Merge pull request #57 from sdcb/dev

Dev

src/BE/Controllers/Admin/AdminMessage/AdminMessageController.cs

@@ -109,6 +109,7 @@ public async Task<ActionResult<ChatsResponseWithMessage>> GetAdminMessage(int ch
                     OutputPrice = x.Usage.OutputCost,
                     ReasoningTokens = x.Usage.ReasoningTokens,
                     Duration = x.Usage.TotalDurationMs - x.Usage.PreprocessDurationMs,
+                    ReasoningDuration = x.Usage.ReasoningDurationMs,
                     FirstTokenLatency = x.Usage.FirstResponseDurationMs,
                     ModelId = x.Usage.UserModel.ModelId,
                     ModelName = x.Usage.UserModel.Model.Name,

src/BE/Controllers/Admin/ModelKeys/ModelKeysController.cs

@@ -149,7 +149,8 @@ public async Task<ActionResult<PossibleModelDto[]>> ListModelKeyPossibleModels(s
                 {
                     bool isVision = model.Contains("qvq", StringComparison.OrdinalIgnoreCase) ||
                         model.Contains("vision", StringComparison.OrdinalIgnoreCase);
-                    short modelReferenceId = isVision ? (short)1401 : (short)1400;
+                    bool isDeepSeek = model.Contains("deepseek", StringComparison.OrdinalIgnoreCase);
+                    short modelReferenceId = isDeepSeek ? (short)1402 : isVision ? (short)1401 : (short)1400;
                     return new PossibleModelDto()
                     {
                         DeploymentName = model,
@@ -196,7 +197,7 @@ public async Task<ActionResult<PossibleModelDto[]>> ListModelKeyPossibleModels(s
                     DeploymentName = x.Models.FirstOrDefault(m => m.ModelKeyId == modelKeyId)!.DeploymentName,
                     ReferenceId = x.Id,
                     ReferenceName = x.Name,
-                    IsLegacy = x.PublishDate == null || x.PublishDate < new DateOnly(2024, 7, 1),
+                    IsLegacy = x.PublishDate != null && x.PublishDate < new DateOnly(2024, 7, 1),
                     IsExists = x.Models.Any(m => m.ModelKeyId == modelKeyId),
                 })
                 .OrderBy(x => (x.IsLegacy ? 1 : 0) + (x.IsExists ? 2 : 0))

src/BE/Controllers/Chats/Chats/ChatController.cs

@@ -400,10 +400,27 @@ private static async Task<ChatSpanResponse> ProcessChatSpan(
         try
         {
             using ChatService s = chatFactory.CreateChatService(userModel.Model);
+            bool responseStated = false, reasoningStarted = false;
             await foreach (InternalChatSegment seg in icc.Run(userBalance.Balance, userModel, s.ChatStreamedFEProcessed(messageToSend, cco, extraDetails, cancellationToken)))
             {
-                if (seg.TextSegment == string.Empty) continue;
-                await writer.WriteAsync(SseResponseLine.Segment(span.Id, seg.TextSegment), cancellationToken);
+                if (!string.IsNullOrEmpty(seg.ReasoningSegment))
+                {
+                    if (!reasoningStarted)
+                    {
+                        await writer.WriteAsync(SseResponseLine.StartReasoning(span.Id), cancellationToken);
+                        reasoningStarted = true;
+                    }
+                    await writer.WriteAsync(SseResponseLine.ReasoningSegment(span.Id, seg.ReasoningSegment), cancellationToken);
+                }
+                if (!string.IsNullOrEmpty(seg.Segment))
+                {
+                    if (!responseStated)
+                    {
+                        await writer.WriteAsync(SseResponseLine.StartResponse(span.Id, icc.ReasoningDurationMs), cancellationToken);
+                        responseStated = true;
+                    }
+                    await writer.WriteAsync(SseResponseLine.Segment(span.Id, seg.Segment), cancellationToken);
+                }
 
                 if (cancellationToken.IsCancellationRequested)
                 {
@@ -422,6 +439,12 @@ private static async Task<ChatSpanResponse> ProcessChatSpan(
             errorText = e.Message;
             logger.LogError(e, "Upstream error: {userMessageId}", req.MessageId);
         }
+        catch (AggregateException e) when (e.InnerException is TaskCanceledException)
+        {
+            // do nothing if cancelled
+            icc.FinishReason = DBFinishReason.Cancelled;
+            errorText = e.InnerException.ToString();
+        }
         catch (TaskCanceledException)
         {
             // do nothing if cancelled
@@ -446,10 +469,6 @@ private static async Task<ChatSpanResponse> ProcessChatSpan(
         {
             ChatId = chat.Id,
             ChatRoleId = (byte)DBChatRole.Assistant,
-            MessageContents =
-            [
-                MessageContent.FromText(icc.FullResponse.TextSegment),
-            ],
             SpanId = span.Id,
             CreatedAt = DateTime.UtcNow,
         };
@@ -461,10 +480,13 @@ private static async Task<ChatSpanResponse> ProcessChatSpan(
         {
             dbAssistantMessage.ParentId = req.MessageId;
         }
+        foreach (MessageContent mc in MessageContent.FromFullResponse(icc.FullResponse, errorText))
+        {
+            dbAssistantMessage.MessageContents.Add(mc);
+        }
 
         if (errorText != null)
         {
-            dbAssistantMessage.MessageContents.Add(MessageContent.FromError(errorText));
             await writer.WriteAsync(SseResponseLine.Error(span.Id, errorText), cancellationToken);
         }
         dbAssistantMessage.Usage = icc.ToUserModelUsage(currentUser.Id, await clientInfoTask, isApi: false);
@@ -490,7 +512,7 @@ private static MessageLiteDto MakeSystemMessage(byte? specifiedSpanId, Chat chat
             ChatRoleId = (byte)DBChatRole.System,
             MessageContents =
             [
-                MessageContent.FromText(span.SystemPrompt!)
+                MessageContent.FromContent(span.SystemPrompt!)
             ],
             CreatedAt = DateTime.UtcNow,
             SpanId = specifiedSpanId,

src/BE/Controllers/Chats/Chats/Dtos/MessageLiteDto.cs

@@ -36,12 +36,11 @@ public async Task<ChatMessage> ToOpenAI(FileUrlProvider fup, CancellationToken c
                 .SelectAwait(async c => await c.ToOpenAI(fup, cancellationToken))
                 .ToArrayAsync(cancellationToken)),
             DBChatRole.Assistant => new AssistantChatMessage(await Content
-                .Where(x => x.ContentTypeId != (byte)DBMessageContentType.Error)
+                .Where(x => x.ContentTypeId != (byte)DBMessageContentType.Error && x.ContentTypeId != (byte)DBMessageContentType.Reasoning)
                 .ToAsyncEnumerable()
                 .SelectAwait(async x => await x.ToOpenAI(fup, cancellationToken))
                 .ToArrayAsync(cancellationToken)),
             _ => throw new NotImplementedException()
         };
     }
 }
-

src/BE/Controllers/Chats/Chats/Dtos/SseResponseKind.cs

@@ -11,4 +11,7 @@ public enum SseResponseKind
     TitleSegment = 5,
     ResponseMessage = 6,
     ChatLeafMessageId = 7,
+    ReasoningSegment = 8,
+    StartResponse = 9,
+    StartReasoning = 10,
 }

src/BE/Controllers/Chats/Chats/Dtos/SseResponseLine.cs

@@ -28,6 +28,36 @@ public static SseResponseLine Segment(byte spanId, string segment)
         };
     }
 
+    public static SseResponseLine ReasoningSegment(byte spanId, string segment)
+    {
+        return new SseResponseLine
+        {
+            SpanId = spanId,
+            Result = segment,
+            Kind = SseResponseKind.ReasoningSegment,
+        };
+    }
+
+    public static SseResponseLine StartResponse(byte spanId, int reasoningTimeMs)
+    {
+        return new SseResponseLine
+        {
+            SpanId = spanId,
+            Result = reasoningTimeMs,
+            Kind = SseResponseKind.StartResponse,
+        };
+    }
+
+    public static SseResponseLine StartReasoning(byte spanId)
+    {
+        return new SseResponseLine
+        {
+            SpanId = spanId,
+            Result = null!,
+            Kind = SseResponseKind.StartReasoning,
+        };
+    }
+
     public static SseResponseLine Error(byte spanId, string error)
     {
         return new SseResponseLine

src/BE/Controllers/Chats/Messages/Dtos/MessageDto.cs

@@ -71,6 +71,9 @@ public record ResponseMessageDto : MessageDto
     [JsonPropertyName("duration")]
     public required int Duration { get; init; }
 
+    [JsonPropertyName("reasoningDuration")]
+    public required int ReasoningDuration { get; init; }
+
     [JsonPropertyName("firstTokenLatency")]
     public required int FirstTokenLatency { get; init; }
 
@@ -99,7 +102,7 @@ public async Task<MessageContent[]> ToMessageContents(FileUrlProvider fup, Cance
     {
         return
         [
-            MessageContent.FromText(Text),
+            MessageContent.FromContent(Text),
             ..(await (FileIds ?? [])
                 .ToAsyncEnumerable()
                 .SelectAwait(async fileId => await fup.CreateFileContent(fileId, cancellationToken))
@@ -113,6 +116,9 @@ public record MessageContentResponse
     [JsonPropertyName("text")]
     public required string Text { get; init; }
 
+    [JsonPropertyName("think"), JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
+    public required string? Think { get; init; }
+
     [JsonPropertyName("fileIds"), JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
     public required FileDto[]? FileIds { get; init; }
 
@@ -132,6 +138,7 @@ public static MessageContentResponse FromSegments(MessageContent[] segments, Fil
         return new MessageContentResponse()
         {
             Text = string.Join("\n", groups[DBMessageContentType.Text].Select(x => x.ToString())),
+            Think = string.Join("\n", groups[DBMessageContentType.Reasoning].Select(x => x.ToString())) switch { "" => null, var x => x },
             FileIds = groups[DBMessageContentType.FileId]
                 .Select(x => fup.CreateFileDto(x.MessageContentFile!.File))
                 .ToArray() switch 
@@ -156,6 +163,7 @@ public record FileDto
 public record ChatMessageTempUsage
 {
     public required int Duration { get; init; }
+    public required int ReasoningDuration { get; init; }
     public required int FirstTokenLatency { get; init; }
     public required decimal InputPrice { get; init; }
     public required int InputTokens { get; init; }
@@ -213,6 +221,7 @@ public MessageDto ToDto(IUrlEncryptionService urlEncryption, FileUrlProvider fup
                 OutputPrice = Usage.OutputPrice,
                 ReasoningTokens = Usage.ReasoningTokens,
                 Duration = Usage.Duration,
+                ReasoningDuration = Usage.ReasoningDuration,
                 FirstTokenLatency = Usage.FirstTokenLatency,
                 ModelId = Usage.ModelId,
                 ModelName = Usage.ModelName,
@@ -240,6 +249,7 @@ public static ChatMessageTemp FromDB(Message assistantMessage)
                 Usage = assistantMessage.Usage == null ? null : new ChatMessageTempUsage()
                 {
                     Duration = assistantMessage.Usage.TotalDurationMs - assistantMessage.Usage.PreprocessDurationMs,
+                    ReasoningDuration = assistantMessage.Usage.ReasoningDurationMs,
                     FirstTokenLatency = assistantMessage.Usage.FirstResponseDurationMs,
                     InputPrice = assistantMessage.Usage.InputCost,
                     InputTokens = assistantMessage.Usage.InputTokens,

src/BE/Controllers/Chats/Messages/MessagesController.cs

@@ -43,6 +43,7 @@ public async Task<ActionResult<MessageDto[]>> GetMessages(string chatId, [FromSe
                     OutputPrice = x.Usage.OutputCost,
                     ReasoningTokens = x.Usage.ReasoningTokens,
                     Duration = x.Usage.TotalDurationMs - x.Usage.PreprocessDurationMs,
+                    ReasoningDuration = x.Usage.ReasoningDurationMs,
                     FirstTokenLatency = x.Usage.FirstResponseDurationMs,
                     ModelId = x.Usage.UserModel.ModelId,
                     ModelName = x.Usage.UserModel.Model.Name,

src/BE/Controllers/OpenAICompatible/Dtos/ChatCompletionChunk.cs

@@ -5,7 +5,10 @@ namespace Chats.BE.Controllers.OpenAICompatible.Dtos;
 public record Delta
 {
     [JsonPropertyName("content")]
-    public required string Content { get; init; }
+    public required string? Content { get; init; }
+
+    [JsonPropertyName("reasoning_content"), JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
+    public required string? ReasoningContent { get; init; }
 }
 
 public record DeltaChoice

src/BE/Controllers/OpenAICompatible/Dtos/FullChatCompletion.cs

@@ -47,7 +47,10 @@ public record ResponseMessage
     public required string Role { get; init; }
 
     [JsonPropertyName("content")]
-    public required string Content { get; init; }
+    public required string? Content { get; init; }
+
+    [JsonPropertyName("reasoning_content"), JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
+    public required string? ReasoningContent { get; init; }
 
     [JsonPropertyName("refusal")]
     public object? Refusal { get; init; }

src/BE/Controllers/OpenAICompatible/OpenAICompatibleController.cs

@@ -48,7 +48,7 @@ public async Task<ActionResult> ChatCompletion([FromBody] JsonObject json, [From
         {
             await foreach (InternalChatSegment seg in icc.Run(userBalance.Balance, userModel, s.ChatStreamedSimulated(cco.Stream, [.. cco.Messages], cco.ToCleanCco(), cancellationToken)))
             {
-                if (seg.TextSegment == string.Empty) continue;
+                if (seg.Segment == string.Empty) continue;
 
                 if (cco.Stream)
                 {

src/BE/DB/ChatsDB.cs

@@ -73,6 +73,8 @@ public ChatsDB(DbContextOptions<ChatsDB> options)
 
     public virtual DbSet<Prompt> Prompts { get; set; }
 
+    public virtual DbSet<ReasoningResponseKind> ReasoningResponseKinds { get; set; }
+
     public virtual DbSet<SmsAttempt> SmsAttempts { get; set; }
 
     public virtual DbSet<SmsRecord> SmsRecords { get; set; }
@@ -330,6 +332,10 @@ protected override void OnModelCreating(ModelBuilder modelBuilder)
                 .OnDelete(DeleteBehavior.ClientSetNull)
                 .HasConstraintName("FK_ModelSetting_ModelProvider");
 
+            entity.HasOne(d => d.ReasoningResponseKind).WithMany(p => p.ModelReferences)
+                .OnDelete(DeleteBehavior.ClientSetNull)
+                .HasConstraintName("FK_ModelReference_ReasoningResponseKind");
+
             entity.HasOne(d => d.Tokenizer).WithMany(p => p.ModelReferences).HasConstraintName("FK_ModelReference_Tokenizer");
         });
 

src/BE/DB/Enums/DBMessageContentType.cs

@@ -6,17 +6,21 @@
 public enum DBMessageContentType : byte
 {
     /// <summary>
-    /// Error content type encoded in UTF-8.
+    /// Error content type, stored in MessageContentText table
     /// </summary>
     Error = 0,
 
     /// <summary>
-    /// Text content type encoded in Unicode.
+    /// Text content type, stored in MessageContentText table
     /// </summary>
     Text = 1,
 
     /// <summary>
-    /// 32bit little-endian integer file ID content type.
+    /// File ID content type, stored in MessageContentFile table
     /// </summary>
     FileId = 2,
+
+    /// <summary>
+    /// Reasoning content type, stored in MessageContentText table
+    Reasoning = 3,
 }

src/BE/DB/Enums/DBModelProvider.cs

@@ -19,4 +19,5 @@ public enum DBModelProvider
     Ollama = 14,
     MiniMax = 15,
     Doubao = 16, 
+    SiliconFlow = 17, 
 }

src/BE/DB/Enums/DBReasoningResponseKind.cs

@@ -0,0 +1,8 @@
+namespace Chats.BE.DB.Enums;
+
+public enum DBReasoningResponseKind : byte
+{
+    NoReasoningOutput = 0, 
+    ReasoningContent = 1, 
+    ThinkTag = 2, 
+}

src/BE/DB/Extensions/MessageContent.cs

@@ -1,5 +1,6 @@
 using Chats.BE.DB.Enums;
 using Chats.BE.Services.FileServices;
+using Chats.BE.Services.Models.Dtos;
 using OpenAI.Chat;
 
 namespace Chats.BE.DB;
@@ -22,16 +23,22 @@ public override string ToString()
         {
             DBMessageContentType.Text => MessageContentText!.Content,
             DBMessageContentType.Error => MessageContentText!.Content,
+            DBMessageContentType.Reasoning => MessageContentText!.Content,
             //DBMessageContentType.FileId => MessageContentUtil.ReadFileId(Content).ToString(), // not supported
             _ => throw new NotSupportedException(),
         };
     }
 
-    public static MessageContent FromText(string text)
+    public static MessageContent FromContent(string text)
     {
         return new MessageContent { MessageContentText = new() { Content = text }, ContentTypeId = (byte)DBMessageContentType.Text };
     }
 
+    public static MessageContent FromReasoningContent(string text)
+    {
+        return new MessageContent { MessageContentText = new() { Content = text }, ContentTypeId = (byte)DBMessageContentType.Reasoning };
+    }
+
     public static MessageContent FromFile(int fileId, File file)
     {
         return new MessageContent { MessageContentFile = new() { FileId = fileId, File = file }, ContentTypeId = (byte)DBMessageContentType.FileId };
@@ -41,4 +48,20 @@ public static MessageContent FromError(string error)
     {
         return new MessageContent { MessageContentText = new() { Content = error }, ContentTypeId = (byte)DBMessageContentType.Error };
     }
+
+    public static IEnumerable<MessageContent> FromFullResponse(InternalChatSegment lastSegment, string? errorText)
+    {
+        if (errorText is not null)
+        {
+            yield return FromError(errorText);
+        }
+        if (lastSegment.ReasoningSegment is not null)
+        {
+            yield return FromReasoningContent(lastSegment.ReasoningSegment);
+        }
+        if (lastSegment.Segment is not null)
+        {
+            yield return FromContent(lastSegment.Segment);
+        }
+    }
 }