@lobehub/chat 1.122.2 → 1.122.3

package/CHANGELOG.md

@@ -2,6 +2,39 @@
 
 # Changelog
 
+### [Version 1.122.3](https://github.com/lobehub/lobe-chat/compare/v1.122.2...v1.122.3)
+
+<sup>Released on **2025-09-04**</sup>
+
+#### 🐛 Bug Fixes
+
+- **misc**: Support base64 image from markdown image syntax.
+
+#### 💄 Styles
+
+- **misc**: Update the price of the o3 model in OpenRouter.
+
+<br/>
+
+<details>
+<summary><kbd>Improvements and Fixes</kbd></summary>
+
+#### What's fixed
+
+- **misc**: Support base64 image from markdown image syntax, closes [#9054](https://github.com/lobehub/lobe-chat/issues/9054) ([d013a16](https://github.com/lobehub/lobe-chat/commit/d013a16))
+
+#### Styles
+
+- **misc**: Update the price of the o3 model in OpenRouter, closes [#9075](https://github.com/lobehub/lobe-chat/issues/9075) ([43ef47c](https://github.com/lobehub/lobe-chat/commit/43ef47c))
+
+</details>
+
+<div align="right">
+
+[![](https://img.shields.io/badge/-BACK_TO_TOP-151515?style=flat-square)](#readme-top)
+
+</div>
+
 ### [Version 1.122.2](https://github.com/lobehub/lobe-chat/compare/v1.122.1...v1.122.2)
 
 <sup>Released on **2025-09-04**</sup>

package/changelog/v1.json

@@ -1,4 +1,16 @@
 [
+  {
+    "children": {
+      "fixes": [
+        "Support base64 image from markdown image syntax."
+      ],
+      "improvements": [
+        "Update the price of the o3 model in OpenRouter."
+      ]
+    },
+    "date": "2025-09-04",
+    "version": "1.122.3"
+  },
   {
     "children": {},
     "date": "2025-09-04",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lobehub/chat",
-  "version": "1.122.2",
+  "version": "1.122.3",
   "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",

package/packages/model-bank/src/aiModels/openrouter.ts

@@ -396,9 +396,9 @@ const openrouterChatModels: AIChatModelCard[] = [
     maxOutput: 100_000,
     pricing: {
       units: [
-        { name: 'textInput_cacheRead', rate: 2.5, strategy: 'fixed', unit: 'millionTokens' },
-        { name: 'textInput', rate: 10, strategy: 'fixed', unit: 'millionTokens' },
-        { name: 'textOutput', rate: 40, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput_cacheRead', rate: 0.5, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput', rate: 2, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textOutput', rate: 8, strategy: 'fixed', unit: 'millionTokens' },
       ],
     },
     releasedAt: '2025-04-17',

package/packages/model-runtime/src/utils/streams/openai/openai.test.ts

@@ -163,6 +163,61 @@ describe('OpenAIStream', () => {
     );
   });
 
+  it('should emit base64_image and strip markdown data:image from text', async () => {
+    const data = [
+      {
+        id: 'img-1',
+        choices: [
+          { index: 0, delta: { role: 'assistant', content: '这是一张图片： ' } },
+        ],
+      },
+      {
+        id: 'img-1',
+        choices: [
+          {
+            index: 0,
+            delta: {
+              content:
+                '![image](data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAIAAAB3D1E1AA==)',
+            },
+          },
+        ],
+      },
+      { id: 'img-1', choices: [{ index: 0, delta: {}, finish_reason: 'stop' }] },
+    ];
+
+    const mockOpenAIStream = new ReadableStream({
+      start(controller) {
+        data.forEach((c) => controller.enqueue(c));
+        controller.close();
+      },
+    });
+
+    const protocolStream = OpenAIStream(mockOpenAIStream);
+
+    const decoder = new TextDecoder();
+    const chunks: string[] = [];
+
+    // @ts-ignore
+    for await (const chunk of protocolStream) {
+      chunks.push(decoder.decode(chunk, { stream: true }));
+    }
+
+    expect(chunks).toEqual(
+      [
+        'id: img-1',
+        'event: text',
+        `data: "这是一张图片： "\n`,
+        'id: img-1',
+        'event: base64_image',
+        `data: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAIAAAB3D1E1AA=="\n`,
+        'id: img-1',
+        'event: stop',
+        `data: "stop"\n`,
+      ].map((i) => `${i}\n`),
+    );
+  });
+
   it('should handle content with tool_calls but is an empty object', async () => {
     // data: {"id":"chatcmpl-A7pokGUqSov0JuMkhiHhWU9GRtAgJ", "object":"chat.completion.chunk", "created":1726430846, "model":"gpt-4o-2024-05-13", "choices":[{"index":0, "delta":{"content":" today", "role":"", "tool_calls":[]}, "finish_reason":"", "logprobs":""}], "prompt_annotations":[{"prompt_index":0, "content_filter_results":null}]}
     const mockOpenAIStream = new ReadableStream({
@@ -2311,4 +2366,86 @@ describe('OpenAIStream', () => {
 
     expect(chunks).toEqual(['id: 6\n', 'event: base64_image\n', `data: "${base64}"\n\n`]);
   });
+
+  it('should handle finish_reason with markdown image in content', async () => {
+    const base64 = 'data:image/png;base64,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';
+    const mockOpenAIStream = new ReadableStream({
+      start(controller) {
+        controller.enqueue({
+          id: 'chatcmpl-test',
+          choices: [
+            {
+              index: 0,
+              delta: { content: `这有一张图片： ![image](${base64})` },
+              finish_reason: 'stop',
+            },
+          ],
+        });
+
+        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: chatcmpl-test\n',
+      'event: text\n',
+      `data: "这有一张图片："\n\n`,
+      'id: chatcmpl-test\n',
+      'event: base64_image\n',
+      `data: "${base64}"\n\n`,
+    ]);
+  });
+
+  it('should handle finish_reason with multiple markdown images in content', async () => {
+    const base64_1 = 'data:image/png;base64,first';
+    const base64_2 = 'data:image/jpeg;base64,second';
+    const mockOpenAIStream = new ReadableStream({
+      start(controller) {
+        controller.enqueue({
+          id: 'chatcmpl-multi',
+          choices: [
+            {
+              index: 0,
+              delta: { content: `![img1](${base64_1}) and ![img2](${base64_2})` },
+              finish_reason: 'stop',
+            },
+          ],
+        });
+
+        controller.close();
+      },
+    });
+
+    const protocolStream = OpenAIStream(mockOpenAIStream);
+
+    const decoder = new TextDecoder();
+    const chunks = [];
+
+    // @ts-ignore
+    for await (const chunk of protocolStream) {
+      chunks.push(decoder.decode(chunk, { stream: true }));
+    }
+
+    expect(chunks).toEqual([
+      'id: chatcmpl-multi\n',
+      'event: text\n',
+      `data: "and"\n\n`, // Remove all markdown base64 image segments
+      'id: chatcmpl-multi\n',
+      'event: base64_image\n',
+      `data: "${base64_1}"\n\n`,
+      'id: chatcmpl-multi\n',
+      'event: base64_image\n',
+      `data: "${base64_2}"\n\n`,
+    ]);
+  });
 });

package/packages/model-runtime/src/utils/streams/openai/openai.ts

@@ -20,6 +20,28 @@ import {
   generateToolCallId,
 } from '../protocol';
 
+// Process markdown base64 images: extract URLs and clean text in one pass
+const processMarkdownBase64Images = (text: string): { cleanedText: string, urls: string[]; } => {
+  if (!text) return { cleanedText: text, urls: [] };
+  
+  const urls: string[] = [];
+  const mdRegex = /!\[[^\]]*]\(\s*(data:image\/[\d+.A-Za-z-]+;base64,[^\s)]+)\s*\)/g;
+  let cleanedText = text;
+  let m: RegExpExecArray | null;
+  
+  // Reset regex lastIndex to ensure we start from the beginning
+  mdRegex.lastIndex = 0;
+  
+  while ((m = mdRegex.exec(text)) !== null) {
+    if (m[1]) urls.push(m[1]);
+  }
+  
+  // Remove all markdown base64 image segments
+  cleanedText = text.replaceAll(mdRegex, '').trim();
+  
+  return { cleanedText, urls };
+};
+
 const transformOpenAIStream = (
   chunk: OpenAI.ChatCompletionChunk,
   streamContext: StreamContext,
@@ -137,7 +159,19 @@ const transformOpenAIStream = (
           return { data: null, id: chunk.id, type: 'text' };
         }
 
-        return { data: item.delta.content, id: chunk.id, type: 'text' };
+
+        const text = item.delta.content as string;
+        const { urls: images, cleanedText: cleaned } = processMarkdownBase64Images(text);
+        if (images.length > 0) {
+          const arr: StreamProtocolChunk[] = [];
+          if (cleaned) arr.push({ data: cleaned, id: chunk.id, type: 'text' });
+          arr.push(
+            ...images.map((url: string) => ({ data: url, id: chunk.id, type: 'base64_image' as const })),
+          );
+          return arr;
+        }
+
+        return { data: text, id: chunk.id, type: 'text' };
       }
 
       // OpenAI Search Preview 模型返回引用源
@@ -284,7 +318,7 @@ const transformOpenAIStream = (
           if (citations) {
             streamContext.returnedCitation = true;
 
-            return [
+            const baseChunks: StreamProtocolChunk[] = [
               {
                 data: {
                   citations: (citations as any[])
@@ -303,6 +337,20 @@ const transformOpenAIStream = (
                 type: streamContext?.thinkingInContent ? 'reasoning' : 'text',
               },
             ];
+            return baseChunks;
+          }
+        }
+
+        // 非思考模式下，额外解析 markdown 中的 base64 图片，按顺序输出 text -> base64_image
+        if (!streamContext?.thinkingInContent) {
+          const { urls, cleanedText: cleaned } = processMarkdownBase64Images(thinkingContent);
+          if (urls.length > 0) {
+            const arr: StreamProtocolChunk[] = [];
+            if (cleaned) arr.push({ data: cleaned, id: chunk.id, type: 'text' });
+            arr.push(
+              ...urls.map((url: string) => ({ data: url, id: chunk.id, type: 'base64_image' as const })),
+            );
+            return arr;
           }
         }