前言

本文通过Codeblaze.SemanticKernel这个项目，学习如何实现ITextEmbeddingGenerationService接口，接入本地嵌入模型.

项目地址：https://github.com/BLaZeKiLL/Codeblaze.SemanticKernel 。

实践

SemanticKernel初看以为只支持OpenAI的各种模型，但其实也提供了强大的抽象能力，可以通过自己实现接口，来实现接入不兼容OpenAI格式的模型.

Codeblaze.SemanticKernel这个项目实现了ITextGenerationService、IChatCompletionService与ITextEmbeddingGenerationService接口，由于现在Ollama的对话已经支持了OpenAI格式，因此可以不用实现ITextGenerationService和IChatCompletionService来接入Ollama中的模型了，但目前Ollama的嵌入还没有兼容OpenAI的格式，因此可以通过实现ITextEmbeddingGenerationService接口，接入Ollama中的嵌入模型.

查看ITextEmbeddingGenerationService接口:

代表了一种生成浮点类型文本嵌入的生成器.

再看看IEmbeddingGenerationService<string, float>接口:

[Experimental("SKEXP0001")]
public interface IEmbeddingGenerationService<TValue, TEmbedding> : IAIService where TEmbedding : unmanaged
{
      Task<IList<ReadOnlyMemory<TEmbedding>>> GenerateEmbeddingsAsync(IList<TValue> data, Kernel? kernel = null, CancellationToken cancellationToken = default(CancellationToken));
}

再看看IAIService接口:

说明我们只要实现了 。

Task<IList<ReadOnlyMemory<TEmbedding>>> GenerateEmbeddingsAsync(IList<TValue> data, Kernel? kernel = null, CancellationToken cancellationToken = default(CancellationToken));
​
IReadOnlyDictionary<string, object?> Attributes { get; }

这个方法和属性就行.

学习Codeblaze.SemanticKernel中是怎么做的.

添加OllamaBase类:

 public interface IOllamaBase
 {
     Task PingOllamaAsync(CancellationToken cancellationToken = new());
 }
 public abstract class OllamaBase<T> : IOllamaBase where T : OllamaBase<T>
 {
     public IReadOnlyDictionary<string, object?> Attributes => _attributes;
     private readonly Dictionary<string, object?> _attributes = new();
     protected readonly HttpClient Http;
     protected readonly ILogger<T> Logger;
​
     protected OllamaBase(string modelId, string baseUrl, HttpClient http, ILoggerFactory? loggerFactory)
     {
         _attributes.Add("model_id", modelId);
         _attributes.Add("base_url", baseUrl);
​
         Http = http;
         Logger = loggerFactory is not null ? loggerFactory.CreateLogger<T>() : NullLogger<T>.Instance;
     }
​
     /// <summary>
     /// Ping Ollama instance to check if the required llm model is available at the instance
     /// </summary>
     /// <param name="cancellationToken"></param>
     public async Task PingOllamaAsync(CancellationToken cancellationToken = new())
     {
         var data = new
         {
             name = Attributes["model_id"]
         };
​
         var response = await Http.PostAsJsonAsync($"{Attributes["base_url"]}/api/show", data, cancellationToken).ConfigureAwait(false);
​
         ValidateOllamaResponse(response);
​
         Logger.LogInformation("Connected to Ollama at {url} with model {model}", Attributes["base_url"], Attributes["model_id"]);
     }
​
     protected void ValidateOllamaResponse(HttpResponseMessage? response)
     {
         try
         {
             response.EnsureSuccessStatusCode();
         }
         catch (HttpRequestException)
         {
             Logger.LogError("Unable to connect to ollama at {url} with model {model}", Attributes["base_url"], Attributes["model_id"]);
         }
     }
 }

注意这个 。

public IReadOnlyDictionary<string, object?> Attributes => _attributes;

实现了接口中的属性.

添加OllamaTextEmbeddingGeneration类:

#pragma warning disable SKEXP0001
    public class OllamaTextEmbeddingGeneration(string modelId, string baseUrl, HttpClient http, ILoggerFactory? loggerFactory)
       : OllamaBase<OllamaTextEmbeddingGeneration>(modelId, baseUrl, http, loggerFactory),
            ITextEmbeddingGenerationService
   {
        public async Task<IList<ReadOnlyMemory<float>>> GenerateEmbeddingsAsync(IList<string> data, Kernel? kernel = null,
            CancellationToken cancellationToken = new())
       {
            var result = new List<ReadOnlyMemory<float>>(data.Count);
​
            foreach (var text in data)
           {
                var request = new
               {
                    model = Attributes["model_id"],
                    prompt = text
               };
​
                var response = await Http.PostAsJsonAsync($"{Attributes["base_url"]}/api/embeddings", request, cancellationToken).ConfigureAwait(false);
​
                ValidateOllamaResponse(response);
​
                var json = JsonSerializer.Deserialize<JsonNode>(await response.Content.ReadAsStringAsync().ConfigureAwait(false));
​
                var embedding = new ReadOnlyMemory<float>(json!["embedding"]?.AsArray().GetValues<float>().ToArray());
​
                result.Add(embedding);
           }
​
            return result;
       }
   }

注意实现了GenerateEmbeddingsAsync方法。实现的思路就是向Ollama中的嵌入接口发送请求，获得embedding数组.

为了在MemoryBuilder中能用还需要添加扩展方法:

#pragma warning disable SKEXP0001
    public static class OllamaMemoryBuilderExtensions
   {
        /// <summary>
        /// Adds Ollama as the text embedding generation backend for semantic memory
        /// </summary>
        /// <param name="builder">kernel builder</param>
        /// <param name="modelId">Ollama model ID to use</param>
        /// <param name="baseUrl">Ollama base url</param>
        /// <returns></returns>
        public static MemoryBuilder WithOllamaTextEmbeddingGeneration(
            this MemoryBuilder builder,
            string modelId,
            string baseUrl
       )
       {
            builder.WithTextEmbeddingGeneration((logger, http) => new OllamaTextEmbeddingGeneration(
                modelId,
                baseUrl,
                http,
                logger
           ));
​
            return builder;
       }       
   }

开始使用

 public async Task<ISemanticTextMemory> GetTextMemory3()
 {
     var builder = new MemoryBuilder();
     var embeddingEndpoint = "http://localhost:11434";
     var cancellationTokenSource = new System.Threading.CancellationTokenSource();
     var cancellationToken = cancellationTokenSource.Token;
     builder.WithHttpClient(new HttpClient());
     builder.WithOllamaTextEmbeddingGeneration("mxbai-embed-large:335m", embeddingEndpoint);
     IMemoryStore memoryStore = await SqliteMemoryStore.ConnectAsync("memstore.db");
     builder.WithMemoryStore(memoryStore);
     var textMemory = builder.Build();
     return textMemory;
 }

  builder.WithOllamaTextEmbeddingGeneration("mxbai-embed-large:335m", embeddingEndpoint);

实现了WithOllamaTextEmbeddingGeneration这个扩展方法，因此可以这么写，使用的是mxbai-embed-large:335m这个向量模型.

我使用WPF简单做了个界面，来试试效果.

找了一个新闻嵌入:

文本向量化存入数据库中:

现在测试RAG效果:

回答的效果也还可以.

大模型使用的是在线api的Qwen/Qwen2-72B-Instruct，嵌入模型使用的是本地Ollama中的mxbai-embed-large:335m.

最后此篇关于SemanticKernel/C#：实现接口，接入本地嵌入模型的文章就讲到这里了,如果你想了解更多关于SemanticKernel/C#：实现接口，接入本地嵌入模型的内容请搜索CFSDN的文章或继续浏览相关文章，希望大家以后支持我的博客！ 。