Use vLLM-served qwen3 model #1009
Description
vLLM-served Qwen3 returned a valid server response, but smart_scraper_graph.run() get None.
log:
http://ai3.18chat.cn:7860/v1
Max input tokens for model openai/tiny not found,
please specify the model_tokens parameter in the llm section of the graph configuration.
Using default token size: 8192
INFO:web-loader:Starting scraping with playwright...
INFO:web-loader:Content scraped
INFO:httpx:HTTP Request: POST http://ai3.18chat.cn:7860/v1/chat/completions "HTTP/1.1 200 OK"
None
✨ Try enhanced version of ScrapegraphAI at https://scrapegraphai.com/oss ✨
config:
graph_config = {
"llm": {
"api_key": api_key,
"model": f"openai/{model_name}",
"temperature": 0.3,
"openai_api_base": base_url,
"max_tokens": 8192,
# "model_tokens": 16000, # use this param get error
},
"verbose": False,
"headless": False,
"loader_kwargs": {
"browser_name": "chromium",
"backend": "playwright",
# "load_state": 'networkidle',
# "requires_js_support": True,
"timeout": 60,
"retry_limit": 1,
}
}
vllm qwen3 response:
[OpenAI API]
{'id': '2479', 'object': 'chat.completion', 'created': 1758147406, 'model': 'gpt', 'choices': [{'index': 0, 'message': {'role': 'assistant', 'content': '{"content": "## 对于普通程序员来说 AI 是什么？AI 究竟用的是什么？\n\n这是一个很有趣的问题，很多人觉得有了 AI 之后是不需要学习，因为 AI 可以帮助他解决大部分问题 ，但是如果你已经使用过一段时间，或者说你正在使用 vibe coding ，那么你的实际体验会告诉你，你需要学习，甚至比之前了解更多。\n\n> 实际上有了 AI 之后，在某些方面我们确实不大需要去认真学习了，比如各种框架的 API ，因为 AI 总能找到合适的 API 来实现需求 。\n\n但是于此而言，作为使用者，我们需要去了解如何使用 AI ，甚至需要知道有什么技术框架，如何组织这种技术框架才能完成我们的需求。\n\n相信大家应该都看过不少 AI 的营销广告，稍微输入几句话，就可以得到精美的页面，甚至完成一键部署，就像是曾经的精美页游广告，但是当自己进入游戏后才发现，为什么我的 AI 并没有想象中聪明？\n\n### 阶段一：AI 的现状与误区\n\n首先第一点就是，现阶段的 AI 确实不够聪明 ，很多时候你看到的精美 AI 营销内容，不外乎两种情况：\n\n- 可远观而不可亵玩的展示品，营销里的都会经过无数次尝试后给出的结果，而这个输出看似精美，但是实际根本不能用，那只是一个 Demo，初见时美若 天仙，但是如果你走近打开灯光，就会发现根本不能用。\n\n- 专业的 vibe coding 大佬们通过各种详尽的规则，让 AI 的智商在当前项目得到了提升。\n\n是的，用好 vibe coding 并不是一件简单的事情，在 vibe coding 领域，给一个项目写 rule 是一件非常耗时的过程，这也是为什么在今年 2025 的 StackOverflow 调查里，有 72% 的开发者对于完全使用 vibe code 不要放心：\n\n因为如果要用好 AI ，你就需要学会配置各种 rule ，但是配置 rule 的时间，大概率你已经把需求写完了 ，这样是目前 vibe coding 的现状。\n\n> 你觉得 AI 不好用，因为你没花费时间去调试规则和提示词 ，都是大家都是同个模型，为什么你用的和他的不 一样，大多数原因就在于这里。\n\n这么说可能不明显，在 AI 画图领域比代码更明显，因为画图是直接输出成果，不像代码这种中间过程，如果要你在 ChatGPT 画出这种图，你会用怎么做？\n\n我们来看官方的提示词 ：\n> 创建一个超现实的3D构图，其中独立的岩石块组合成【主题】。风格：破碎板岩 + 砂岩碎片，每个部分明显分离，留有微小间隙；经过雕琢的边缘，可见的石纹。外观：极简的米白色背景，柔和的自左上方向下的工作室灯光，细腻的阴影，暗板岩中点缀着微暖的赭色色调。构图：居中，整洁，正面视角。添加几片微小碎片以增加深度。无文字，无额外道具。格式：【比例】。使各部分清晰分离且易于辨识。\n\n可以看到，这里面都是细节，你需要告诉 AI ：\n- 风格：破碎板岩 + 砂岩碎片\n- 每个部分明显分离，留有微小间隙\n- 可见的石纹\n- 灯光规则\n- 各种色调\n- 构图方式\n- 地面细节\n- ···\n\n也许看起来是短短一段话，但是如果你没有专业知识打底，事实上你很难调试出这样一个图生图的效果，这其实在编程领域也是一样。\n\n举一个另外很有趣的误区，那就是问模型它自己叫啥，然后发现好像和自己用的不一致，然后觉得自己被骗了，这其实也是幻觉所致，因为从训练的角度看，训练 Claude4 的时候，互联网语料还没有 Claude4 这个东西，不管是 Trae 还是 Cursor 都会出现这个问题，而官方之所以能正确，是在系统提示词方面增加了配置。\n\n### 为什么需要 Rules\n\n- AI 有时会偏离技术栈，或使用不一致的技术导致失败，例如想用 SQL，但 AI 突然换成 JSON 文件存储，导致问题\n- Rules 相当于 AI 的 \"system prompt\"，可清晰告诉 Agent 使用技术栈和编码偏 好，确保一致性\n\n### 如何配置 Rules\n\n- 比如在 Cursor 的 project rules，会生成一个 cursor 文件夹，内部包含 .mdc（或类似）规则文件\n- 可同时加入用户级（通用）规则与项目级规则，推荐优先使用项目级设置，便于管理\n\n#### 常见规则类型和设置内容\n\n| 必备规则 | 原因 |\n|---|---|\n| 技术栈限制 | 明确指定语言与工具，例如：使用 Flutter 3.29.3 版本 API，使用 Provider 6.1.2 作为状态管理框架，使用 dio 5.8.0+1 作为网络请求，使用 cherry_toast 1.12.0 作为提示框架·····这里的核心目的就是限定范围，越是限制越详细，幻觉就会越来越低，错误也就越来越少。 |\n| 简洁优先 | 要求 AI 优先 选择简单方案，避免代码重复，检查是否已有实现，不要盲目新建功能 |\n| 环境区分 | 明确分隔 Dev、Test 和 Prod，确保修改只在 Dev 或 Test 环境执行，生产环境稳定 |\n| 范围控制 | 只修改被明确请求 的代码部分，不应产生副作用或引入新技术；如确需改动，务必先移除旧实现，避免重复逻辑，这在 AI 使用过程中是经常发生的，因为它时不时就会出现随机乱改的情况，所以类似要求必须有 |\n| 避免脚本残留 | 不要在代码库中留下单次运行的脚本；执行完应删除，或者直接在代码中 inline 而非文件留存，这在 Agent 执行过程中经常存在，不管是 Github Copilot Agent 还是 Cursor |\n| 文件大小限制 | 限制当个文件大小，比如文件超过 200–300 行需及早重构，防止庞大代码文件影响 AI 理解与测试，因为 AI 的上下文限制，AI 实际上是没办法完全理解一个超大型代码文件的，所以不生成超大型代码文件是基本要求 |\n| API Key 安全 | 不允许 Agent 覆盖本地 .env 文件或 API 密钥配置，避免泄露或重置问题 |\n| 测试覆盖 | 强制为所有重要功能编写测试，让每个功能都可以单独运行单元测试 |\n| 稳定性优先 | 稳定架构已有时，不允许 AI 引入大规模改结构除非明确指示；关注可能受影响模块 |\n\n#### 其他建议\n\n- 上下文管理：随着对话内容变多，AI 性能可能下降或者越觉增加，所以尽量在适当时候开启新 Chat，载入关键规则和代码片段 ，防止 context overflow 导致效率下降\n- 请求粒度：保持任务微小、明确（例如“添加一个功能”），便于 AI 针对执行，不易引入错误\n- 尽量不要使用 AI 直接对中大型项目进行重构，AI 最不擅长做的事是 对中大型项目进行重构和迁移，因为如果你不给出详细的规则，实现步骤，审查逻辑和 fallback 规则，那么它就会充满自己的想法，迁移过程很难控制。所以，不要给出像“帮我构建整个功能”这样冗长的提示 ，这样绝 对会让 AI 会开始产生幻觉，然后胡乱造作，这样基本是在浪费 Token 去赌 AI 会不会有幻觉，AI 很难处理超长上下文的东西，所以才有 Agent 的场景，Agent 一般会帮你拆分需求，但是如果都让 Agent 去拆分，那么可能每次拆分的结果可能都不大一样，也就是幻觉的概率问题。所以，在使用 AI 过程中，你必须学会将任何想要添加的功能分解成几个阶段，尤其是在构建复杂功能时，与其一次性给出一个冗长的提示，不如根据你的用例将其分解成 3-5 个甚至更多的请求。\n\n> AI 是帮助专业的人变得更专业。\n\n最后，一些框架也会提供它们的 AI 模板，例如 Flutter 就通过 [docs.flutter.dev/ai/ai-rules](https://link.juejin.cn?target=https%3A%2F%2Fdocs.flutter.dev%2Fai%2Fai-rules \"https://docs.flutter.dev/ai/ai-rules\") 提供了对应的模板：\n\n# AI rules for Flutter\n\nYou are an expert in Flutter and Dart development. Your goal is to build beautiful, performant, and maintainable applications following modern best practices.\n\n## Project Structure\n\n* Assumes a standard Flutter project structure with lib/main.dart as the primary application entry point.\n\n## Package Management\n\n* If a new feature requires an external package, the AI will identify the most suitable and stable package from pub.dev.\n* To add a regular dependency, it will execute flutter pub add.\n* To add a development dependency, it will execute flutter pub add dev:.\n\n## Code Quality\n\n* Adhere to maintainable code structure and separation of concerns (e.g., UI logic separate from business logic).\n* Adhere to meaningful and consistent naming conventions.\n\n## Dart Best Practices\n\n* Follow the official Effective Dart guidelines.\n* Define related classes within the same library file. For large libraries, export smaller, private libraries from a single top-level library.\n* Group related libraries in the same folder.\n* Add documentation comments to all public APIs, including classes, constructors, methods, and top-level functions.\n* Write clear comments for complex or non-obvious code. Avoid over-commenting. Don't add trailing comments.\n* Ensure proper use of async/await for asynchronous operations with robust error handling.\n* Use pattern matching features where they simplify the code.\n\n## Flutter Best Practices\n\n* Widgets (especially StatelessWidget) are immutable; when the UI needs to change, Flutter rebuilds the widget tree.\n* Prefer composing smaller widgets over extending existing ones.\n* Use small, private Widget classes instead of private helper methods that return a Widget.\n* Break down large build() methods into smaller, reusable private Widget classes.\n* Use ListView.builder to create lazy-loaded lists for performance.\n* Use const constructors for widgets and in build() methods whenever possible to optimize performance.\n* Avoid performing expensive operations, like network calls or complex computations, directly within build() methods.\n\n## Application Architecture\n\n* Aim for separation of concerns similar to MVC/MVVM, with defined Model, View, and ViewModel/Controller roles.\n\n### State Management\n\n* Default to Flutter's built-in state management solutions. Do not use a third-party package unless explicitly requested.\n* Use Streams and StreamBuilder for handling a sequence of asynchronous events.\n* Use Futures and FutureBuilder for handling a single asynchronous operation that will complete in the future.\n* Use ValueNotifier with ValueListenableBuilder for simple, local state that involves a single value.\n\ndart\\n// Define a ValueNotifier to hold the state.\\nfinal ValueNotifier _counter = ValueNotifier(0);\\n// Use ValueListenableBuilder to listen and rebuild.\\nValueListenableBuilder(\\n valueListenable: _counter,\\n builder: (context, value, child) {\\n return Text(\'Count: $value\');\\n },\\n);\\n\n\n* For state that is more complex or shared across multiple widgets, use ChangeNotifier.\n* Use ListenableBuilder to listen to changes from a ChangeNotifier or other Listenable.\n* When a more robust solution is needed, structure the app using the Model-View-ViewModel (MVVM) pattern.\n* Use manual dependency injection via constructors to make a class's dependencies explicit in its API.\n* Use provider for dependency injection to make services, repositories, or complex state objects available to the UI layer without tight coupling (note: new-rules.md generally defaults against third-party packages for state management unless explicitly requested).\n\n### Data Flow\n\n* Define data structures (classes) to represent the data used in the application.\n* Abstract data sources (e.g., API calls, database operations) using Repositories/Services to promote testability.\n\n### Routing\n\n* Use go_router for declarative navigation, deep linking, and web support.\n\ndart\\n// 1. Add the dependency\\n// flutter pub add go_router\\n// 2. Configure the router\\nfinal GoRouter _router = GoRouter(\\n routes: [\\n GoRoute(\\n path: \\\'/\\\',\\n builder: (context, state) => const HomeScreen(),\\n routes: [\\n GoRoute(\\n path: \'details/:id\',\\n // Route with a path parameter\\n builder: (context, state) {\\n final String id = state.pathParameters[\'id\']!;\\n return DetailScreen(id: id);\\n },\\n ),\\n ],\\n ),\\n ],\\n);\\n// 3. Use it in your MaterialApp\\nMaterialApp.router(\\n routerConfig: _router,\\n);\\n\n\n* Use the built-in Navigator for short-lived screens that do not need to be deep-linkable, such as dialogs or temporary views.\n\ndart\\n// Push a new screen onto the stack\\nNavigator.push(\\n context,\\n MaterialPageRoute(builder: (context) => const DetailsScreen()),\\n);\\n// Pop the current screen to go back\\nNavigator.pop(context);\\n\n\n### Data Handling & Serialization\n\n* Use json_serializable and json_annotation for parsing and encoding JSON data.\n* When encoding data, use fieldRename: FieldRename.snake to convert Dart's camelCase fields to snake_case JSON keys.\n\ndart\\n// In your model file\\nimport \'package:json_annotation/json_annotation.dart\';\\npart \'user.g.dart\';\\n@JsonSerializable(fieldRename: FieldRename.snake)\\nclass User {\\n final String firstName;\\n final String lastName;\\n User({required this.firstName, required this.lastName});\\n factory User.fromJson(Map json) => _$UserFromJson(json);\\n Map toJson() => _$UserToJson(this);\\n}\\n\n\n### Logging\n\n* Use the log function from dart:developer for structured logging that integrates with Dart DevTools.\n\ndart\\nimport \'dart:developer\' as developer;\\n// For simple messages\\ndeveloper.log(\'User logged in successfully.\');\\n// For structured error logging\\ntry {\\n // ... code that might fail \\n} catch (e, s) {\\n developer.log( \'Failed to fetch data\',\\n name: \'myapp.network\',\\n level: 1000, // SEVERE error: e, stackTrace: s, \\n );\\n}\\n\n\n## Error Handling\n\n* Implement mechanisms to gracefully handle errors across the application (e.g., using try-catch blocks, Either types for functional error handling, or global error handlers).\n\n## Code Generation\n\n* Use build_runner for all code generation tasks, such as for json_serializable.\n* After modifying files that require code generation, run the build command:\n\nshell\\ndart run build_runner build --delete-conflicting-outputs\n\n## Testing\n\n* Use package:test for unit tests.\n* Use package:flutter_test for widget tests.\n* Use package:integration_test for integration tests.\n* Prefer using package:checks for more expressive and readable assertions over the default matchers.\n\n## Visual Design & Theming\n\n* Build beautiful and intuitive user interfaces that follow modern design guidelines.\n* Ensure the app is mobile responsive and adapts to different screen sizes, working perfectly on mobile and web.\n* If there are multiple pages for the user to interact with, provide an intuitive and easy navigation bar or controls.\n* Stress and emphasize font sizes to ease understanding, e.g., hero text, section headlines, list headlines, keywords in paragraphs.\n* Apply subtle noise texture to the main background to add a premium, tactile feel.\n* Multi-layered drop shadows create a strong sense of depth; cards have a soft, deep shadow to look \"lifted.\"\n* Incorporate icons to enhance the user’s understanding and the logical navigation of the app.\n* Buttons, checkboxes, sliders, lists, charts, graphs, and other interactive elements have a shadow with elegant use of color to create a \"glow\" effect.\n\n### Theming\n\n* Define a centralized ThemeData object to ensure a consistent application-wide style.\n* Use Material 3 by setting useMaterial3: true in your ThemeData.\n* Implement support for both light and dark themes, ideal for a user-facing theme toggle (ThemeMode.light, ThemeMode.dark, ThemeMode.system).\n* Generate harmonious color palettes from a single color using ColorScheme.fromSeed.\n\ndart\\nfinal ThemeData lightTheme = ThemeData(\\n useMaterial3: true,\\n colorScheme: ColorScheme.fromSeed(\\n seedColor: Colors.deepPurple,\\n brightness: Brightness.light,\\n ),\\n // ... other theme properties\\n);\\n\n\n* Include a wide range of color concentrations and hues in the palette to create a vibrant and energetic look and feel.\n* Use specific theme properties (e.g., appBarTheme, elevatedButtonTheme) to customize the appearance of individual Material components.\n* For custom fonts, use the google_fonts package. Define a TextTheme to apply fonts consistently.\n\ndart\\n// 1. Add the dependency\\n// flutter pub add google_fonts\\n// 2. Define a TextTheme with a custom font\\nfinal TextTheme appTextTheme = TextTheme(\\n displayLarge: GoogleFonts.oswald(fontSize: 57, fontWeight: FontWeight.bold),\\n titleLarge: GoogleFonts.roboto(fontSize: 22, fontWeight: FontWeight.w500),\\n bodyMedium: GoogleFonts.openSans(fontSize: 14),\\n);\\n\n\n### Assets and Images\n\n* If images are needed, make them relevant and meaningful, with appropriate size, layout, and licensing (e.g., freely available). Provide placeholder images if real ones are not available.\n* Declare all asset paths in your pubspec.yaml file.\n\nyaml\\nflutter:\\n uses-material-design: true\\n assets:\\n - assets/images/\\n\n\n* Use Image.asset to display local images from your asset bundle.\n\ndart\\nImage.asset(\'assets/images/placeholder.png\')\\n\n\n* Displays an icon from an ImageProvider, useful for custom icons not in Icons class.\n* Use Image.network to display images from a URL, and always include loadingBuilder and errorBuilder for a better user experience.\n\ndart\\nImage.network( \'https://picsum.photos/200/300\',\\n loadingBuilder: (context, child, progress) {\\n if (progress == null) return child;\\n return const Center(child: CircularProgressIndicator());\\n },\\n errorBuilder: (context, error, stackTrace) {\\n return const Icon(Icons.error);\\n },\\n)\\n\n\n## Accessibility (A11Y)\n\n* Implement accessibility features to empower all users, assuming a wide variety of users with different physical abilities, mental abilities, age groups, education levels, and learning styles."}', 'reasoning_content': None, 'tool_calls': None}, 'logprobs': None, 'finish_reason': 'stop'}], 'usage': {'prompt_tokens': 6735, 'total_tokens': 11104, 'completion_tokens': 4369}}