练习 - 提供上下文提示

10 分钟

在本练习中，你将使用对话历史记录向大型语言模型（LLM）提供上下文。还可以调整代码，使用户能够继续聊天，就像真正的聊天机器人一样。让我们开始吧！

修改代码以使用 do-while 循环接受用户的输入：

string input;

do 
{
    Console.WriteLine("What would you like to do?");
    input = Console.ReadLine();

    // ...
}
while (!string.IsNullOrWhiteSpace(input));

现在，你可以继续对话，直到用户输入一个空白行。

通过修改 SuggestDestinations 事例捕获有关用户行程的详细信息：

case "SuggestDestinations":
    chatHistory.AppendLine("User:" + input);
    var recommendations = await kernel.InvokePromptAsync(input!);
    Console.WriteLine(recommendations);
    break;

将 SuggestActivities 示例中的行程详细信息与以下代码一起使用：
```
 case "SuggestActivities":
    var chatSummary = await kernel.InvokeAsync(
        "ConversationSummaryPlugin", 
        "SummarizeConversation", 
        new() {{ "input", chatHistory.ToString() }});
    break;
```
在此代码中，你将使用内置 SummarizeConversation 函数来汇总与用户的聊天。接下来，让我们使用摘要来建议目标上的活动。

使用以下代码扩展 SuggestActivities 案例：

var activities = await kernel.InvokePromptAsync(
    input,
    new () {
        {"input", input},
        {"history", chatSummary},
        {"ToolCallBehavior", ToolCallBehavior.AutoInvokeKernelFunctions}
});

chatHistory.AppendLine("User:" + input);
chatHistory.AppendLine("Assistant:" + activities.ToString());

Console.WriteLine(activities);
break;

在此代码中，将 input 和 chatSummary 添加为内核参数。然后，内核调用提示并将其路由到 SuggestActivities 插件。还可以将用户的输入和助手的响应追加到聊天历史记录并显示结果。接下来，需要将 chatSummary 变量添加到 SuggestActivities 插件。

导航到 Prompts/SuggestActivities/config.json 并在 Visual Studio Code 中打开该文件

在 input_variables下，为聊天历史记录添加变量：

"input_variables": [
  {
      "name": "history",
      "description": "Some background information about the user",
      "required": false
  },
  {
      "name": "destination",
      "description": "The destination a user wants to visit",
      "required": true
  }
  ]

导航到 Prompts/SuggestActivities/skprompt.txt 并打开该文件

添加提示以使用聊天历史记录：

You are an experienced travel agent. 
You are helpful, creative, and very friendly. 
Consider the traveler's background: {{$history}}

按原样保留其余提示。现在，插件使用聊天历史记录向 LLM 提供上下文。

检查你的工作

在此任务中，运行应用程序并验证代码是否正常工作。

将更新的开关事例与以下代码进行比较：

case "SuggestDestinations":
        chatHistory.AppendLine("User:" + input);
        var recommendations = await kernel.InvokePromptAsync(input!);
        Console.WriteLine(recommendations);
        break;
case "SuggestActivities":

    var chatSummary = await kernel.InvokeAsync(
        "ConversationSummaryPlugin", 
        "SummarizeConversation", 
        new() {{ "input", chatHistory.ToString() }});

    var activities = await kernel.InvokePromptAsync(
        input!,
        new () {
            {"input", input},
            {"history", chatSummary},
            {"ToolCallBehavior", ToolCallBehavior.AutoInvokeKernelFunctions}
    });

    chatHistory.AppendLine("User:" + input);
    chatHistory.AppendLine("Assistant:" + activities.ToString());

    Console.WriteLine(activities);
    break;

在终端中输入 dotnet run。出现提示时，输入类似于以下内容的一些文本：
```
What would you like to do?
How much is 60 USD in new zealand dollars?
```

您将看到类似于以下内容的输出：

$60 USD is approximately $97.88 in New Zealand Dollars (NZD)
What would you like to do?

输入提示以获取目标建议，附带一些上下文提示，例如：

What would you like to do?
I'm planning an anniversary trip with my spouse, but they are currently using a wheelchair and accessibility is a must. What are some destinations that would be romantic for us?

你应该会收到一些输出，其中包含可访问目标的建议。

输入活动建议的提示，例如：

What would you like to do?
What are some things to do in Barcelona?

您应该收到符合之前背景的建议，例如，类似以下在巴塞罗那的无障碍活动：

1. Visit the iconic Sagrada Família: This breathtaking basilica is an iconic symbol of Barcelona's architecture and is known for its unique design by Antoni Gaudí.

2. Explore Park Güell: Another masterpiece by Gaudí, this park offers stunning panoramic views of the city, intricate mosaic work, and whimsical architectural elements.

3. Visit the Picasso Museum: Explore the extensive collection of artworks by the iconic painter Pablo Picasso, showcasing his different periods and styles.

注意

如果代码未生成预期的输出，则可以在 解决方案 文件夹中查看代码。

可以使用不同的提示和上下文提示继续测试应用程序。出色的工作！你已成功向 LLM 提供上下文提示，并调整了代码，以允许用户继续对话。

练习 - 提供上下文提示

检查你的工作

反馈