如何拿到接口返回的消耗token

SemanticKernel

以下引用自官方案例

1. Text模型
   1. 使用Kernel

          FunctionResult functionResult = await kernel.InvokePromptAsync(promptTestDataGeneration);
          CompletionsUsage? usage = FunctionResult.Metadata?["Usage"] as CompletionsUsage;
          //回答消耗的Token
          Console.WriteLine(usage.CompletionTokens);
          //提示词消耗的Token
      	Console.WriteLine(usage.PromptTokens);
      	//一共消耗的Token
      	Console.WriteLine(usage.TotalTokens);
          ```
      

   2. 直接使用AzureOpenAIChatCompletionService

      	var service = new AzureOpenAIChatCompletionService("deployment", "https://endpoint", "api-key", "model-id", this._httpClient);
      	var result = await service.GetChatMessageContentsAsync([], settings, kernel);
      	var usage = result[0].Metadata?["Usage"] as CompletionsUsage;
      	//回答消耗的Token
          Console.WriteLine(usage.CompletionTokens);
          //提示词消耗的Token
      	Console.WriteLine(usage.PromptTokens);
      	//一共消耗的Token
      	Console.WriteLine(usage.TotalTokens);
      	```
      

2. 暂不支持Embedding模型

客制方法

1. 当前SK仅支持OpenAI和AzureOpenAI的模型调用，如何需要其他模型供应商目前有两种办法解决
2. 使用OneAPI做一个中转成AzureOpenAI然后使用上述SK的方法使用。
3. 自定义方法发送请求，使用AddKeyedSingleton注入。

   1. 定义IAzureTextEmbeddingCompletionService接口，引用ITextEmbeddingGenerationService

      public interface IAzureTextEmbeddingCompletionService : ITextEmbeddingGenerationService
      {
      	Task<(List<float>, TextEmbeddingUsage)> GenerateEmbeddingsByUsageAsync(IList<string> data, Kernel kernel = null, CancellationToken cancellationToken = default);
      }
      

   2. 定义AzureTextEmbeddingCompletionService实现IAzureTextEmbeddingCompletionService接口

      public class AzureTextEmbeddingCompletionService : IAzureTextEmbeddingCompletionService
      {
      	private readonly SemanticAzureOpenAIConfig config;
      	private readonly Dictionary<string, object?> _attributes = new();
      	private readonly ModelClient client;
      	public IReadOnlyDictionary<string, object> Attributes => _attributes;
      
      	public AzureTextEmbeddingCompletionService(ModelOptions options)
      	{
      		config = new SemanticAzureOpenAIConfig()
      		{
      			ApiKey = options.ApiKey,
      			Endpoint = options.Endpoint,
      			DeploymentName = "text-embedding-ada-002",
      			ApiVersion = options.ModelVersion
      		};
      		client = new(config.ApiKey, ModelType.AzureOpenAI, config.Endpoint);
      	}
      
      	public async Task<IList<ReadOnlyMemory<float>>> GenerateEmbeddingsAsync(IList<string> data, Kernel kernel = null, CancellationToken cancellationToken = default)
      	{
      		var result = await client.AzureOpenAI.GenerateEmbeddingsAsync(config.DeploymentName, data.First());
      		var list = result.Data[0].Embedding.ToList();
      		IList<ReadOnlyMemory<float>> readOnlyMemoryList = list.Select(f => new ReadOnlyMemory<float>(new[] { f })).ToList();
      		return readOnlyMemoryList;
      	}
      
      	public async Task<(List<float>, TextEmbeddingUsage)> GenerateEmbeddingsByUsageAsync(IList<string> data, Kernel kernel = null, CancellationToken cancellationToken = default)
      	{
      		var result = await client.AzureOpenAI.GenerateEmbeddingsAsync(config.DeploymentName, data.First());
      		var list = result.Data[0].Embedding.ToList();
      		IList<ReadOnlyMemory<float>> readOnlyMemoryList = list.Select(f => new ReadOnlyMemory<float>(new[] { f })).ToList();
      		return (list, result.Usage);
      	}
      }
      //定义入参类
      public class ModelOptions
      {
      	public string AppId { get; set; }
      
      	public string ApiKey { get; set; }
      
      	public string ApiSecret { get; set; }
      	public string Endpoint { get; set; }
      	public string DeploymentName { get; set; }
      	public ChatHistory ChatHistory { get; set; } = new ChatHistory();
      
      	public string ModelVersion { get; set; }
      }
      

   3. ModelClient的实现参考对话模型的对接方式实现，原理是一致的这里不详述。(.Net接入AzureOpenAI、OpenAI、通义千问、智谱AI、讯飞星火、文心一言大语言模型。 | FaceMan)

   4. 创建Kernel实例，注入IAzureTextEmbeddingCompletionService服务

      var builder = Kernel.CreateBuilder();
      var options = new ModelOptions()
      	{
      					Endpoint = "YourEndpoint",
      					ApiKey = "YourApiKey",
      					DeploymentName = "YourDeploymentName",,
      					ChatHistory = "YourChatHistory",//Embedding可不填
      	};
      	builder.Services.AddKeyedSingleton<IAzureTextEmbeddingCompletionService>("AzureOpenAIEmbedding", new AzureTextEmbeddingCompletionService(options));
      

   5. 获取Embedding服务，获取Token

      var service = _kernel.GetRequiredService<IAzureTextEmbeddingCompletionService>();
      var res = await service.GenerateEmbeddingsByUsageAsync
      		(new List<string> { partitionContent }, cancellationToken: cancellationToken)
      		.ConfigureAwait(false);
      // item1 向量，item2消耗的token
      float[] result = res.Item1.ToArray<float>();
      totalToken += res.Item2.TotalTokens;
      Embedding embedding = new Embedding(result);