using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Connectors.OpenAI;
var builder = Kernel.CreateBuilder;
// 选择你的提供商：OpenAI 或 Azure OpenAI
builder.AddOpenAIChatCompletion(
modelId: "gpt-4o-mini", // 或你喜欢的模型
apiKey: Environment.GetEnvironmentVariable("OPENAI_API_KEY")
);
var kernel = builder.Build;

using Microsoft.SemanticKernel.SemanticFunctions;
var prompt = @"你是一个简洁的解释器。
用 3 个要点向初学者解释 '{topic}' 的概念。";
var explainFunc = kernel.CreateFunctionFromPrompt(prompt);
var result = await explainFunc.InvokeAsync(kernel, new { ["topic"] = "vector embeddings" });
Console.WriteLine(result);

public class WeatherPlugin
{
[KernelFunction, Description("获取城市天气")]
public async Task<string> GetWeatherAsync(
[Description("城市名称")] string city)
{
// 在这里调用你的天气 API
var temp = 22; // 占位符
return $"Weather in {city}: {temp}°C and clear";
}
}
// 注册插件
var weather = new WeatherPlugin;
kernel.Plugins.AddFromObject(weather, pluginName: "weather");

var prompt = @"
你是一个分类引擎。
任务：将 `message` 分类为 [Billing, Tech Support, Sales] 中的一个。
返回严格的 JSON：{ \"label\": string, \"confidence\": number }
message: {message}
";
var classify = kernel.CreateFunctionFromPrompt(prompt);
var output = await classify.InvokeAsync(kernel, new { ["message"] = "I can’t log into my account." });
Console.WriteLine(output); // {"label":"Tech Support","confidence":0.89}

using Microsoft.SemanticKernel.Memory;
var memory = new MemoryBuilder
.WithMemoryStore(new VolatileMemoryStore) // 替换为向量数据库（Qdrant、Pinecone、Azure AI Search）
.WithTextEmbeddingGeneration(
new OpenAITextEmbeddingGeneration("text-embedding-3-small", Environment.GetEnvironmentVariable("OPENAI_API_KEY")))
.Build;
await memory.SaveInformationAsync(
collection: "policies", id: "refund-policy",
text: "Customers can request a refund within 30 days of purchase with proof of receipt."
);
// 稍后：检索并填充到提示中
var results = memory.SearchAsync("refund window", collection: "policies", limit: 3, minRelevanceScore: 0.7);
await foreach (var item in results)
{
Console.WriteLine($"Relevant: {item.Metadata.Text}");
}

// 1) 像以前一样定义插件（语义 + 原生）
// 2) 使用规划器（API 界面可能因版本而异）
var goal = "总结附加的策略，对风险进行分类，并通过电子邮件发送报告";
// 假设我们有插件：files, summarize, classify, email
// 规划器将组装一个计划：files.Load → summarize.Run → classify.Run → email.Send
// 顺序执行计划，验证步骤之间的 JSON 输出

var kernel = Kernel.CreateBuilder
.AddOpenAIChatCompletion("gpt-4o-mini", apiKey)
.Build;
var fastKernel = Kernel.CreateBuilder
.AddOpenAIChatCompletion("gpt-4o-mini", apiKey) // 速度优化
.Build;
// 将更简单的提示路由到 fastKernel；将复杂的任务路由到 kernel

record Classification(string label, double confidence);
bool TryParseClassification(string text, out Classification cls)
{
try { cls = System.Text.Json.JsonSerializer.Deserialize<Classification>(text)!; return true; }
catch { cls = default!; return false; }
}

// 1) 将文档提取到内存中
await memory.SaveInformationAsync("handbook", "vacation-policy",
"Employees accrue 1.5 days of PTO per month and can carry over 5 days.");
// 2) 检索问题的上下文
var top = memory.SearchAsync("carry over PTO", "handbook", limit: 3, minRelevanceScore: 0.75);
var contexts = new List<string>;
await foreach (var r in top) contexts.Add(r.Metadata.Text);
// 3) 询问上下文并请求引用
var qaPrompt = @"
你严格从提供的上下文中回答。如果缺少，请说你不知道。
包括像 [source i] 这样的内联引用，使用从 1 开始的上下文项目索引。
Context:
1) {{ctx1}}
2) {{ctx2}}
Question: {{q}}
";
var qa = kernel.CreateFunctionFromPrompt(qaPrompt);
var variables = new KernelArguments
{
["ctx1"] = contexts.ElementAtOrDefault(0) ?? "",
["ctx2"] = contexts.ElementAtOrDefault(1) ?? "",
["q"] = "How many PTO days can I carry over?"
};
var answer = await qa.InvokeAsync(kernel, variables);
Console.WriteLine(answer);