使用C#构建一个同时问多个LLM并总结的小工具

mingupupup2025-05-08 16:11

前言

在AI编程时代，如果自己能够知道一些可行的解决方案，那么描述清楚交给AI，可以有很大的帮助。

但是我们往往不知道真正可行的解决方案是什么？

我自己有过这样的经历，遇到一个需求，我不知道有哪些解决方案，就去问AI，然后AI输出一大堆东西，我一个个去试，然后再换个AI问，又提出了不同的解决方案。

在换AI问与一个个试的过程中好像浪费了很多时间。

突然出现了一个想法，不是可以一下子把问题丢给多个AI，然后再总结一下出现最多的三个方案。那么这三个方案可行的概率会大一点。然后再丢给Cursor或者Cline等AI编程工具帮我们实现一下。

这样做的缺点是比起直接在网页上问，调用API需要耗费Token，但是硅基流动给我赠送了很多额度还没用完，随便玩一下。

实现效果：

实现方案

实现方案也很简单，如下图所示：

先设计一下布局：

xaml 复制代码 
<UserControl xmlns="https://github.com/avaloniaui"
             xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
             xmlns:d="http://schemas.microsoft.com/expression/blend/2008"
             xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"
             mc:Ignorable="d" d:DesignWidth="800" d:DesignHeight="450"
			 xmlns:vm="using:AIE_Studio.ViewModels"
			 x:DataType="vm:DuoWenViewModel"
             x:Class="AIE_Studio.Views.DuoWenView">
	<StackPanel>
		<TextBox Text="{Binding Question}"></TextBox>
		<Button Content="提问" Command="{Binding DuoWenStreamingParallelCommand}" Margin="5"/>
		<ScrollViewer VerticalScrollBarVisibility="Auto">
		<Grid>
			<Grid.RowDefinitions>
				<RowDefinition Height="*"/>
				<RowDefinition Height="*"/>
			</Grid.RowDefinitions>
			<Grid.ColumnDefinitions>
				<ColumnDefinition Width="*"/>
				<ColumnDefinition Width="*"/>
				<ColumnDefinition Width="*"/>
			</Grid.ColumnDefinitions>
			<!-- Row 1, Column 1 -->
			<StackPanel Grid.Row="0" Grid.Column="0">
				<TextBlock Text="{Binding Title1}" Margin="5"/>
				<ScrollViewer VerticalScrollBarVisibility="Auto">
					<TextBox Text="{Binding Result1}" AcceptsReturn="True" Margin="5" Height="300"/>
				</ScrollViewer>
			</StackPanel>
			<!-- Row 1, Column 2 -->
			<StackPanel Grid.Row="0" Grid.Column="1">
				<TextBlock Text="{Binding Title2}" Margin="5"/>
				<ScrollViewer VerticalScrollBarVisibility="Auto">
					<TextBox Text="{Binding Result2}" AcceptsReturn="True" Margin="5" Height="300"/>
				</ScrollViewer>
			</StackPanel>
			<!-- Row 1, Column 3 -->
			<StackPanel Grid.Row="0" Grid.Column="2">
				<TextBlock Text="{Binding Title3}" Margin="5"/>
				<ScrollViewer VerticalScrollBarVisibility="Auto">
					<TextBox Text="{Binding Result3}" AcceptsReturn="True" Margin="5" Height="300"/>
				</ScrollViewer>
			</StackPanel>
			<!-- Row 2, Column 1 -->
			<StackPanel Grid.Row="1" Grid.Column="0">
				<TextBlock Text="{Binding Title4}" Margin="5"/>
				<ScrollViewer VerticalScrollBarVisibility="Auto">
					<TextBox Text="{Binding Result4}" AcceptsReturn="True" Margin="5" Height="300"/>
				</ScrollViewer>
			</StackPanel>
			<!-- Row 2, Column 2 -->
			<StackPanel Grid.Row="1" Grid.Column="1">
				<TextBlock Text="{Binding Title5}" Margin="5"/>
				<ScrollViewer VerticalScrollBarVisibility="Auto">
					<TextBox Text="{Binding Result5}" AcceptsReturn="True" Margin="5" Height="300"/>
				</ScrollViewer>
			</StackPanel>
			<!-- Row 2, Column 3 -->
			<StackPanel Grid.Row="1" Grid.Column="2">
				<TextBlock Text="{Binding Title6}" Margin="5"/>
				<ScrollViewer VerticalScrollBarVisibility="Auto">
					<TextBox Text="{Binding Result6}" AcceptsReturn="True" Margin="5" Height="300"/>
				</ScrollViewer>
			</StackPanel>
		</Grid>
		</ScrollViewer>
	</StackPanel>
</UserControl>

在ViewModel中先来看一下最原始的显示结果的方式：

csharp 复制代码 
 [RelayCommand]
 private async Task DuoWen()
 {
     ApiKeyCredential apiKeyCredential = new ApiKeyCredential("your api key");

     OpenAIClientOptions openAIClientOptions = new OpenAIClientOptions();
     openAIClientOptions.Endpoint = new Uri("https://api.siliconflow.cn/v1");
   
     IChatClient client1 =
     new OpenAI.Chat.ChatClient("Qwen/Qwen2.5-72B-Instruct", apiKeyCredential, openAIClientOptions).AsChatClient();

     var result1 = await client1.GetResponseAsync(Question);

     Result1 = result1.ToString();

     IChatClient client2 =
     new OpenAI.Chat.ChatClient("Qwen/Qwen3-235B-A22B", apiKeyCredential, openAIClientOptions).AsChatClient();

     var result2 = await client2.GetResponseAsync(Question);

     Result2 = result2.ToString();

     IChatClient client3 =
     new OpenAI.Chat.ChatClient("THUDM/GLM-Z1-32B-0414", apiKeyCredential, openAIClientOptions).AsChatClient();

     var result3 = await client3.GetResponseAsync(Question);

     Result3 = result3.ToString();

     IChatClient client4 =
     new OpenAI.Chat.ChatClient("THUDM/GLM-4-32B-0414", apiKeyCredential, openAIClientOptions).AsChatClient();

     var result4 = await client4.GetResponseAsync(Question);

     Result4 = result4.ToString();

     IChatClient client5 =
    new OpenAI.Chat.ChatClient("deepseek-ai/DeepSeek-R1", apiKeyCredential, openAIClientOptions).AsChatClient();

     var result5 = await client5.GetResponseAsync(Question);

     Result5 = result5.ToString();

     IChatClient client6 =
     new OpenAI.Chat.ChatClient("deepseek-ai/DeepSeek-V3", apiKeyCredential, openAIClientOptions).AsChatClient();

     var result6 = await client6.GetResponseAsync(Question);

     Result6 = result6.ToString();

这种最简单的方式是非流式的并且也不是并行的，你会发现一个结束了才会继续向下一个提问。

但至少已经成功显示结果了，现在想要实现的是有一个窗体进行总结。

窗体设计：

xaml 复制代码 
<Window xmlns="https://github.com/avaloniaui"
        xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
        xmlns:d="http://schemas.microsoft.com/expression/blend/2008"
        xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"
        mc:Ignorable="d" d:DesignWidth="600" d:DesignHeight="450"
		xmlns:vm="using:AIE_Studio.ViewModels"
        x:Class="AIE_Studio.Views.ShowResultWindow"
		x:DataType="vm:ShowResultWindowViewModel"
        Title="ShowResultWindow">
	<StackPanel>
		<TextBlock Text="最终结果：" Margin="5" />
		<ScrollViewer VerticalScrollBarVisibility="Auto">
			<TextBox Text="{Binding ReceivedValue}" AcceptsReturn="True" Margin="5" Height="400"/>
		</ScrollViewer>
	</StackPanel>
</Window>

窗体的ViewModel：

csharp 复制代码 
public partial class ShowResultWindowViewModel : ViewModelBase
{
    [ObservableProperty]
    private string? receivedValue;      
}

然后只要在全部都有结果之后，再进行一下总结即可。

csharp 复制代码 
IChatClient client7 =
new OpenAI.Chat.ChatClient("Qwen/Qwen2.5-72B-Instruct", apiKeyCredential, openAIClientOptions).AsChatClient();
List<Microsoft.Extensions.AI.ChatMessage> messages = new List<Microsoft.Extensions.AI.ChatMessage>();

string prompt = $"""
      请分析以下各个助手给出的方案，选择其中提到最多的3种方案。
      助手1：{result1}
      助手2：{result2}
      助手3：{result3}
      助手4：{result4}
      助手5：{result5}
      助手6：{result6}
      """;
messages.Add(new Microsoft.Extensions.AI.ChatMessage(ChatRole.User, prompt));
var result7 = await client7.GetResponseAsync(messages);

var showWindow = _serviceProvider.GetRequiredService<ShowResultWindow>();
var showWindowViewModel = _serviceProvider.GetRequiredService<ShowResultWindowViewModel>();
showWindowViewModel.ReceivedValue = result7.ToString();
showWindow.DataContext = showWindowViewModel;
showWindow.Show();

以上就成功实现了。

但是还是有可以改进的地方，首先是并行，一个一个问不如同时问。

csharp 复制代码 
 [RelayCommand]
 private async Task DuoWenParallel()
 {
     ApiKeyCredential apiKeyCredential = new ApiKeyCredential("your api key");

     OpenAIClientOptions openAIClientOptions = new OpenAIClientOptions();
     openAIClientOptions.Endpoint = new Uri("https://api.siliconflow.cn/v1");

     // 创建一个列表来存储所有的任务
     var tasks = new List<Task<string>>();

     // 向每个助手发送请求并将任务添加到列表中
     tasks.Add(GetResponseFromClient("Qwen/Qwen2.5-72B-Instruct", apiKeyCredential, openAIClientOptions));
     tasks.Add(GetResponseFromClient("Qwen/Qwen3-235B-A22B", apiKeyCredential, openAIClientOptions));
     tasks.Add(GetResponseFromClient("THUDM/GLM-Z1-32B-0414", apiKeyCredential, openAIClientOptions));
     tasks.Add(GetResponseFromClient("THUDM/GLM-4-32B-0414", apiKeyCredential, openAIClientOptions));
     tasks.Add(GetResponseFromClient("deepseek-ai/DeepSeek-R1", apiKeyCredential, openAIClientOptions));
     tasks.Add(GetResponseFromClient("deepseek-ai/DeepSeek-V3", apiKeyCredential, openAIClientOptions));

     // 等待所有任务完成
     var results = await Task.WhenAll(tasks);

     // 将结果分配给相应的属性
     Result1 = results[0];
     Result2 = results[1];
     Result3 = results[2];
     Result4 = results[3];
     Result5 = results[4];
     Result6 = results[5];
 }

  private async Task<string> GetResponseFromClient(string model, ApiKeyCredential apiKeyCredential, OpenAIClientOptions options)
  {
      IChatClient client = new OpenAI.Chat.ChatClient(model, apiKeyCredential, options).AsChatClient();
      var result = await client.GetResponseAsync(Question);
      return result.ToString();
  }

现在虽然是并行了，但是只有等到所有助手都回答了之后，才会统一显示，用户体验也不好。

改成流式：

csharp 复制代码 
[RelayCommand]
private async Task DuoWenStreaming()
{
    ApiKeyCredential apiKeyCredential = new ApiKeyCredential("your api key");

    OpenAIClientOptions openAIClientOptions = new OpenAIClientOptions();
    openAIClientOptions.Endpoint = new Uri("https://api.siliconflow.cn/v1");

    //string question = "C#如何获取鼠标滑动选中的值？请告诉我一些可能的方案，每个方案只需用一句话描述即可，不用展开说明。";

    IChatClient client1 =
    new OpenAI.Chat.ChatClient("Qwen/Qwen2.5-72B-Instruct", apiKeyCredential, openAIClientOptions).AsChatClient();

    await foreach (var item in client1.GetStreamingResponseAsync(Question))
    {
       Result1 += item.ToString();
    }          
}

现在查看效果：

最后再改造成流式+并行就好了。

csharp 复制代码 
 [RelayCommand]
 private async Task DuoWenStreamingParallel()
 {
     ApiKeyCredential apiKeyCredential = new ApiKeyCredential("your api key");

     OpenAIClientOptions openAIClientOptions = new OpenAIClientOptions();
     openAIClientOptions.Endpoint = new Uri("https://api.siliconflow.cn/v1");

     // Clear previous results
     Result1 = Result2 = Result3 = Result4 = Result5 = Result6 = string.Empty;

     // Create a list of tasks for parallel processing
     var tasks = new List<Task>
     {
         ProcessStreamingResponse("Qwen/Qwen2.5-72B-Instruct", apiKeyCredential, openAIClientOptions, (text) => Result1 += text),
         ProcessStreamingResponse("Qwen/Qwen3-235B-A22B", apiKeyCredential, openAIClientOptions, (text) => Result2 += text),
         ProcessStreamingResponse("THUDM/GLM-Z1-32B-0414", apiKeyCredential, openAIClientOptions, (text) => Result3 += text),
         ProcessStreamingResponse("THUDM/GLM-4-32B-0414", apiKeyCredential, openAIClientOptions, (text) => Result4 += text),
         ProcessStreamingResponse("deepseek-ai/DeepSeek-R1", apiKeyCredential, openAIClientOptions, (text) => Result5 += text),
         ProcessStreamingResponse("deepseek-ai/DeepSeek-V3", apiKeyCredential, openAIClientOptions, (text) => Result6 += text)
     };

     // Wait for all streaming responses to complete
     await Task.WhenAll(tasks);

     IChatClient client7 =
     new OpenAI.Chat.ChatClient("Qwen/Qwen2.5-72B-Instruct", apiKeyCredential, openAIClientOptions).AsChatClient();
     List<Microsoft.Extensions.AI.ChatMessage> messages = new List<Microsoft.Extensions.AI.ChatMessage>();

     string prompt = $"""
           请分析以下各个助手给出的方案，选择其中提到最多的3种方案。
           助手1：{Result1}
           助手2：{Result2}
           助手3：{Result3}
           助手4：{Result4}
           助手5：{Result5}
           助手6：{Result6}
           """;
     messages.Add(new Microsoft.Extensions.AI.ChatMessage(ChatRole.User, prompt));
     var result7 = await client7.GetResponseAsync(messages);

     var showWindow = _serviceProvider.GetRequiredService<ShowResultWindow>();
     var showWindowViewModel = _serviceProvider.GetRequiredService<ShowResultWindowViewModel>();
     showWindowViewModel.ReceivedValue = result7.ToString();
     showWindow.DataContext = showWindowViewModel;
     showWindow.Show();
 }

private async Task ProcessStreamingResponse(string model, ApiKeyCredential apiKeyCredential, OpenAIClientOptions options, Action<string> updateResult)
{
    IChatClient client = new OpenAI.Chat.ChatClient(model, apiKeyCredential, options).AsChatClient();
    
    await foreach (var item in client.GetStreamingResponseAsync(Question))
    {
        updateResult(item.ToString());
    }
}

这里使用了一个带有一个参数的委托来更新每个助手回复的结果。

现在再查看效果：

Qwen/Qwen3-235B-A22B、THUDM/GLM-Z1-32B-0414、deepseek-ai/DeepSeek-R1有思考过程，返回结果比较慢。

目前Microsoft.Extensions.AI.OpenAI好像还无法获取思考内容。

等待久一会之后，可以看到结果都出来了：

然后总结窗口会显示最终的总结内容：

确定方案之后可以让Cursor或者Cline帮我们写一下试试。

热门推荐
01从零安装 LLaMA-Factory 微调 Qwen 大模型成功及所有的坑02KGG转MP3工具|非KGM文件|解密音频03YOLOv8入门 | 重要性能衡量指标、训练结果评价及分析及影响mAP的因素【发论文关注的指标】04DeepSeek各版本说明与优缺点分析05【SpeedAI科研小助手】2分钟极速解决知网维普重复率、AIGC率过高，一键全文降！文件格式不变，公式都保留的！06组基轨迹建模 GBTM的介绍与实现（Stata 或 R）07Coze扣子平台完整体验和实践（附国内和国际版对比）08VMware虚拟机安装Win7专业版保姆级教程（附镜像包）09苍穹外卖面试总结10ImageJ软件使用教程（二）：图像测量