Merge pull request #87 from cnblogs/support-deepseek-reasoning

feat: support reasoning models

Author: ikesnowy

README.md

@@ -47,7 +47,7 @@ builder.AddDashScopeClient(builder.Configuration);
 ```json
 {
     "DashScope": {
-        "ApiKey": "your-api-key"
+        "ApiKey": "your-api-key",
     }
 }
 ```
@@ -66,21 +66,54 @@ public class YourService(IDashScopeClient client)
 
 # Supported APIs
 
-- Text Embedding API - `dashScopeClient.GetTextEmbeddingsAsync()`
-- Text Generation API(qwen-turbo, qwen-max, etc.) - `dashScopeClient.GetQwenCompletionAsync()` and `dashScopeClient.GetQWenCompletionStreamAsync()`
-- BaiChuan Models - Use `dashScopeClient.GetBaiChuanTextCompletionAsync()`
-- LLaMa2 Models - `dashScopeClient.GetLlama2TextCompletionAsync()`
-- Multimodal Generation API(qwen-vl-max, etc.) - `dashScopeClient.GetQWenMultimodalCompletionAsync()` and `dashScopeClient.GetQWenMultimodalCompletionStreamAsync()`
+- Text Embedding API - `GetTextEmbeddingsAsync()`
+- Text Generation API(qwen-turbo, qwen-max, etc.) - `GetQWenCompletionAsync()` and `GetQWenCompletionStreamAsync()`
+- DeepSeek Models - `GetDeepSeekCompletionAsync()` and `GetDeepSeekCompletionStreamAsync()`
+- BaiChuan Models - Use `GetBaiChuanTextCompletionAsync()`
+- LLaMa2 Models - `GetLlama2TextCompletionAsync()`
+- Multimodal Generation API(qwen-vl-max, etc.) - `GetQWenMultimodalCompletionAsync()` and `GetQWenMultimodalCompletionStreamAsync()`
 - Wanx Models(Image generation, background generation, etc)
   - Image Synthesis - `CreateWanxImageSynthesisTaskAsync()` and `GetWanxImageSynthesisTaskAsync()`
   - Image Generation - `CreateWanxImageGenerationTaskAsync()` and `GetWanxImageGenerationTaskAsync()`
   - Background Image Generation - `CreateWanxBackgroundGenerationTaskAsync()` and `GetWanxBackgroundGenerationTaskAsync()`
-- File API that used by Qwen-Long - `dashScopeClient.UploadFileAsync()` and `dashScopeClient.DeleteFileAsync`
+- File API that used by Qwen-Long - `UploadFileAsync()` and `DeleteFileAsync`
 - Application call - `GetApplicationResponseAsync()` and `GetApplicationResponseStreamAsync()`
 
 # Examples
 
-Visit [tests](./test) for more usage of each api.
+Visit [snapshots](./test/Cnblogs.DashScope.Sdk.UnitTests/Utils/Snapshots.cs) for calling samples.
+
+Visit [tests](./test/Cnblogs.DashScope.Sdk.UnitTests) for more usage of each api.
+
+## General Text Completion API
+
+Use `client.GetTextCompletionAsync` and  `client.GetTextCompletionStreamAsync` to access text generation api directly.
+
+```csharp
+var completion = await dashScopeClient.GetTextCompletionAsync(
+            new ModelRequest<TextGenerationInput, ITextGenerationParameters>
+            {
+                Model = "your-model-name",
+                Input = new TextGenerationInput { Prompt = prompt },
+                Parameters = new TextGenerationParameters()
+                {
+                    // control parameters as you wish.
+                    EnableSearch = true
+                }
+            });
+var completions = dashScopeClient.GetTextCompletionStreamAsync(
+            new ModelRequest<TextGenerationInput, ITextGenerationParameters>
+            {
+                Model = "your-model-name",
+                Input = new TextGenerationInput { Messages = [TextChatMessage.System("you are a helpful assistant"), TextChatMessage.User("How are you?")] },
+                Parameters = new TextGenerationParameters()
+                {
+                    // control parameters as you wish.
+                    EnableSearch = true,
+                    IncreamentalOutput = true
+                }
+            });
+```
 
 ## Single Text Completion
 
@@ -90,6 +123,19 @@ var completion = await client.GetQWenCompletionAsync(QWenLlm.QWenMax, prompt);
 Console.WriteLine(completion.Output.Text);
 ```
 
+## Reasoning
+
+Use `completion.Output.Choices![0].Message.ReasoningContent` to access the reasoning content from model.
+
+```csharp
+var history = new List<ChatMessage>
+{
+    ChatMessage.User("Calculate 1+1")
+};
+var completion = await client.GetDeepSeekChatCompletionAsync(DeepSeekLlm.DeepSeekR1, history);
+Console.WriteLine(completion.Output.Choices[0]!.Message.ReasoningContent);
+```
+
 ## Multi-round chat
 
 ```csharp

README.zh-Hans.md

@@ -66,23 +66,59 @@ public class YourService(IDashScopeClient client)
 
 # 支持的 API
 
-- 通用文本向量 - `dashScopeClient.GetTextEmbeddingsAsync()`
-- 通义千问（`qwen-turbo`， `qwen-max` 等） - `dashScopeClient.GetQwenCompletionAsync()` and `dashScopeClient.GetQWenCompletionStreamAsync()`
-- 百川开源大模型 - Use `dashScopeClient.GetBaiChuanTextCompletionAsync()`
-- LLaMa2 大语言模型 - `dashScopeClient.GetLlama2TextCompletionAsync()`
-- 通义千问 VL 和通义千问 Audio(`qwen-vl-max`， `qwen-audio`) - `dashScopeClient.GetQWenMultimodalCompletionAsync()` and `dashScopeClient.GetQWenMultimodalCompletionStreamAsync()`
+- 通用文本向量 - `GetTextEmbeddingsAsync()`
+- 通义千问（`qwen-turbo`， `qwen-max` 等） - `GetQWenCompletionAsync()` 和 `GetQWenCompletionStreamAsync()`
+- DeepSeek 系列模型（`deepseek-r1`，`deepseek-v3` 等） - `GetDeepSeekChatCompletionAsync()` 和 `GetDeepSeekChatCompletionStreamAsync()`
+- 百川开源大模型 - `GetBaiChuanTextCompletionAsync()`
+- LLaMa2 大语言模型 - `GetLlama2TextCompletionAsync()`
+- 通义千问 VL 和通义千问 Audio(`qwen-vl-max`， `qwen-audio`) - `GetQWenMultimodalCompletionAsync()` 和 `GetQWenMultimodalCompletionStreamAsync()`
 - 通义万相系列
-    - 文生图 - `CreateWanxImageSynthesisTaskAsync()` and `GetWanxImageSynthesisTaskAsync()`
-    - 人像风格重绘 - `CreateWanxImageGenerationTaskAsync()` and `GetWanxImageGenerationTaskAsync()`
-    - 图像背景生成 - `CreateWanxBackgroundGenerationTaskAsync()` and `GetWanxBackgroundGenerationTaskAsync()`
-- 适用于 QWen-Long 的文件 API `dashScopeClient.UploadFileAsync()` and `dashScopeClient.DeleteFileAsync`
-- 应用调用 `dashScopeClient.GetApplicationResponseAsync` 和 `dashScopeClient.GetApplicationResponseStreamAsync()`
+    - 文生图 - `CreateWanxImageSynthesisTaskAsync()` 和 `GetWanxImageSynthesisTaskAsync()`
+    - 人像风格重绘 - `CreateWanxImageGenerationTaskAsync()` 和 `GetWanxImageGenerationTaskAsync()`
+    - 图像背景生成 - `CreateWanxBackgroundGenerationTaskAsync()` 和 `GetWanxBackgroundGenerationTaskAsync()`
+- 适用于 QWen-Long 的文件 API `UploadFileAsync()` 和 `DeleteFileAsync`
+- 应用调用 `GetApplicationResponseAsync` 和 `GetApplicationResponseStreamAsync()`
 - 其他使用相同 Endpoint 的模型
 
 # 示例
 
+查看 [Snapshots.cs](./test/Cnblogs.DashScope.Sdk.UnitTests/Utils/Snapshots.cs) 获得 API 调用参数示例.
+
 查看 [测试](./test) 获得更多 API 使用示例。
 
+## 文本生成
+
+使用 `dashScopeClient.GetTextCompletionAsync` 和 `dashScopeClient.GetTextCompletionStreamAsync` 来直接访问文本生成接口。
+
+相关文档：https://help.aliyun.com/zh/model-studio/user-guide/text-generation/
+
+```csharp
+var completion = await dashScopeClient.GetTextCompletionAsync(
+            new ModelRequest<TextGenerationInput, ITextGenerationParameters>
+            {
+                Model = "your-model-name",
+                Input = new TextGenerationInput { Prompt = prompt },
+                Parameters = new TextGenerationParameters()
+                {
+                    // control parameters as you wish.
+                    EnableSearch = true
+                }
+            });
+
+var completions = dashScopeClient.GetTextCompletionStreamAsync(
+            new ModelRequest<TextGenerationInput, ITextGenerationParameters>
+            {
+                Model = "your-model-name",
+                Input = new TextGenerationInput { Messages = [TextChatMessage.System("you are a helpful assistant"), TextChatMessage.User("How are you?")] },
+                Parameters = new TextGenerationParameters()
+                {
+                    // control parameters as you wish.
+                    EnableSearch = true,
+                    IncreamentalOutput = true
+                }
+            });
+```
+
 ## 单轮对话
 
 ```csharp
@@ -108,6 +144,19 @@ var completion = await client.GetQWenChatCompletionAsync(QWenLlm.QWenMax, histor
 Console.WriteLine(completion.Output.Choices[0].Message.Content); // The number is 42
 ```
 
+## 推理
+
+使用推理模型时，模型的思考过程可以通过 `ReasoningContent` 属性获取。
+
+```csharp
+var history = new List<TextChatMessage>
+{
+    TextChatMessage.User("Calculate 1+1")
+};
+var completion = await client.GetDeepSeekChatCompletionAsync(DeepSeekLlm.DeepSeekR1, history);
+Console.WriteLine(completion.Output.Choices[0]!.Message.ReasoningContent);
+```
+
 ## 工具调用
 
 创建一个可供模型使用的方法。

src/Cnblogs.DashScope.AI/DashScopeChatClient.cs

@@ -183,7 +183,7 @@ public async IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
                 {
                     update.Contents.Add(
                         new UsageContent(
-                            new UsageDetails()
+                            new UsageDetails
                             {
                                 InputTokenCount = response.Usage.InputTokens,
                                 OutputTokenCount = response.Usage.OutputTokens,
@@ -208,7 +208,7 @@ public async IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
                     RawRepresentation = completion.Messages[0].RawRepresentation,
                     CreatedAt = completion.CreatedAt,
                     FinishReason = completion.FinishReason,
-                    ModelId = completion.ModelId,
+                    ModelId = completion.ModelId
                 };
             }
             else
@@ -467,6 +467,7 @@ private IEnumerable<TextChatMessage> ToTextChatMessages(
                 from.Text,
                 from.AuthorName,
                 null,
+                null,
                 functionCall.Count > 0 ? functionCall : null);
         }
     }

src/Cnblogs.DashScope.Core/TextChatMessage.cs

@@ -10,13 +10,15 @@ namespace Cnblogs.DashScope.Core;
 /// <param name="Content">The content of this message.</param>
 /// <param name="Name">Used when role is tool, represents the function name of this message generated by.</param>
 /// <param name="Partial">Notify model that next message should use this message as prefix.</param>
+/// <param name="ReasoningContent">Reasoning content for reasoning model.</param>
 /// <param name="ToolCalls">Calls to the function.</param>
 [method: JsonConstructor]
 public record TextChatMessage(
     string Role,
     string Content,
     string? Name = null,
     bool? Partial = null,
+    string? ReasoningContent = null,
     List<ToolCall>? ToolCalls = null) : IMessage<string>
 {
     /// <summary>
@@ -84,11 +86,17 @@ public static TextChatMessage System(string content)
     /// <param name="content">The content of the message.</param>
     /// <param name="partial">When set to true, content of this message would be the prefix of next model output.</param>
     /// <param name="name">Author name.</param>
+    /// <param name="reasoningContent">Think content when reasoning.</param>
     /// <param name="toolCalls">Tool calls by model.</param>
     /// <returns></returns>
-    public static TextChatMessage Assistant(string content, bool? partial = null, string? name = null, List<ToolCall>? toolCalls = null)
+    public static TextChatMessage Assistant(
+        string content,
+        bool? partial = null,
+        string? name = null,
+        string? reasoningContent = null,
+        List<ToolCall>? toolCalls = null)
     {
-        return new TextChatMessage(DashScopeRoleNames.Assistant, content, name, partial, toolCalls);
+        return new TextChatMessage(DashScopeRoleNames.Assistant, content, name, partial, reasoningContent, toolCalls);
     }
 
     /// <summary>

src/Cnblogs.DashScope.Sdk/DeepSeek/DeepSeekLlm.cs

@@ -0,0 +1,17 @@
+namespace Cnblogs.DashScope.Sdk.DeepSeek;
+
+/// <summary>
+/// DeepSeek models.
+/// </summary>
+public enum DeepSeekLlm
+{
+    /// <summary>
+    /// deepseek-v3 model.
+    /// </summary>
+    DeepSeekV3 = 1,
+
+    /// <summary>
+    /// deepseek-r1 model.
+    /// </summary>
+    DeepSeekR1 = 2
+}

src/Cnblogs.DashScope.Sdk/DeepSeek/DeepSeekLlmName.cs

@@ -0,0 +1,14 @@
+namespace Cnblogs.DashScope.Sdk.DeepSeek;
+
+internal static class DeepSeekLlmName
+{
+    public static string GetModelName(this DeepSeekLlm model)
+    {
+        return model switch
+        {
+            DeepSeekLlm.DeepSeekR1 => "deepseek-r1",
+            DeepSeekLlm.DeepSeekV3 => "deepseek-v3",
+            _ => ThrowHelper.UnknownModelName(nameof(model), model)
+        };
+    }
+}

src/Cnblogs.DashScope.Sdk/DeepSeek/DeepSeekTextGenerationApi.cs

@@ -0,0 +1,87 @@
+using Cnblogs.DashScope.Core;
+
+namespace Cnblogs.DashScope.Sdk.DeepSeek;
+
+/// <summary>
+/// Extensions for calling DeepSeek models, see: https://help.aliyun.com/zh/model-studio/developer-reference/deepseek
+/// </summary>
+public static class DeepSeekTextGenerationApi
+{
+    private static TextGenerationParameters StreamingParameters { get; } = new() { IncrementalOutput = true };
+
+    /// <summary>
+    /// Get text completion from deepseek model.
+    /// </summary>
+    /// <param name="client">The <see cref="IDashScopeClient"/>.</param>
+    /// <param name="model">The model name.</param>
+    /// <param name="messages">The context messages.</param>
+    /// <returns></returns>
+    public static async Task<ModelResponse<TextGenerationOutput, TextGenerationTokenUsage>>
+        GetDeepSeekChatCompletionAsync(
+            this IDashScopeClient client,
+            DeepSeekLlm model,
+            IEnumerable<TextChatMessage> messages)
+    {
+        return await client.GetDeepSeekChatCompletionAsync(model.GetModelName(), messages);
+    }
+
+    /// <summary>
+    /// Get text completion from deepseek model.
+    /// </summary>
+    /// <param name="client">The <see cref="IDashScopeClient"/>.</param>
+    /// <param name="model">The model name.</param>
+    /// <param name="messages">The context messages.</param>
+    /// <returns></returns>
+    public static async Task<ModelResponse<TextGenerationOutput, TextGenerationTokenUsage>>
+        GetDeepSeekChatCompletionAsync(
+            this IDashScopeClient client,
+            string model,
+            IEnumerable<TextChatMessage> messages)
+    {
+        return await client.GetTextCompletionAsync(
+            new ModelRequest<TextGenerationInput, ITextGenerationParameters>
+            {
+                Model = model,
+                Input = new TextGenerationInput { Messages = messages },
+                Parameters = null
+            });
+    }
+
+    /// <summary>
+    /// Get streamed completion from deepseek model.
+    /// </summary>
+    /// <param name="client"></param>
+    /// <param name="model"></param>
+    /// <param name="messages"></param>
+    /// <returns></returns>
+    public static IAsyncEnumerable<ModelResponse<TextGenerationOutput, TextGenerationTokenUsage>>
+        GetDeepSeekChatCompletionStreamAsync(
+            this IDashScopeClient client,
+            DeepSeekLlm model,
+            IEnumerable<TextChatMessage> messages)
+    {
+        return client.GetDeepSeekChatCompletionStreamAsync(model.GetModelName(), messages);
+    }
+
+    /// <summary>
+    /// Get streamed completion from deepseek model.
+    /// </summary>
+    /// <param name="client"></param>
+    /// <param name="model"></param>
+    /// <param name="messages"></param>
+    /// <returns></returns>
+    public static IAsyncEnumerable<ModelResponse<TextGenerationOutput, TextGenerationTokenUsage>>
+        GetDeepSeekChatCompletionStreamAsync(
+            this IDashScopeClient client,
+            string model,
+            IEnumerable<TextChatMessage> messages)
+    {
+        return client.GetTextCompletionStreamAsync(
+            new ModelRequest<TextGenerationInput, ITextGenerationParameters>
+            {
+                Model = model,
+                Input = new TextGenerationInput { Messages = messages },
+                Parameters = StreamingParameters
+            });
+    }
+}

src/Cnblogs.DashScope.Sdk/QWen/QWenLlm.cs

@@ -133,5 +133,10 @@ public enum QWenLlm
     /// <summary>
     /// qvq-72b-preview
     /// </summary>
-    QwQ72BPreview = 25
+    QvQ72BPreview = 25,
+
+    /// <summary>
+    /// qwq-32b
+    /// </summary>
+    QwQ32B = 26
 }

src/Cnblogs.DashScope.Sdk/QWen/QWenLlmNames.cs

@@ -30,7 +30,8 @@ public static string GetModelName(this QWenLlm llm)
             QWenLlm.QWenPlusLatest => "qwen-plus-latest",
             QWenLlm.QWenTurboLatest => "qwen-turbo-latest",
             QWenLlm.QwQ32BPreview => "qwq-32b-preview",
-            QWenLlm.QwQ72BPreview => "qwq-72b-preview",
+            QWenLlm.QvQ72BPreview => "qvq-72b-preview",
+            QWenLlm.QwQ32B => "qwq-32b",
             _ => ThrowHelper.UnknownModelName(nameof(llm), llm)
         };
     }