Merge pull request #25 from AIDotNet/feature/embedding_format

- 支持根据请求的向量类型自动帮忙转换返回的嵌入模型的类型

src/Thor.Abstractions/ObjectModels/ObjectModels/ResponseModels/EmbeddingCreateResponse.cs

@@ -1,25 +1,91 @@
-using System.Text.Json.Serialization;
+using System.Buffers;
+using System.Text.Json;
+using System.Text.Json.Serialization;
 using Thor.Abstractions.Dtos;
 
 namespace Thor.Abstractions.ObjectModels.ObjectModels.ResponseModels;
 
 public record EmbeddingCreateResponse : ThorBaseResponse
 {
-    [JsonPropertyName("model")] 
-    public string Model { get; set; }
+    [JsonPropertyName("model")] public string Model { get; set; }
 
-    [JsonPropertyName("data")] 
-    public List<EmbeddingResponse> Data { get; set; }
+    [JsonPropertyName("data")] public List<EmbeddingResponse> Data { get; set; } = [];
 
-    [JsonPropertyName("usage")] 
-    public ThorUsageResponse Usage { get; set; }
+    /// <summary>
+    /// 类型转换，如果类型是base64,则将float[]转换为base64,如果是空或是float和原始类型一样，则不转换
+    /// </summary>
+    public void ConvertEmbeddingData(string? encodingFormat)
+    {
+        if (Data.Count == 0)
+        {
+            return;
+        }
+
+        switch (encodingFormat)
+        {
+            // 判断第一个是否是float[]，如果是则不转换
+            case null or "float" when Data[0].Embedding is float[]:
+                return;
+            // 否则转换成float[]
+            case null or "float":
+            {
+                foreach (var embeddingResponse in Data)
+                {
+                    if (embeddingResponse.Embedding is string base64)
+                    {
+                        embeddingResponse.Embedding = Convert.FromBase64String(base64);
+                    }
+                }
+
+                return;
+            }
+            // 判断第一个是否是string，如果是则不转换
+            case "base64" when Data[0].Embedding is string:
+                return;
+            // 否则转换成base64
+            case "base64":
+            {
+                foreach (var embeddingResponse in Data)
+                {
+                    if (embeddingResponse.Embedding is JsonElement str)
+                    {
+                        if (str.ValueKind == JsonValueKind.Array)
+                        {
+                            var floats = str.EnumerateArray().Select(element => element.GetSingle()).ToArray();
+
+                            embeddingResponse.Embedding = ConvertFloatArrayToBase64(floats);
+                        }
+                    }
+                }
+
+                break;
+            }
+        }
+    }
+
+    public static string ConvertFloatArrayToBase64(float[] floatArray)
+    {
+        // 将 float[] 转换成 byte[]
+        byte[] byteArray = ArrayPool<byte>.Shared.Rent(floatArray.Length * sizeof(float));
+        try
+        {
+            Buffer.BlockCopy(floatArray, 0, byteArray, 0, byteArray.Length);
+
+            // 将 byte[] 转换成 base64 字符串
+            return Convert.ToBase64String(byteArray);
+        }
+        finally
+        {
+            ArrayPool<byte>.Shared.Return(byteArray);
+        }
+    }
+
+    [JsonPropertyName("usage")] public ThorUsageResponse Usage { get; set; }
 }
 
 public record EmbeddingResponse
 {
-    [JsonPropertyName("index")] 
-    public int? Index { get; set; }
+    [JsonPropertyName("index")] public int? Index { get; set; }
 
-    [JsonPropertyName("embedding")] 
-    public object Embedding { get; set; }
+    [JsonPropertyName("embedding")] public object Embedding { get; set; }
 }

src/Thor.Service/Program.cs

@@ -4,6 +4,7 @@
 using Microsoft.AspNetCore.Mvc;
 using Serilog;
 using Thor.Abstractions.Chats.Dtos;
+using Thor.Abstractions.Embeddings.Dtos;
 using Thor.Abstractions.ObjectModels.ObjectModels.RequestModels;
 using Thor.AzureOpenAI.Extensions;
 using Thor.BuildingBlocks.Data;
@@ -227,7 +228,6 @@
 
         var loggerDbContext = scope.ServiceProvider.GetRequiredService<LoggerDbContext>();
         await loggerDbContext.Database.EnsureCreatedAsync();
-
     }
 // 由于没有生成迁移记录，所以使用EnsureCreated
     else if (string.Equals(dbType, "postgresql") || string.Equals(dbType, "pgsql") ||
@@ -241,7 +241,6 @@
 
         var loggerDbContext = scope.ServiceProvider.GetRequiredService<LoggerDbContext>();
         await loggerDbContext.Database.EnsureCreatedAsync();
-
     }
 
 
@@ -707,8 +706,9 @@ await service.CompletionsAsync(context))
         .WithDescription("Get completions from OpenAI")
         .WithOpenApi();
 
-    app.MapPost("/v1/embeddings", async (ChatService embeddingService, HttpContext context) =>
-            await embeddingService.EmbeddingAsync(context))
+    app.MapPost("/v1/embeddings",
+            async (ChatService embeddingService, HttpContext context, ThorEmbeddingInput module) =>
+                await embeddingService.EmbeddingAsync(context, module))
         .WithDescription("OpenAI")
         .WithDescription("Embedding")
         .WithOpenApi();
@@ -721,10 +721,7 @@ await imageService.CreateImageAsync(context, request))
         .WithOpenApi();
 
 
-    app.MapGet("/v1/models", async (HttpContext context) =>
-        {
-            return await ModelService.GetAsync(context);
-        })
+    app.MapGet("/v1/models", async (HttpContext context) => { return await ModelService.GetAsync(context); })
         .WithDescription("获取模型列表")
         .RequireAuthorization()
         .WithOpenApi();

src/Thor.Service/Service/ChatService.cs

@@ -162,15 +162,10 @@ await loggerService.CreateConsumeAsync(string.Format(ConsumerTemplate, rate, 0),
         }
     }
 
-    public async ValueTask EmbeddingAsync(HttpContext context)
+    public async ValueTask EmbeddingAsync(HttpContext context,ThorEmbeddingInput module)
     {
         try
         {
-            using var body = new MemoryStream();
-            await context.Request.Body.CopyToAsync(body);
-
-            var module = JsonSerializer.Deserialize<ThorEmbeddingInput>(body.ToArray());
-
             if (module == null) throw new Exception("模型校验异常");
 
             await rateLimitModelService.CheckAsync(module!.Model, context);
@@ -246,6 +241,8 @@ await userService.ConsumeAsync(user!.Id, (long)quota, requestToken, token?.Key,
                     module.Model);
             }
 
+            stream.ConvertEmbeddingData(module.EncodingFormat);
+
             await context.Response.WriteAsJsonAsync(stream);
         }
         catch (RateLimitException)