@lobehub/chat 1.84.25 → 1.84.27

package/CHANGELOG.md

@@ -2,6 +2,56 @@
 
 # Changelog
 
+### [Version 1.84.27](https://github.com/lobehub/lobe-chat/compare/v1.84.26...v1.84.27)
+
+<sup>Released on **2025-05-09**</sup>
+
+#### 💄 Styles
+
+- **misc**: Add reasoning tokens and token usage statistics for Google Gemini.
+
+<br/>
+
+<details>
+<summary><kbd>Improvements and Fixes</kbd></summary>
+
+#### Styles
+
+- **misc**: Add reasoning tokens and token usage statistics for Google Gemini, closes [#7501](https://github.com/lobehub/lobe-chat/issues/7501) ([b466b42](https://github.com/lobehub/lobe-chat/commit/b466b42))
+
+</details>
+
+<div align="right">
+
+[![](https://img.shields.io/badge/-BACK_TO_TOP-151515?style=flat-square)](#readme-top)
+
+</div>
+
+### [Version 1.84.26](https://github.com/lobehub/lobe-chat/compare/v1.84.25...v1.84.26)
+
+<sup>Released on **2025-05-08**</sup>
+
+#### 💄 Styles
+
+- **misc**: Add qwen3 for ollama.
+
+<br/>
+
+<details>
+<summary><kbd>Improvements and Fixes</kbd></summary>
+
+#### Styles
+
+- **misc**: Add qwen3 for ollama, closes [#7746](https://github.com/lobehub/lobe-chat/issues/7746) ([806d905](https://github.com/lobehub/lobe-chat/commit/806d905))
+
+</details>
+
+<div align="right">
+
+[![](https://img.shields.io/badge/-BACK_TO_TOP-151515?style=flat-square)](#readme-top)
+
+</div>
+
 ### [Version 1.84.25](https://github.com/lobehub/lobe-chat/compare/v1.84.24...v1.84.25)
 
 <sup>Released on **2025-05-08**</sup>

package/changelog/v1.json

@@ -1,4 +1,22 @@
 [
+  {
+    "children": {
+      "improvements": [
+        "Add reasoning tokens and token usage statistics for Google Gemini."
+      ]
+    },
+    "date": "2025-05-09",
+    "version": "1.84.27"
+  },
+  {
+    "children": {
+      "improvements": [
+        "Add qwen3 for ollama."
+      ]
+    },
+    "date": "2025-05-08",
+    "version": "1.84.26"
+  },
   {
     "children": {
       "fixes": [

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lobehub/chat",
-  "version": "1.84.25",
+  "version": "1.84.27",
   "description": "Lobe Chat - an open-source, high-performance chatbot framework that supports speech synthesis, multimodal, and extensible Function Call plugin system. Supports one-click free deployment of your private ChatGPT/LLM web application.",
   "keywords": [
     "framework",
@@ -121,7 +121,7 @@
   "dependencies": {
     "@ant-design/icons": "^5.6.1",
     "@ant-design/pro-components": "^2.8.7",
-    "@anthropic-ai/sdk": "^0.39.0",
+    "@anthropic-ai/sdk": "^0.41.0",
     "@auth/core": "^0.38.0",
     "@aws-sdk/client-bedrock-runtime": "^3.779.0",
     "@aws-sdk/client-s3": "^3.779.0",
@@ -199,7 +199,7 @@
     "langfuse": "^3.37.1",
     "langfuse-core": "^3.37.1",
     "lodash-es": "^4.17.21",
-    "lucide-react": "^0.503.0",
+    "lucide-react": "^0.508.0",
     "mammoth": "^1.9.0",
     "mdast-util-to-markdown": "^2.1.2",
     "modern-screenshot": "^4.6.0",

package/src/config/aiModels/ollama.ts

@@ -89,11 +89,22 @@ const ollamaChatModels: AIChatModelCard[] = [
     description:
       'QwQ 是 Qwen 系列的推理模型。与传统的指令调优模型相比，QwQ 具备思考和推理的能力，能够在下游任务中，尤其是困难问题上，显著提升性能。QwQ-32B 是中型推理模型，能够在与最先进的推理模型（如 DeepSeek-R1、o1-mini）竞争时取得可观的表现。',
     displayName: 'QwQ 32B',
-    enabled: true,
     id: 'qwq',
     releasedAt: '2024-11-28',
     type: 'chat',
   },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    contextWindowTokens: 65_536,
+    description: 'Qwen3 是阿里巴巴的新一代大规模语言模型，以优异的性能支持多元化的应用需求。',
+    displayName: 'Qwen3 7B',
+    enabled: true,
+    id: 'qwen3',
+    type: 'chat',
+  },
+
   {
     contextWindowTokens: 128_000,
     description: 'Qwen2.5 是阿里巴巴的新一代大规模语言模型，以优异的性能支持多元化的应用需求。',
@@ -115,7 +126,6 @@ const ollamaChatModels: AIChatModelCard[] = [
     contextWindowTokens: 128_000,
     description: 'Qwen2.5 是阿里巴巴的新一代大规模语言模型，以优异的性能支持多元化的应用需求。',
     displayName: 'Qwen2.5 7B',
-    enabled: true,
     id: 'qwen2.5',
     type: 'chat',
   },

package/src/config/aiModels/vertexai.ts

@@ -98,9 +98,9 @@ const vertexaiChatModels: AIChatModelCard[] = [
     type: 'chat',
   },
   {
-    abilities: { 
-      functionCall: true, 
-      vision: true 
+    abilities: {
+      functionCall: true,
+      vision: true
     },
     contextWindowTokens: 1_000_000 + 8192,
     description: 'Gemini 1.5 Flash 002 是一款高效的多模态模型，支持广泛应用的扩展。',
@@ -115,9 +115,9 @@ const vertexaiChatModels: AIChatModelCard[] = [
     type: 'chat',
   },
   {
-    abilities: { 
-      functionCall: true, 
-      vision: true 
+    abilities: {
+      functionCall: true,
+      vision: true
     },
     contextWindowTokens: 2_000_000 + 8192,
     description:

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

@@ -106,6 +106,7 @@ export class LobeGoogleAI implements LobeRuntimeAI {
 
       const contents = await this.buildGoogleMessages(payload.messages);
 
+      const inputStartAt = Date.now();
       const geminiStreamResult = await this.client
         .getGenerativeModel(
           {
@@ -161,7 +162,7 @@ export class LobeGoogleAI implements LobeRuntimeAI {
 
       // Convert the response into a friendly text-stream
       const Stream = this.isVertexAi ? VertexAIStream : GoogleGenerativeAIStream;
-      const stream = Stream(prod, options?.callback);
+      const stream = Stream(prod, { callbacks: options?.callback, inputStartAt });
 
       // Respond with the stream
       return StreamingResponse(stream, { headers: options?.headers });

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

@@ -34,10 +34,12 @@ describe('GoogleGenerativeAIStream', () => {
     const onCompletionMock = vi.fn();
 
     const protocolStream = GoogleGenerativeAIStream(mockGoogleStream, {
-      onStart: onStartMock,
-      onText: onTextMock,
-      onToolsCalling: onToolCallMock,
-      onCompletion: onCompletionMock,
+      callbacks: {
+        onStart: onStartMock,
+        onText: onTextMock,
+        onToolsCalling: onToolCallMock,
+        onCompletion: onCompletionMock,
+      },
     });
 
     const decoder = new TextDecoder();
@@ -187,7 +189,7 @@ describe('GoogleGenerativeAIStream', () => {
       // usage
       'id: chat_1\n',
       'event: usage\n',
-      `data: {"inputImageTokens":258,"inputTextTokens":8,"totalInputTokens":266,"totalTokens":266}\n\n`,
+      `data: {"inputImageTokens":258,"inputTextTokens":8,"outputTextTokens":0,"totalInputTokens":266,"totalOutputTokens":0,"totalTokens":266}\n\n`,
     ]);
   });
 
@@ -276,7 +278,100 @@ describe('GoogleGenerativeAIStream', () => {
         // usage
         'id: chat_1',
         'event: usage',
-        `data: {"inputTextTokens":19,"totalInputTokens":19,"totalOutputTokens":11,"totalTokens":30}\n`,
+        `data: {"inputTextTokens":19,"outputTextTokens":11,"totalInputTokens":19,"totalOutputTokens":11,"totalTokens":30}\n`,
+      ].map((i) => i + '\n'),
+    );
+  });
+
+  it('should handle stop with content and thought', async () => {
+    vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('1');
+
+    const data = [
+      {
+        candidates: [
+          {
+            content: { parts: [{ text: '234' }], role: 'model' },
+            safetyRatings: [
+              { category: 'HARM_CATEGORY_HATE_SPEECH', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_DANGEROUS_CONTENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_HARASSMENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT', probability: 'NEGLIGIBLE' },
+            ],
+          },
+        ],
+        text: () => '234',
+        usageMetadata: {
+          promptTokenCount: 19,
+          candidatesTokenCount: 3,
+          totalTokenCount: 122,
+          promptTokensDetails: [{ modality: 'TEXT', tokenCount: 19 }],
+          thoughtsTokenCount: 100,
+        },
+        modelVersion: 'gemini-2.0-flash-exp-image-generation',
+      },
+      {
+        text: () => '567890\n',
+        candidates: [
+          {
+            content: { parts: [{ text: '567890\n' }], role: 'model' },
+            finishReason: 'STOP',
+            safetyRatings: [
+              { category: 'HARM_CATEGORY_HATE_SPEECH', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_DANGEROUS_CONTENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_HARASSMENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT', probability: 'NEGLIGIBLE' },
+            ],
+          },
+        ],
+        usageMetadata: {
+          promptTokenCount: 19,
+          candidatesTokenCount: 11,
+          totalTokenCount: 131,
+          promptTokensDetails: [{ modality: 'TEXT', tokenCount: 19 }],
+          candidatesTokensDetails: [{ modality: 'TEXT', tokenCount: 11 }],
+          thoughtsTokenCount: 100,
+        },
+        modelVersion: 'gemini-2.0-flash-exp-image-generation',
+      },
+    ];
+
+    const mockGoogleStream = new ReadableStream({
+      start(controller) {
+        data.forEach((item) => {
+          controller.enqueue(item);
+        });
+
+        controller.close();
+      },
+    });
+
+    const protocolStream = GoogleGenerativeAIStream(mockGoogleStream);
+
+    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: chat_1',
+        'event: text',
+        'data: "234"\n',
+
+        'id: chat_1',
+        'event: text',
+        `data: "567890\\n"\n`,
+        // stop
+        'id: chat_1',
+        'event: stop',
+        `data: "STOP"\n`,
+        // usage
+        'id: chat_1',
+        'event: usage',
+        `data: {"inputTextTokens":19,"outputReasoningTokens":100,"outputTextTokens":11,"totalInputTokens":19,"totalOutputTokens":111,"totalTokens":131}\n`,
       ].map((i) => i + '\n'),
     );
   });

package/src/libs/agent-runtime/utils/streams/google-ai.ts

@@ -11,6 +11,7 @@ import {
   StreamToolCallChunkData,
   createCallbacksTransformer,
   createSSEProtocolTransformer,
+  createTokenSpeedCalculator,
   generateToolCallId,
 } from './protocol';
 
@@ -19,31 +20,62 @@ const transformGoogleGenerativeAIStream = (
   context: StreamContext,
 ): StreamProtocolChunk | StreamProtocolChunk[] => {
   // maybe need another structure to add support for multiple choices
+  const candidate = chunk.candidates?.[0];
+  const usage = chunk.usageMetadata;
+  const usageChunks: StreamProtocolChunk[] = [];
+  if (candidate?.finishReason && usage) {
+    const outputReasoningTokens = (usage as any).thoughtsTokenCount || undefined;
+    const totalOutputTokens = (usage.candidatesTokenCount ?? 0) + (outputReasoningTokens ?? 0);
+
+    usageChunks.push(
+      { data: candidate.finishReason, id: context?.id, type: 'stop' },
+      {
+        data: {
+          // TODO: Google SDK 0.24.0 don't have promptTokensDetails types
+          inputImageTokens: (usage as any).promptTokensDetails?.find(
+            (i: any) => i.modality === 'IMAGE',
+          )?.tokenCount,
+          inputTextTokens: (usage as any).promptTokensDetails?.find(
+            (i: any) => i.modality === 'TEXT',
+          )?.tokenCount,
+          outputReasoningTokens,
+          outputTextTokens: totalOutputTokens - (outputReasoningTokens ?? 0),
+          totalInputTokens: usage.promptTokenCount,
+          totalOutputTokens,
+          totalTokens: usage.totalTokenCount,
+        } as ModelTokensUsage,
+        id: context?.id,
+        type: 'usage',
+      },
+    );
+  }
+
   const functionCalls = chunk.functionCalls?.();
 
   if (functionCalls) {
-    return {
-      data: functionCalls.map(
-        (value, index): StreamToolCallChunkData => ({
-          function: {
-            arguments: JSON.stringify(value.args),
-            name: value.name,
-          },
-          id: generateToolCallId(index, value.name),
-          index: index,
-          type: 'function',
-        }),
-      ),
-      id: context.id,
-      type: 'tool_calls',
-    };
+    return [
+      {
+        data: functionCalls.map(
+          (value, index): StreamToolCallChunkData => ({
+            function: {
+              arguments: JSON.stringify(value.args),
+              name: value.name,
+            },
+            id: generateToolCallId(index, value.name),
+            index: index,
+            type: 'function',
+          }),
+        ),
+        id: context.id,
+        type: 'tool_calls',
+      },
+      ...usageChunks,
+    ];
   }
 
   const text = chunk.text?.();
 
-  if (chunk.candidates) {
-    const candidate = chunk.candidates[0];
-
+  if (candidate) {
     // return the grounding
     if (candidate.groundingMetadata) {
       const { webSearchQueries, groundingChunks } = candidate.groundingMetadata;
@@ -64,31 +96,15 @@ const transformGoogleGenerativeAIStream = (
           id: context.id,
           type: 'grounding',
         },
+        ...usageChunks,
       ];
     }
 
     if (candidate.finishReason) {
       if (chunk.usageMetadata) {
-        const usage = chunk.usageMetadata;
         return [
           !!text ? { data: text, id: context?.id, type: 'text' } : undefined,
-          { data: candidate.finishReason, id: context?.id, type: 'stop' },
-          {
-            data: {
-              // TODO: Google SDK 0.24.0 don't have promptTokensDetails types
-              inputImageTokens: (usage as any).promptTokensDetails?.find(
-                (i: any) => i.modality === 'IMAGE',
-              )?.tokenCount,
-              inputTextTokens: (usage as any).promptTokensDetails?.find(
-                (i: any) => i.modality === 'TEXT',
-              )?.tokenCount,
-              totalInputTokens: usage.promptTokenCount,
-              totalOutputTokens: usage.candidatesTokenCount,
-              totalTokens: usage.totalTokenCount,
-            } as ModelTokensUsage,
-            id: context?.id,
-            type: 'usage',
-          },
+          ...usageChunks,
         ].filter(Boolean) as StreamProtocolChunk[];
       }
       return { data: candidate.finishReason, id: context?.id, type: 'stop' };
@@ -117,13 +133,21 @@ const transformGoogleGenerativeAIStream = (
   };
 };
 
+export interface GoogleAIStreamOptions {
+  callbacks?: ChatStreamCallbacks;
+  inputStartAt?: number;
+}
+
 export const GoogleGenerativeAIStream = (
   rawStream: ReadableStream<EnhancedGenerateContentResponse>,
-  callbacks?: ChatStreamCallbacks,
+  { callbacks, inputStartAt }: GoogleAIStreamOptions = {},
 ) => {
   const streamStack: StreamContext = { id: 'chat_' + nanoid() };
 
   return rawStream
-    .pipeThrough(createSSEProtocolTransformer(transformGoogleGenerativeAIStream, streamStack))
+    .pipeThrough(
+      createTokenSpeedCalculator(transformGoogleGenerativeAIStream, { inputStartAt, streamStack }),
+    )
+    .pipeThrough(createSSEProtocolTransformer((c) => c, streamStack))
     .pipeThrough(createCallbacksTransformer(callbacks));
 };

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

@@ -298,17 +298,37 @@ export const TOKEN_SPEED_CHUNK_ID = 'output_speed';
  */
 export const createTokenSpeedCalculator = (
   transformer: (chunk: any, stack: StreamContext) => StreamProtocolChunk | StreamProtocolChunk[],
-  { streamStack, inputStartAt }: { inputStartAt?: number; streamStack?: StreamContext } = {},
+  { inputStartAt, streamStack }: { inputStartAt?: number; streamStack?: StreamContext } = {},
 ) => {
   let outputStartAt: number | undefined;
+  let outputThinking: boolean | undefined;
 
   const process = (chunk: StreamProtocolChunk) => {
     let result = [chunk];
-    // if the chunk is the first text chunk, set as output start
-    if (!outputStartAt && chunk.type === 'text') outputStartAt = Date.now();
+    // if the chunk is the first text or reasoning chunk, set as output start
+    if (!outputStartAt && (chunk.type === 'text' || chunk.type === 'reasoning')) {
+      outputStartAt = Date.now();
+    }
+
+    /**
+     * 部分 provider 在正式输出 reasoning 前，可能会先输出 content 为空字符串的 chunk，
+     * 其中 reasoning 可能为 null，会导致判断是否输出思考内容错误，所以过滤掉 null 或者空字符串。
+     * 也可能是某些特殊 token，所以不修改 outputStartAt 的逻辑。
+     */
+    if (
+      outputThinking === undefined &&
+      (chunk.type === 'text' || chunk.type === 'reasoning') &&
+      typeof chunk.data === 'string' &&
+      chunk.data.length > 0
+    ) {
+      outputThinking = chunk.type === 'reasoning';
+    }
     // if the chunk is the stop chunk, set as output finish
     if (inputStartAt && outputStartAt && chunk.type === 'usage') {
-      const outputTokens = chunk.data?.totalOutputTokens || chunk.data?.outputTextTokens;
+      const totalOutputTokens = chunk.data?.totalOutputTokens || chunk.data?.outputTextTokens;
+      const reasoningTokens = chunk.data?.outputReasoningTokens || 0;
+      const outputTokens =
+        (outputThinking ?? false) ? totalOutputTokens : totalOutputTokens - reasoningTokens;
       result.push({
         data: {
           tps: (outputTokens / (Date.now() - outputStartAt)) * 1000,

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

@@ -103,10 +103,12 @@ describe('VertexAIStream', () => {
     const onCompletionMock = vi.fn();
 
     const protocolStream = VertexAIStream(mockGoogleStream, {
-      onStart: onStartMock,
-      onText: onTextMock,
-      onToolsCalling: onToolCallMock,
-      onCompletion: onCompletionMock,
+      callbacks: {
+        onStart: onStartMock,
+        onText: onTextMock,
+        onToolsCalling: onToolCallMock,
+        onCompletion: onCompletionMock,
+      },
     });
 
     const decoder = new TextDecoder();
@@ -136,6 +138,7 @@ describe('VertexAIStream', () => {
 
   it('tool_calls', async () => {
     vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('1');
+
     const rawChunks = [
       {
         candidates: [
@@ -204,10 +207,12 @@ describe('VertexAIStream', () => {
     const onCompletionMock = vi.fn();
 
     const protocolStream = VertexAIStream(mockGoogleStream, {
-      onStart: onStartMock,
-      onText: onTextMock,
-      onToolsCalling: onToolCallMock,
-      onCompletion: onCompletionMock,
+      callbacks: {
+        onStart: onStartMock,
+        onText: onTextMock,
+        onToolsCalling: onToolCallMock,
+        onCompletion: onCompletionMock,
+      },
     });
 
     const decoder = new TextDecoder();
@@ -223,10 +228,106 @@ describe('VertexAIStream', () => {
       'id: chat_1\n',
       'event: tool_calls\n',
       `data: [{"function":{"arguments":"{\\"city\\":\\"杭州\\"}","name":"realtime-weather____fetchCurrentWeather"},"id":"realtime-weather____fetchCurrentWeather_0","index":0,"type":"function"}]\n\n`,
+      'id: chat_1\n',
+      'event: stop\n',
+      'data: "STOP"\n\n',
+      'id: chat_1\n',
+      'event: usage\n',
+      'data: {"outputTextTokens":9,"totalInputTokens":95,"totalOutputTokens":9,"totalTokens":104}\n\n',
     ]);
 
     expect(onStartMock).toHaveBeenCalledTimes(1);
     expect(onToolCallMock).toHaveBeenCalledTimes(1);
     expect(onCompletionMock).toHaveBeenCalledTimes(1);
   });
+
+  it('should handle stop with content', async () => {
+    vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('1');
+
+    const data = [
+      {
+        candidates: [
+          {
+            content: { parts: [{ text: '234' }], role: 'model' },
+            safetyRatings: [
+              { category: 'HARM_CATEGORY_HATE_SPEECH', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_DANGEROUS_CONTENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_HARASSMENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT', probability: 'NEGLIGIBLE' },
+            ],
+          },
+        ],
+        text: () => '234',
+        usageMetadata: {
+          promptTokenCount: 20,
+          totalTokenCount: 20,
+          promptTokensDetails: [{ modality: 'TEXT', tokenCount: 20 }],
+        },
+        modelVersion: 'gemini-2.0-flash-exp-image-generation',
+      },
+      {
+        text: () => '567890\n',
+        candidates: [
+          {
+            content: { parts: [{ text: '567890\n' }], role: 'model' },
+            finishReason: 'STOP',
+            safetyRatings: [
+              { category: 'HARM_CATEGORY_HATE_SPEECH', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_DANGEROUS_CONTENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_HARASSMENT', probability: 'NEGLIGIBLE' },
+              { category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT', probability: 'NEGLIGIBLE' },
+            ],
+          },
+        ],
+        usageMetadata: {
+          promptTokenCount: 19,
+          candidatesTokenCount: 11,
+          totalTokenCount: 30,
+          promptTokensDetails: [{ modality: 'TEXT', tokenCount: 19 }],
+          candidatesTokensDetails: [{ modality: 'TEXT', tokenCount: 11 }],
+        },
+        modelVersion: 'gemini-2.0-flash-exp-image-generation',
+      },
+    ];
+
+    const mockGoogleStream = new ReadableStream({
+      start(controller) {
+        data.forEach((item) => {
+          controller.enqueue(item);
+        });
+
+        controller.close();
+      },
+    });
+
+    const protocolStream = VertexAIStream(mockGoogleStream);
+
+    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: chat_1',
+        'event: text',
+        'data: "234"\n',
+
+        'id: chat_1',
+        'event: text',
+        `data: "567890\\n"\n`,
+        // stop
+        'id: chat_1',
+        'event: stop',
+        `data: "STOP"\n`,
+        // usage
+        'id: chat_1',
+        'event: usage',
+        `data: {"inputTextTokens":19,"outputTextTokens":11,"totalInputTokens":19,"totalOutputTokens":11,"totalTokens":30}\n`,
+      ].map((i) => i + '\n'),
+    );
+  });
 });

package/src/libs/agent-runtime/utils/streams/vertex-ai.ts

@@ -1,27 +1,58 @@
 import { EnhancedGenerateContentResponse, GenerateContentResponse } from '@google/generative-ai';
 
+import { ModelTokensUsage } from '@/types/message';
 import { nanoid } from '@/utils/uuid';
 
-import { ChatStreamCallbacks } from '../../types';
+import { type GoogleAIStreamOptions } from './google-ai';
 import {
   StreamContext,
   StreamProtocolChunk,
   createCallbacksTransformer,
   createSSEProtocolTransformer,
+  createTokenSpeedCalculator,
   generateToolCallId,
 } from './protocol';
 
 const transformVertexAIStream = (
   chunk: GenerateContentResponse,
-  stack: StreamContext,
-): StreamProtocolChunk => {
+  context: StreamContext,
+): StreamProtocolChunk | StreamProtocolChunk[] => {
   // maybe need another structure to add support for multiple choices
-  const candidates = chunk.candidates;
+  const candidate = chunk.candidates?.[0];
+  const usage = chunk.usageMetadata;
+  const usageChunks: StreamProtocolChunk[] = [];
+  if (candidate?.finishReason && usage) {
+    const outputReasoningTokens = (usage as any).thoughtsTokenCount || undefined;
+    const totalOutputTokens = (usage.candidatesTokenCount ?? 0) + (outputReasoningTokens ?? 0);
+
+    usageChunks.push(
+      { data: candidate.finishReason, id: context?.id, type: 'stop' },
+      {
+        data: {
+          // TODO: Google SDK 0.24.0 don't have promptTokensDetails types
+          inputImageTokens: (usage as any).promptTokensDetails?.find(
+            (i: any) => i.modality === 'IMAGE',
+          )?.tokenCount,
+          inputTextTokens: (usage as any).promptTokensDetails?.find(
+            (i: any) => i.modality === 'TEXT',
+          )?.tokenCount,
+          outputReasoningTokens,
+          outputTextTokens: totalOutputTokens - (outputReasoningTokens ?? 0),
+          totalInputTokens: usage.promptTokenCount,
+          totalOutputTokens,
+          totalTokens: usage.totalTokenCount,
+        } as ModelTokensUsage,
+        id: context?.id,
+        type: 'usage',
+      },
+    );
+  }
 
+  const candidates = chunk.candidates;
   if (!candidates)
     return {
       data: '',
-      id: stack?.id,
+      id: context?.id,
       type: 'text',
     };
 
@@ -32,44 +63,58 @@ const transformVertexAIStream = (
     if (part.functionCall) {
       const functionCall = part.functionCall;
 
-      return {
-        data: [
-          {
-            function: {
-              arguments: JSON.stringify(functionCall.args),
-              name: functionCall.name,
+      return [
+        {
+          data: [
+            {
+              function: {
+                arguments: JSON.stringify(functionCall.args),
+                name: functionCall.name,
+              },
+              id: generateToolCallId(0, functionCall.name),
+              index: 0,
+              type: 'function',
             },
-            id: generateToolCallId(0, functionCall.name),
-            index: 0,
-            type: 'function',
-          },
-        ],
-        id: stack?.id,
-        type: 'tool_calls',
-      };
+          ],
+          id: context?.id,
+          type: 'tool_calls',
+        },
+        ...usageChunks,
+      ];
+    }
+
+    if (item.finishReason) {
+      if (chunk.usageMetadata) {
+        return [
+          !!part.text ? { data: part.text, id: context?.id, type: 'text' } : undefined,
+          ...usageChunks,
+        ].filter(Boolean) as StreamProtocolChunk[];
+      }
+      return { data: item.finishReason, id: context?.id, type: 'stop' };
     }
 
     return {
       data: part.text,
-      id: stack?.id,
+      id: context?.id,
       type: 'text',
     };
   }
 
   return {
     data: '',
-    id: stack?.id,
+    id: context?.id,
     type: 'stop',
   };
 };
 
 export const VertexAIStream = (
   rawStream: ReadableStream<EnhancedGenerateContentResponse>,
-  callbacks?: ChatStreamCallbacks,
+  { callbacks, inputStartAt }: GoogleAIStreamOptions = {},
 ) => {
   const streamStack: StreamContext = { id: 'chat_' + nanoid() };
 
   return rawStream
-    .pipeThrough(createSSEProtocolTransformer(transformVertexAIStream, streamStack))
+    .pipeThrough(createTokenSpeedCalculator(transformVertexAIStream, { inputStartAt, streamStack }))
+    .pipeThrough(createSSEProtocolTransformer((c) => c, streamStack))
     .pipeThrough(createCallbacksTransformer(callbacks));
 };

package/src/libs/langchain/loaders/epub/index.ts

@@ -2,13 +2,15 @@ import { EPubLoader as Loader } from '@langchain/community/document_loaders/fs/e
 import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';
 
 import { TempFileManager } from '@/server/utils/tempFileManager';
+import { nanoid } from '@/utils/uuid';
 
 import { loaderConfig } from '../config';
 
 export const EPubLoader = async (content: Uint8Array) => {
-  const tempManager = new TempFileManager();
+  const tempManager = new TempFileManager('epub-');
+
   try {
-    const tempPath = await tempManager.writeTempFile(content);
+    const tempPath = await tempManager.writeTempFile(content, `${nanoid()}.epub`);
     const loader = new Loader(tempPath);
     const documents = await loader.load();
 

package/src/libs/mcp/__tests__/__snapshots__/index.test.ts.snap

@@ -21,9 +21,11 @@ exports[`MCPClient > Stdio Transport > should list tools via stdio 1`] = `
     "name": "echo",
   },
   {
-    "annotations": {},
     "description": "Lists all available tools and methods",
     "inputSchema": {
+      "$schema": "http://json-schema.org/draft-07/schema#",
+      "additionalProperties": false,
+      "properties": {},
       "type": "object",
     },
     "name": "debug",

package/src/server/routers/async/file.ts

@@ -35,7 +35,7 @@ const fileProcedure = asyncAuthedProcedure.use(async (opts) => {
       chunkService: new ChunkService(ctx.userId),
       embeddingModel: new EmbeddingModel(ctx.serverDB, ctx.userId),
       fileModel: new FileModel(ctx.serverDB, ctx.userId),
-      fileService: new FileService(),
+      fileService: new FileService(ctx.serverDB, ctx.userId),
     },
   });
 });

package/src/server/routers/lambda/__tests__/file.test.ts

@@ -0,0 +1,213 @@
+import { TRPCError } from '@trpc/server';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { fileRouter } from '@/server/routers/lambda/file';
+import { AsyncTaskStatus } from '@/types/asyncTask';
+
+// Patch: Use actual router context middleware to inject the correct models/services
+function createCallerWithCtx(partialCtx: any = {}) {
+  // All mocks are spies
+  const fileModel = {
+    checkHash: vi.fn().mockResolvedValue({ isExist: true }),
+    create: vi.fn().mockResolvedValue({ id: 'test-id' }),
+    findById: vi.fn().mockResolvedValue(undefined),
+    query: vi.fn().mockResolvedValue([]),
+    delete: vi.fn().mockResolvedValue(undefined),
+    deleteMany: vi.fn().mockResolvedValue([]),
+    clear: vi.fn().mockResolvedValue({} as any),
+  };
+
+  const fileService = {
+    getFullFileUrl: vi.fn().mockResolvedValue('full-url'),
+    deleteFile: vi.fn().mockResolvedValue(undefined),
+    deleteFiles: vi.fn().mockResolvedValue(undefined),
+  };
+
+  const chunkModel = {
+    countByFileIds: vi.fn().mockResolvedValue([{ id: 'test-id', count: 5 }]),
+    countByFileId: vi.fn().mockResolvedValue(5),
+  };
+
+  const asyncTaskModel = {
+    findByIds: vi.fn().mockResolvedValue([
+      {
+        id: 'test-task-id',
+        status: AsyncTaskStatus.Success,
+      },
+    ]),
+    findById: vi.fn(),
+    delete: vi.fn(),
+  };
+
+  const ctx = {
+    serverDB: {} as any,
+    userId: 'test-user',
+    asyncTaskModel,
+    chunkModel,
+    fileModel,
+    fileService,
+    ...partialCtx,
+  };
+
+  return { ctx, caller: fileRouter.createCaller(ctx) };
+}
+
+vi.mock('@/config/db', () => ({
+  serverDBEnv: {
+    REMOVE_GLOBAL_FILE: false,
+  },
+}));
+
+vi.mock('@/database/models/asyncTask', () => ({
+  AsyncTaskModel: vi.fn(() => ({
+    findById: vi.fn(),
+    findByIds: vi.fn(),
+    delete: vi.fn(),
+  })),
+}));
+
+vi.mock('@/database/models/chunk', () => ({
+  ChunkModel: vi.fn(() => ({
+    countByFileId: vi.fn(),
+    countByFileIds: vi.fn(),
+  })),
+}));
+
+vi.mock('@/database/models/file', () => ({
+  FileModel: vi.fn(() => ({
+    checkHash: vi.fn(),
+    create: vi.fn(),
+    delete: vi.fn(),
+    deleteMany: vi.fn(),
+    findById: vi.fn(),
+    query: vi.fn(),
+    clear: vi.fn(),
+  })),
+}));
+
+vi.mock('@/server/services/file', () => ({
+  FileService: vi.fn(() => ({
+    getFullFileUrl: vi.fn(),
+    deleteFile: vi.fn(),
+    deleteFiles: vi.fn(),
+  })),
+}));
+
+describe('fileRouter', () => {
+  let ctx: any;
+  let caller: any;
+  let mockFile: any;
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+
+    mockFile = {
+      id: 'test-id',
+      name: 'test.txt',
+      url: 'test-url',
+      createdAt: new Date(),
+      updatedAt: new Date(),
+      accessedAt: new Date(),
+      userId: 'test-user',
+      size: 100,
+      fileType: 'text',
+      metadata: {},
+      fileHash: null,
+      clientId: null,
+      chunkTaskId: null,
+      embeddingTaskId: null,
+    };
+
+    // Use actual context with default mocks
+    ({ ctx, caller } = createCallerWithCtx());
+  });
+
+  describe('checkFileHash', () => {
+    it('should handle when fileModel.checkHash returns undefined', async () => {
+      ctx.fileModel.checkHash.mockResolvedValue(undefined);
+      await expect(caller.checkFileHash({ hash: 'test-hash' })).resolves.toBeUndefined();
+    });
+  });
+
+  describe('createFile', () => {
+    it('should throw if fileModel.checkHash returns undefined', async () => {
+      ctx.fileModel.checkHash.mockResolvedValue(undefined);
+      await expect(
+        caller.createFile({
+          hash: 'test-hash',
+          fileType: 'text',
+          name: 'test.txt',
+          size: 100,
+          url: 'test-url',
+          metadata: {},
+        }),
+      ).rejects.toThrow();
+    });
+  });
+
+  describe('findById', () => {
+    it('should throw error when file not found', async () => {
+      ctx.fileModel.findById.mockResolvedValue(null);
+
+      await expect(caller.findById({ id: 'invalid-id' })).rejects.toThrow(TRPCError);
+    });
+  });
+
+  describe('getFileItemById', () => {
+    it('should throw error when file not found', async () => {
+      ctx.fileModel.findById.mockResolvedValue(null);
+
+      await expect(caller.getFileItemById({ id: 'invalid-id' })).rejects.toThrow(TRPCError);
+    });
+  });
+
+  describe('getFiles', () => {
+    it('should handle fileModel.query returning undefined', async () => {
+      ctx.fileModel.query.mockResolvedValue(undefined);
+
+      await expect(caller.getFiles({})).rejects.toThrow();
+    });
+  });
+
+  describe('removeFile', () => {
+    it('should do nothing when file not found', async () => {
+      ctx.fileModel.delete.mockResolvedValue(null);
+
+      await caller.removeFile({ id: 'invalid-id' });
+
+      expect(ctx.fileService.deleteFile).not.toHaveBeenCalled();
+    });
+  });
+
+  describe('removeFiles', () => {
+    it('should do nothing when no files found', async () => {
+      ctx.fileModel.deleteMany.mockResolvedValue([]);
+
+      await caller.removeFiles({ ids: ['invalid-1', 'invalid-2'] });
+
+      expect(ctx.fileService.deleteFiles).not.toHaveBeenCalled();
+    });
+  });
+
+  describe('removeFileAsyncTask', () => {
+    it('should do nothing when file not found', async () => {
+      ctx.fileModel.findById.mockResolvedValue(null);
+
+      await caller.removeFileAsyncTask({ id: 'test-id', type: 'chunk' });
+
+      expect(ctx.asyncTaskModel.delete).not.toHaveBeenCalled();
+    });
+
+    it('should do nothing when task id is missing', async () => {
+      ctx.fileModel.findById.mockResolvedValue(mockFile);
+
+      await caller.removeFileAsyncTask({ id: 'test-id', type: 'embedding' });
+
+      expect(ctx.asyncTaskModel.delete).not.toHaveBeenCalled();
+
+      await caller.removeFileAsyncTask({ id: 'test-id', type: 'chunk' });
+
+      expect(ctx.asyncTaskModel.delete).not.toHaveBeenCalled();
+    });
+  });
+});

package/src/server/routers/lambda/file.ts

@@ -19,7 +19,7 @@ const fileProcedure = authedProcedure.use(serverDatabase).use(async (opts) => {
       asyncTaskModel: new AsyncTaskModel(ctx.serverDB, ctx.userId),
       chunkModel: new ChunkModel(ctx.serverDB, ctx.userId),
       fileModel: new FileModel(ctx.serverDB, ctx.userId),
-      fileService: new FileService(),
+      fileService: new FileService(ctx.serverDB, ctx.userId),
     },
   });
 });

package/src/server/routers/lambda/importer.ts

@@ -10,10 +10,12 @@ import { ImportResultData, ImporterEntryData } from '@/types/importer';
 
 const importProcedure = authedProcedure.use(serverDatabase).use(async (opts) => {
   const { ctx } = opts;
-  const dataImporterService = new DataImporterRepos(ctx.serverDB, ctx.userId);
 
   return opts.next({
-    ctx: { dataImporterService, fileService: new FileService() },
+    ctx: {
+      dataImporterService: new DataImporterRepos(ctx.serverDB, ctx.userId),
+      fileService: new FileService(ctx.serverDB, ctx.userId),
+    },
   });
 });
 

package/src/server/routers/lambda/message.ts

@@ -16,7 +16,7 @@ const messageProcedure = authedProcedure.use(serverDatabase).use(async (opts) =>
 
   return opts.next({
     ctx: {
-      fileService: new FileService(),
+      fileService: new FileService(ctx.serverDB, ctx.userId),
       messageModel: new MessageModel(ctx.serverDB, ctx.userId),
     },
   });
@@ -102,7 +102,7 @@ export const messageRouter = router({
       const serverDB = await getServerDB();
 
       const messageModel = new MessageModel(serverDB, ctx.userId);
-      const fileService = new FileService();
+      const fileService = new FileService(serverDB, ctx.userId);
 
       return messageModel.query(input, {
         postProcessUrl: (path) => fileService.getFullFileUrl(path),

package/src/server/routers/lambda/ragEval.ts

@@ -40,7 +40,7 @@ const ragEvalProcedure = authedProcedure
         datasetRecordModel: new EvalDatasetRecordModel(ctx.userId),
         evaluationModel: new EvalEvaluationModel(ctx.userId),
         evaluationRecordModel: new EvaluationRecordModel(ctx.userId),
-        fileService: new FileService(),
+        fileService: new FileService(ctx.serverDB, ctx.userId),
       },
     });
   });

package/src/server/routers/lambda/user.ts

@@ -1,6 +1,6 @@
 import { UserJSON } from '@clerk/backend';
+import { v4 as uuidv4 } from 'uuid';
 import { z } from 'zod';
-import { v4 as uuidv4 } from 'uuid'; // 需要添加此导入
 
 import { enableClerk } from '@/const/auth';
 import { isDesktop } from '@/const/version';
@@ -28,7 +28,7 @@ const userProcedure = authedProcedure.use(serverDatabase).use(async ({ ctx, next
   return next({
     ctx: {
       clerkAuth: new ClerkAuth(),
-      fileService: new FileService(),
+      fileService: new FileService(ctx.serverDB, ctx.userId),
       nextAuthDbAdapter: LobeNextAuthDbAdapter(ctx.serverDB),
       userModel: new UserModel(ctx.serverDB, ctx.userId),
     },

package/src/server/services/file/index.ts

@@ -1,3 +1,12 @@
+import { TRPCError } from '@trpc/server';
+
+import { serverDBEnv } from '@/config/db';
+import { FileModel } from '@/database/models/file';
+import { FileItem } from '@/database/schemas';
+import { LobeChatDatabase } from '@/database/type';
+import { TempFileManager } from '@/server/utils/tempFileManager';
+import { nanoid } from '@/utils/uuid';
+
 import { FileServiceImpl, createFileServiceModule } from './impls';
 
 /**
@@ -5,8 +14,16 @@ import { FileServiceImpl, createFileServiceModule } from './impls';
  * 使用模块化实现方式，提供文件操作服务
  */
 export class FileService {
+  private userId: string;
+  private fileModel: FileModel;
+
   private impl: FileServiceImpl = createFileServiceModule();
 
+  constructor(db: LobeChatDatabase, userId: string) {
+    this.userId = userId;
+    this.fileModel = new FileModel(db, userId);
+  }
+
   /**
    * 删除文件
    */
@@ -62,4 +79,33 @@ export class FileService {
   public async getFullFileUrl(url?: string | null, expiresIn?: number): Promise<string> {
     return this.impl.getFullFileUrl(url, expiresIn);
   }
+
+  async downloadFileToLocal(
+    fileId: string,
+  ): Promise<{ cleanup: () => void; file: FileItem; filePath: string }> {
+    const file = await this.fileModel.findById(fileId);
+    if (!file) {
+      throw new TRPCError({ code: 'BAD_REQUEST', message: 'File not found' });
+    }
+
+    let content: Uint8Array | undefined;
+    try {
+      content = await this.getFileByteArray(file.url);
+    } catch (e) {
+      console.error(e);
+      // if file not found, delete it from db
+      if ((e as any).Code === 'NoSuchKey') {
+        await this.fileModel.delete(fileId, serverDBEnv.REMOVE_GLOBAL_FILE);
+        throw new TRPCError({ code: 'BAD_REQUEST', message: 'File not found' });
+      }
+    }
+
+    if (!content) throw new TRPCError({ code: 'BAD_REQUEST', message: 'File content is empty' });
+
+    const dir = nanoid();
+    const tempManager = new TempFileManager(dir);
+
+    const filePath = await tempManager.writeTempFile(content, file.name);
+    return { cleanup: () => tempManager.cleanup(), file, filePath };
+  }
 }

package/src/server/utils/tempFileManager.ts

@@ -1,7 +1,6 @@
 import { existsSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
-import { v4 as uuidv4 } from 'uuid';
 
 /**
  * 安全存储临时文件工具类
@@ -10,21 +9,19 @@ export class TempFileManager {
   private readonly tempDir: string;
   private filePaths: Set<string> = new Set();
 
-  constructor() {
+  constructor(dirname: string) {
     // 创建唯一临时目录 (跨平台安全)
-    this.tempDir = mkdtempSync(join(tmpdir(), 'epub-'));
+    this.tempDir = mkdtempSync(join(tmpdir(), dirname));
     // 注册退出清理钩子
     this.registerCleanupHook();
   }
 
   /**
    * 将 Uint8Array 写入临时文件
-   * @param data 文件数据
-   * @param ext 文件扩展名 (默认 .epub)
-   * @returns 临时文件绝对路径
+
    */
-  async writeTempFile(data: Uint8Array, ext = '.epub'): Promise<string> {
-    const filePath = join(this.tempDir, `${uuidv4()}${ext}`);
+  async writeTempFile(data: Uint8Array, name: string): Promise<string> {
+    const filePath = join(this.tempDir, name);
 
     try {
       writeFileSync(filePath, data);