@lobehub/chat 1.49.16 → 1.50.0

package/src/libs/agent-runtime/github/index.ts

@@ -2,7 +2,7 @@ import { LOBE_DEFAULT_MODEL_LIST } from '@/config/modelProviders';
 import type { ChatModelCard } from '@/types/llm';
 
 import { AgentRuntimeErrorType } from '../error';
-import { o1Models, pruneO1Payload } from '../openai';
+import { pruneReasoningPayload, reasoningModels } from '../openai';
 import { ModelProvider } from '../types';
 import {
   CHAT_MODELS_BLOCK_LIST,
@@ -37,8 +37,8 @@ export const LobeGithubAI = LobeOpenAICompatibleFactory({
     handlePayload: (payload) => {
       const { model } = payload;
 
-      if (o1Models.has(model)) {
-        return { ...pruneO1Payload(payload), stream: false } as any;
+      if (reasoningModels.has(model)) {
+        return { ...pruneReasoningPayload(payload), stream: false } as any;
       }
 
       return { ...payload, stream: payload.stream ?? true };

package/src/libs/agent-runtime/openai/index.ts

@@ -2,21 +2,23 @@ import { ChatStreamPayload, ModelProvider, OpenAIChatMessage } from '../types';
 import { LobeOpenAICompatibleFactory } from '../utils/openaiCompatibleFactory';
 
 // TODO: 临时写法，后续要重构成 model card 展示配置
-export const o1Models = new Set([
+export const reasoningModels = new Set([
   'o1-preview',
   'o1-preview-2024-09-12',
   'o1-mini',
   'o1-mini-2024-09-12',
   'o1',
   'o1-2024-12-17',
+  'o3-mini',
+  'o3-mini-2025-01-31',
 ]);
 
-export const pruneO1Payload = (payload: ChatStreamPayload) => ({
+export const pruneReasoningPayload = (payload: ChatStreamPayload) => ({
   ...payload,
   frequency_penalty: 0,
   messages: payload.messages.map((message: OpenAIChatMessage) => ({
     ...message,
-    role: message.role === 'system' ? 'user' : message.role,
+    role: message.role === 'system' ? 'developer' : message.role,
   })),
   presence_penalty: 0,
   temperature: 1,
@@ -29,8 +31,8 @@ export const LobeOpenAI = LobeOpenAICompatibleFactory({
     handlePayload: (payload) => {
       const { model } = payload;
 
-      if (o1Models.has(model)) {
-        return pruneO1Payload(payload) as any;
+      if (reasoningModels.has(model)) {
+        return pruneReasoningPayload(payload) as any;
       }
 
       return { ...payload, stream: payload.stream ?? true };

package/src/libs/agent-runtime/utils/streams/openai.test.ts

@@ -754,6 +754,7 @@ describe('OpenAIStream', () => {
         ].map((i) => `${i}\n`),
       );
     });
+
     it('should handle reasoning in litellm', async () => {
       const data = [
         {
@@ -954,5 +955,206 @@ describe('OpenAIStream', () => {
         ].map((i) => `${i}\n`),
       );
     });
+
+    it('should handle reasoning in siliconflow', async () => {
+      const data = [
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { role: 'assistant', reasoning_content: '', content: '' },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { reasoning_content: '您好', content: '' },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { reasoning_content: '！', content: '' },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { content: '你好', reasoning_content: null },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { content: '很高兴', reasoning_cont: null },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { content: '为您', reasoning_content: null },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { content: '提供', reasoning_content: null },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { content: '帮助。', reasoning_content: null },
+              logprobs: null,
+              finish_reason: null,
+            },
+          ],
+        },
+        {
+          id: '1',
+          object: 'chat.completion.chunk',
+          created: 1737563070,
+          model: 'deepseek-reasoner',
+          system_fingerprint: 'fp_1c5d8833bc',
+          choices: [
+            {
+              index: 0,
+              delta: { content: '', reasoning_content: null },
+              logprobs: null,
+              finish_reason: 'stop',
+            },
+          ],
+          usage: {
+            prompt_tokens: 6,
+            completion_tokens: 104,
+            total_tokens: 110,
+            prompt_tokens_details: { cached_tokens: 0 },
+            completion_tokens_details: { reasoning_tokens: 70 },
+            prompt_cache_hit_tokens: 0,
+            prompt_cache_miss_tokens: 6,
+          },
+        },
+      ];
+
+      const mockOpenAIStream = new ReadableStream({
+        start(controller) {
+          data.forEach((chunk) => {
+            controller.enqueue(chunk);
+          });
+
+          controller.close();
+        },
+      });
+
+      const protocolStream = OpenAIStream(mockOpenAIStream);
+
+      const decoder = new TextDecoder();
+      const chunks = [];
+
+      // @ts-ignore
+      for await (const chunk of protocolStream) {
+        chunks.push(decoder.decode(chunk, { stream: true }));
+      }
+
+      expect(chunks).toEqual(
+        [
+          'id: 1',
+          'event: reasoning',
+          `data: ""\n`,
+          'id: 1',
+          'event: reasoning',
+          `data: "您好"\n`,
+          'id: 1',
+          'event: reasoning',
+          `data: "！"\n`,
+          'id: 1',
+          'event: text',
+          `data: "你好"\n`,
+          'id: 1',
+          'event: text',
+          `data: "很高兴"\n`,
+          'id: 1',
+          'event: text',
+          `data: "为您"\n`,
+          'id: 1',
+          'event: text',
+          `data: "提供"\n`,
+          'id: 1',
+          'event: text',
+          `data: "帮助。"\n`,
+          'id: 1',
+          'event: stop',
+          `data: "stop"\n`,
+        ].map((i) => `${i}\n`),
+      );
+    });
   });
 });

package/src/libs/agent-runtime/utils/streams/openai.ts

@@ -37,9 +37,8 @@ export const transformOpenAIStream = (
     return { data: errorData, id: 'first_chunk_error', type: 'error' };
   }
 
-  // maybe need another structure to add support for multiple choices
-
   try {
+    // maybe need another structure to add support for multiple choices
     const item = chunk.choices[0];
     if (!item) {
       return { data: chunk, id: chunk.id, type: 'data' };
@@ -88,12 +87,10 @@ export const transformOpenAIStream = (
       return { data: item.finish_reason, id: chunk.id, type: 'stop' };
     }
 
-    if (typeof item.delta?.content === 'string') {
-      return { data: item.delta.content, id: chunk.id, type: 'text' };
-    }
-
-    // DeepSeek reasoner 会将 thinking 放在 reasoning_content 字段中
-    // litellm 处理 reasoning content 时 不会设定 content = null
+    // DeepSeek reasoner will put thinking in the reasoning_content field
+    // litellm will not set content = null when processing reasoning content
+    // en: siliconflow has encountered a situation where both content and reasoning_content are present, so the parsing order go ahead
+    // refs: https://github.com/lobehub/lobe-chat/issues/5681
     if (
       item.delta &&
       'reasoning_content' in item.delta &&
@@ -102,6 +99,10 @@ export const transformOpenAIStream = (
       return { data: item.delta.reasoning_content, id: chunk.id, type: 'reasoning' };
     }
 
+    if (typeof item.delta?.content === 'string') {
+      return { data: item.delta.content, id: chunk.id, type: 'text' };
+    }
+
     // 无内容情况
     if (item.delta && item.delta.content === null) {
       return { data: item.delta, id: chunk.id, type: 'data' };

package/src/locales/default/discover.ts

@@ -127,6 +127,10 @@ export default {
         title: '话题新鲜度',
       },
       range: '范围',
+      reasoning_effort: {
+        desc: '此设置用于控制模型在生成回答前的推理强度。低强度优先响应速度并节省 Token，高强度提供更完整的推理，但会消耗更多 Token 并降低响应速度。默认值为中，平衡推理准确性与响应速度。',
+        title: '推理强度',
+      },
       temperature: {
         desc: '此设置影响模型回应的多样性。较低的值会导致更可预测和典型的回应，而较高的值则鼓励更多样化和不常见的回应。当值设为0时，模型对于给定的输入总是给出相同的回应。',
         title: '随机性',

package/src/locales/default/setting.ts

@@ -202,6 +202,9 @@ export default {
     enableMaxTokens: {
       title: '开启单次回复限制',
     },
+    enableReasoningEffort: {
+      title: '开启推理强度调整',
+    },
     frequencyPenalty: {
       desc: '值越大，越有可能降低重复字词',
       title: '频率惩罚度',
@@ -218,6 +221,15 @@ export default {
       desc: '值越大，越有可能扩展到新话题',
       title: '话题新鲜度',
     },
+    reasoningEffort: {
+      desc: '值越大，推理能力越强，但可能会增加响应时间和 Token 消耗',
+      options: {
+        high: '高',
+        low: '低',
+        medium: '中',
+      },
+      title: '推理强度',
+    },
     temperature: {
       desc: '值越大，回复越随机',
       title: '随机性',

package/src/store/chat/slices/aiChat/actions/generateAIChat.ts

@@ -421,6 +421,11 @@ export const generateAIChat: StateCreator<
       ? agentConfig.params.max_tokens
       : undefined;
 
+    // 5. handle reasoning_effort
+    agentConfig.params.reasoning_effort = chatConfig.enableReasoningEffort
+      ? agentConfig.params.reasoning_effort
+      : undefined;
+
     let isFunctionCall = false;
     let msgTraceId: string | undefined;
     let output = '';

package/src/types/agent/index.ts

@@ -68,6 +68,11 @@ export interface LobeAgentChatConfig {
   enableHistoryCount?: boolean;
   enableMaxTokens?: boolean;
 
+  /**
+   * 自定义推理强度
+   */
+  enableReasoningEffort?: boolean;
+
   /**
    * 历史消息条数
    */
@@ -82,6 +87,7 @@ export const AgentChatConfigSchema = z.object({
   enableCompressHistory: z.boolean().optional(),
   enableHistoryCount: z.boolean().optional(),
   enableMaxTokens: z.boolean().optional(),
+  enableReasoningEffort: z.boolean().optional(),
   historyCount: z.number().optional(),
 });
 

package/src/types/llm.ts

@@ -159,6 +159,11 @@ export interface LLMParams {
    * 生成文本的随机度量，用于控制文本的创造性和多样性
    * @default 1
    */
+  reasoning_effort?: string;
+  /**
+   * 控制模型推理能力
+   * @default medium
+   */
   temperature?: number;
   /**
    * 控制生成文本中最高概率的单个 token