化 Semantic Kernel。

private Kernel _kernel = null;
private OpenAIPromptExecutionSettings _settings = null;
private IChatCompletionService _chatCompletion = null;
private ChatHistory _chatHistory = null;

private void ChatForm_Load(object sender, EventArgs e)
{
    // 1. 加载配置
    var configuration = new ConfigurationBuilder().AddJsonFile($"appsettings.{ConfigConstants.LLM_API_PROVIDER}.json").Build();
    var config = configuration;
    
    var openAiConfiguration = new OpenAiConfiguration(
        config["LLM_API_PROVIDER"],
        config["LLM_API_MODEL"],
        config["LLM_API_BASE_URL"],
        config["LLM_API_KEY"]);

    // 2. 创建带自定义路由的 HttpClient
    var openAiClient = new HttpClient(new CustomLlmApiHandler(openAiConfiguration.Provider, openAiConfiguration.EndPoint));
    
    // 3. 构建 Kernel
    _kernel = Kernel.CreateBuilder()
        .AddOpenAIChatCompletion(openAiConfiguration.ModelId, openAiConfiguration.ApiKey, httpClient: openAiClient)
        .Build();

    _chatCompletion = _kernel.GetRequiredService<IChatCompletionService>();
    _chatHistory = new ChatHistory();
    
    // 4. 设置系统提示词，定义 Agent 角色
    _chatHistory.AddSystemMessage("You are one WorkOrder Assistant. You help users manage manufacturing work orders.");
}

接下来，注册 Functions 到 Kernel 中。这是让 LLM 拥有'手脚'的关键步骤。

_kernel.Plugins.Add(KernelPluginFactory.CreateFromFunctions("WorkOrderHelperPlugin",
    new List<KernelFunction>
    {
        // 注册查询函数
        _kernel.CreateFunctionFromMethod((string orderName) =>
        {
            var workOrderRepository = new WorkOrderService();
            return workOrderRepository.GetWorkOrderInfo(orderName);
        }, "GetWorkOrderInfo", "Get WorkOrder's Detail Information"),
        
        // 注册减少数量函数
        _kernel.CreateFunctionFromMethod((string orderName, int newQuantity) =>
        {
            var workOrderRepository = new WorkOrderService();
            return workOrderRepository.ReduceWorkOrderQuantity(orderName, newQuantity);
        }, "ReduceWorkOrderQuantity", "Reduce WorkOrder's Quantity to new Quantity"),
        
        // 注册更新状态函数
        _kernel.CreateFunctionFromMethod((string orderName, string newStatus) =>
        {
            var workOrderRepository = new WorkOrderService();
            return workOrderRepository.UpdateWorkOrderStatus(orderName, newStatus);
        }, "UpdateWorkOrderStatus", "Update WorkOrder's Status to new Status")
     }));

开启自动调用 Function，告诉大模型可以自行决定调用相关 Functions，而且大模型会自行决定根据什么顺序来调用，这就是大模型作为 Agent 大脑的规划能力。当然，我们还可以通过定制化 Planner 来增强 agent 的规划能力。

_settings = new OpenAIPromptExecutionSettings
{
   ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions
};

5. 发送用户提示词

最后，实现用户交互逻辑。区分是否启用 Function Calling 模式进行测试。

private void btnSend_Click(object sender, EventArgs e)
{
    _chatHistory.AddUserMessage(tbxPrompt.Text);
    ChatMessageContent chatResponse = null;
    tbxResponse.Clear();

    if (cbxUseFunctionCalling.Checked)
    {
        Task.Run(() =>
        {
            ShowProcessMessage("AI is handling your request now...");
            try
            {
                chatResponse = _chatCompletion.GetChatMessageContentAsync(_chatHistory, _settings, _kernel)
                    .GetAwaiter()
                    .GetResult();
                UpdateResponseContent(chatResponse.ToString());
                ShowProcessMessage("AI Response:");
            }
            catch (Exception ex)
            {
                UpdateResponseContent($"Error: {ex.Message}");
            }
        });
    }
    else
    {
        Task.Run(() =>
        {
            ShowProcessMessage("AI is handling your request now...");
            try
            {
                chatResponse = _chatCompletion.GetChatMessageContentAsync(_chatHistory, null, _kernel)
                    .GetAwaiter()
                    .GetResult();
                UpdateResponseContent(chatResponse.ToString());
                ShowProcessMessage("AI Response:");
            }
            catch (Exception ex)
            {
                UpdateResponseContent($"Error: {ex.Message}");
            }
        });
    }
}

最佳实践与安全建议

在将 Agent 投入生产环境之前，还需要考虑以下关键点：

1. 安全与权限控制

• API Key 管理：切勿将 API Key 硬编码在代码库中。应使用环境变量、Azure Key Vault 或 AWS Secrets Manager。
• 权限最小化：Agent 调用的工具函数应遵循最小权限原则。例如，工单更新接口不应允许删除工单，除非有明确授权。
• 输入过滤：对用户输入的 Prompt 进行安全检查，防止 Prompt Injection 攻击导致 Agent 执行恶意操作。

2. 错误处理与重试

• 网络异常：LLM 调用和外部 API 调用都可能超时或失败。需要实现指数退避重试机制。
• 业务逻辑校验：如示例代码所示，在 Function 内部进行严格的参数校验，防止脏数据写入。
• 日志记录：详细记录 Agent 的思考过程（Thought）、工具调用参数及返回结果，便于问题排查和审计。

3. 性能优化

• 缓存策略：对于相同的查询请求，可以在 Service 层增加缓存，减少 LLM 调用次数和延迟。
• 流式输出：对于长文本回复，建议使用 Stream 模式，提升用户体验。

总结

本文简单介绍了 AI Agent 的基本概念和工作方式，对比了高代码手搓和低代码拖拉拽两种开发模式。最后，通过 C# + Semantic Kernel + 智谱 GLM-4 模型演示了如何快速开发一个简易的 AI Agent。

虽然这只是个 Demo，但希望对你快速了解 Agent 有所帮助。未来，随着 RAG（检索增强生成）技术的引入和更复杂的 Planner 设计，Agent 将在企业级应用中发挥更大的价值。开发者应关注如何平衡自动化程度与人工干预的边界，确保系统的安全可控。

扩展阅读

• Microsoft Semantic Kernel 官方文档
• OpenAI Function Calling 指南
• LangChain 框架架构分析
• 大模型在企业内部署的最佳实践