发布日期: 2025-08-28
版本号: v0.6.26

0.6.26版本（2025-08-28）引入了多项功能增强、错误修复和系统改进。在功能方面，新增了细粒度的聊天交互权限控制，允许管理员对单个聊天操作设置权限；支持自定义AI响应中的推理标签检测；添加了下拉刷新功能以解决长时间生成响应时的界面超时问题；文件上传API现在支持后台或同步处理模式；增强了Azure Document Intelligence的认证支持，兼容DefaultAzureCredential；为仅SSO部署添加了初始管理员注册和强制登录表单的配置选项；引入了搜索查询生成缓存机制以减少重复的LLM调用；允许配置工具调用的最大重试次数；提供了精简版Docker镜像选项以减小镜像体积；在工作区提示词页面添加了Shift+删除的快捷键功能；优化了PDF导出以减小文件大小；通过pip安装时默认采用精简依赖，需手动添加全部依赖；进行了后端重构、无障碍功能增强以及多语言本地化更新。在修复方面，解决了聊天生成错误时无用户反馈、不适用的图像生成步骤设置显示、Datalab文档加载器路径错误、引用对象错误导致界面无响应、群组用户添加API失效、外部工具服务器错误处理、代码块复制按钮内容、PDF导出内容不匹配、Windows侧边栏鼠标指针图标、Safari浏览器输入法组合键、混合搜索参数处理及除零错误、右向左文本对齐、文件内容预览模态框空白以及Playwright超时配置错误等一系列问题。此外，还移除了后续问题建议严格遵循聊天主要语言设置的限制，使建议生成更为灵活。

更新内容 （中文）

[0.6.26] - 2025-08-28

新增

  • 🛂 细粒度聊天互动权限：为“继续响应”、“重新生成响应”、“评分响应”和“删除消息”等单个聊天操作添加了精细的权限控制。管理员现在可以按用户组或通过环境变量设置系统默认值来配置这些权限，通过防止响应继续中可能的系统提示泄露，增强安全性与治理能力，并实现对用户AI响应交互的精确控制。
  • 🧠 自定义推理标签配置：为AI模型响应添加了可配置的推理标签检测，允许管理员和用户自定义系统识别和处理推理内容的方式。用户现在可以通过高级参数界面定义自定义推理标签对、使用默认标签（如"think"和"reasoning"）或完全禁用推理检测，增强了对AI思维过程可见性的控制。
  • 📱 下拉刷新支持：添加了下拉刷新功能，允许用户通过在导航栏区域向下拉动轻松刷新界面。这解决了在长AI响应生成过程中暂时切换应用时出现的超时问题，无需关闭并重新启动PWA。
  • 📁 可配置文件上传处理模式：在文件上传API端点添加了"process_in_background"查询参数，允许客户端在异步（默认）和同步文件处理之间进行选择。设置"process_in_background=false"会强制上传请求等待直到提取和嵌入完成，返回立即可用的文件，并简化了偏好阻塞调用而非轮询工作流的后端API消费者的集成。
  • 🔐 Azure Document Intelligence DefaultAzureCredential支持：除了API密钥认证外，添加了使用DefaultAzureCredential进行Azure Document Intelligence身份验证的支持，实现了与Azure Entra ID和托管身份认证的无缝集成，适用于企业Azure环境。
  • 🔐 身份验证引导增强：添加了"ENABLE_INITIAL_ADMIN_SIGNUP"环境变量和"?form=true" URL参数，以在仅SSO部署中启用初始管理员用户创建和强制显示登录表单。这解决了当登录表单被禁用时管理员无法创建首个用户的引导问题，允许正确初始化SSO配置的部署，无需临时更改配置。
  • ⚡ 查询生成缓存：添加了"ENABLE_QUERIES_CACHE"环境变量，以启用生成搜索查询的请求范围缓存。当网络搜索和文件检索都激活时，为网络搜索生成的查询会自动重用于文件检索，消除了重复的LLM API调用，减少了令牌使用和成本，同时保持搜索质量。
  • 🔧 可配置工具调用重试限制：添加了"CHAT_RESPONSE_MAX_TOOL_CALL_RETRIES"环境变量，以控制在安全停止聊天会话之前允许的最大连续工具调用次数。这取代了之前硬编码的10次限制，允许管理员为需要大量工具交互的复杂工作流配置更高的限制。
  • 📦 精简Docker镜像变体：通过"USE_SLIM"构建参数添加了新的精简Docker镜像选项，该选项排除了嵌入式AI模型和Ollama安装，将镜像大小减少了约1GB。此变体使AI模型在外部管理的环境中能够更快地拉取和部署镜像，对自动扩展集群和分布式部署特别有益。
  • 🗂️ 工作区提示的Shift删除功能：为工作区提示页面的快速提示删除添加了键盘快捷键支持。按住Shift并悬停在任何提示上即可显示垃圾桶图标进行即时删除，在所有工作区部分（模型、工具、函数以及现在的提示）引入了一致的交互模式，简化了提示管理工作流程。
  • ♿ 无障碍性增强：通过改进关键平台组件的键盘导航、ARIA标签和屏幕阅读器兼容性，增强了用户界面的可访问性。
  • 📄 优化PDF导出以减小文件大小：PDF导出现在已显著优化，生成的文件更小，便于更快下载和更轻松地共享或归档您的聊天和文档。
  • 📦 精简默认安装，可选完整依赖：通过pip安装Open WebUI现在默认为更精简的包；PostgreSQL支持不再默认包含——只需使用’pip install open-webui[all]‘即可包含所有可选依赖项，以实现完整的功能兼容性。
  • 🔄 通用后端重构：实施了各种后端改进以提升性能、稳定性和安全性，为所有用户提供更具弹性和可靠性的平台。
  • 🌐 本地化与国际化改进：增强并扩展了芬兰语、西班牙语、日语、波兰语、葡萄牙语（巴西）和中文的翻译，包括缺失翻译和拼写错误更正，为这些语言的使用者在整个界面中提供更自然和专业的用户体验。

修复

  • ⚠️ 聊天错误反馈已恢复：修复了各种后端错误（工具服务器故障、API连接问题、格式错误的响应）会导致聊天无限加载而不提供用户反馈的问题。现在系统在聊天生成过程中发生故障时会正确显示错误消息，使用户能够理解问题并根据需要重试，而不是无限期等待。
  • 🖼️ 图像生成步骤设置可见性已修复：修复了一个UI问题，即为不支持此参数的OpenAI和Gemini图像生成引擎错误地显示了"设置步骤"配置选项。此设置现在仅对兼容的引擎（如ComfyUI和Automatic1111）可见，消除了用户对无效配置选项的困惑。
  • 📄 Datalab Marker API文档加载器已修复：通过纠正托管Datalab服务和自托管Marker服务器的URL路径处理，修复了损坏的Datalab Marker API文档加载器功能。从加载器代码中删除了硬编码的"/marker"路径，并恢复了正确的默认URL结构，确保PDF和文档处理在两种部署类型下均能正常工作。
  • 📋 引用错误处理已改进：修复了来自外部工具、管道或过滤器的格式错误的引用或源对象会导致JavaScript错误并使聊天界面完全无响应的问题。系统现在能优雅地处理缺失或未定义的引用属性，即使工具生成有缺陷的引用事件，也能让对话正常加载。
  • 👥 组用户添加API端点已修复：修复了"/api/v1/groups/id/{group_id}/users/add" API端点会接受请求但不报错，却无法实际将用户添加到组中的问题。系统现在能正确初始化和去重用户ID列表，确保通过API调用正确地将用户添加到组中和从组中移除。
  • 🛠️ 外部工具服务器错误处理已改进：修复了不可达或配置错误的外部工具服务器会导致JavaScript错误并阻止界面正确加载的问题。系统现在能优雅地处理连接失败，显示适当的错误消息，并在保持正常连接的全部功能的同时，过滤掉无法访问的服务器。
  • 📋 代码块复制按钮内容已修复：修复了代码块中的复制按钮会复制原始AI生成的代码而不是任何用户编辑内容的问题，确保复制功能始终捕获用户修改后的当前显示代码。
  • 📄 PDF导出内容不匹配已修复：解决了在查看另一个聊天时导出一个聊天的PDF会错误地使用当前查看的聊天内容而非目标聊天内容生成PDF的问题。此外，通过改进的画布切片、更好的内存管理和增强的图像质量优化了PDF生成算法，同时从单个聊天菜单中移除了有问题的PDF导出选项，以防止进一步混淆。
  • 🖱️ Windows侧边栏光标图标已更正：修复了Windows系统上令人困惑的光标图标问题，侧边栏切换按钮显示调整大小光标（ew-resize）而非适当的指针光标。侧边栏按钮现在在Windows上显示标准指针光标，消除了用户对于按钮是展开/折叠侧边栏还是调整其大小的困惑。
  • 📝 Safari输入法组合错误已修复：解决了在Safari上使用输入法（IME）组合中文文本时按Enter键会过早发送消息而非完成文本组合的问题。系统现在能正确检测组合状态，并忽略组合结束后立即发生的keydown事件，确保所有浏览器上流畅的多语言文本输入。
  • 🔍 混合搜索参数处理已修复：修复了集合查询请求中的"hybrid"参数未被正确评估的问题，导致系统忽略用户指定的混合搜索偏好，仅检查全局配置。此外解决了当BM25Retriever在空文档列表上调用时混合搜索中出现的除以零错误，确保了所有集合状态下搜索功能的稳健性。
  • 💬 消息中的RTL文本方向已修复：修复了用户消息和AI响应中从右到左（RTL）语言的文本对齐问题，确保根据用户语言设置正确设置文本方向。代码块现在无论用户语言偏好如何，均一致使用从左到右（LTR）方向，保持所有支持语言的代码可读性。
  • 📁 模态框中的文件内容预览已恢复：修复了点击上传文件时会显示空白预览模态框的问题，即使文件已成功处理并可用于AI上下文。文件内容现在在预览模态框中正确显示，确保用户可以按照预期验证和查看其上传的文档。
  • 🌐 Playwright超时配置已更正：修复了Playwright超时值从毫秒到秒的转换不正确，并额外乘以1000倍，导致网络加载超时过长的问题。超时参数现在按预期正确使用配置的毫秒值，确保响应迅速的网络搜索和文档加载操作。

变更

  • 🔄 后续问题语言约束已移除：后续问题建议不再严格遵守聊天的主要语言设置，允许基于对话上下文和相关性生成更灵活、多样化的建议，可能包含不同语言的问题，而非强制的语言匹配。

更新内容 （原始）

[0.6.26] - 2025-08-28

Added

  • 🛂 Granular Chat Interaction Permissions: Added fine-grained permission controls for individual chat actions including “Continue Response”, “Regenerate Response”, “Rate Response”, and “Delete Messages”. Administrators can now configure these permissions per user group or set system defaults via environment variables, providing enhanced security and governance by preventing potential system prompt leakage through response continuation and enabling precise control over user interactions with AI responses.
  • 🧠 Custom Reasoning Tags Configuration: Added configurable reasoning tag detection for AI model responses, allowing administrators and users to customize how the system identifies and processes reasoning content. Users can now define custom reasoning tag pairs, use default tags like “think” and “reasoning”, or disable reasoning detection entirely through the Advanced Parameters interface, providing enhanced control over AI thought process visibility.
  • 📱 Pull-to-Refresh Support: Added pull-to-refresh functionality allowing user to easily refresh the interface by pulling down on the navbar area. This resolves timeout issues that occurred when temporarily switching away from the app during long AI response generations, eliminating the need to close and relaunch the PWA.
  • 📁 Configurable File Upload Processing Mode: Added “process_in_background” query parameter to the file upload API endpoint, allowing clients to choose between asynchronous (default) and synchronous file processing. Setting “process_in_background=false” forces the upload request to wait until extraction and embedding complete, returning immediately usable files and simplifying integration for backend API consumers that prefer blocking calls over polling workflows.
  • 🔐 Azure Document Intelligence DefaultAzureCredential Support: Added support for authenticating with Azure Document Intelligence using DefaultAzureCredential in addition to API key authentication, enabling seamless integration with Azure Entra ID and managed identity authentication for enterprise Azure environments.
  • 🔐 Authentication Bootstrapping Enhancements: Added “ENABLE_INITIAL_ADMIN_SIGNUP” environment variable and “?form=true” URL parameter to enable initial admin user creation and forced login form display in SSO-only deployments. This resolves bootstrap issues where administrators couldn’t create the first user when login forms were disabled, allowing proper initialization of SSO-configured deployments without requiring temporary configuration changes.
  • ⚡ Query Generation Caching: Added “ENABLE_QUERIES_CACHE” environment variable to enable request-scoped caching of generated search queries. When both web search and file retrieval are active, queries generated for web search are automatically reused for file retrieval, eliminating duplicate LLM API calls and reducing token usage and costs while maintaining search quality.
  • 🔧 Configurable Tool Call Retry Limit: Added “CHAT_RESPONSE_MAX_TOOL_CALL_RETRIES” environment variable to control the maximum number of sequential tool calls allowed before safety stopping a chat session. This replaces the previous hardcoded limit of 10, enabling administrators to configure higher limits for complex workflows requiring extensive tool interactions.
  • 📦 Slim Docker Image Variant: Added new slim Docker image option via “USE_SLIM” build argument that excludes embedded AI models and Ollama installation, reducing image size by approximately 1GB. This variant enables faster image pulls and deployments for environments where AI models are managed externally, particularly beneficial for auto-scaling clusters and distributed deployments.
  • 🗂️ Shift-to-Delete Functionality for Workspace Prompts: Added keyboard shortcut support for quick prompt deletion on the Workspace Prompts page. Hold Shift and hover over any prompt to reveal a trash icon for instant deletion, bringing consistent interaction patterns across all workspace sections (Models, Tools, Functions, and now Prompts) and streamlining prompt management workflows.
  • ♿ Accessibility Enhancements: Enhanced user interface accessibility with improved keyboard navigation, ARIA labels, and screen reader compatibility across key platform components.
  • 📄 Optimized PDF Export for Smaller File Size: PDF exports are now significantly optimized, producing much smaller files for faster downloads and easier sharing or archiving of your chats and documents.
  • 📦 Slimmed Default Install with Optional Full Dependencies: Installing Open WebUI via pip now defaults to a slimmer package; PostgreSQL support is no longer included by default—simply use ‘pip install open-webui[all]’ to include all optional dependencies for full feature compatibility.
  • 🔄 General Backend Refactoring: Implemented various backend improvements to enhance performance, stability, and security, ensuring a more resilient and reliable platform for all users.
  • 🌐 Localization & Internationalization Improvements: Enhanced and expanded translations for Finnish, Spanish, Japanese, Polish, Portuguese (Brazil), and Chinese, including missing translations and typo corrections, providing a more natural and professional user experience for speakers of these languages across the entire interface.

Fixed

  • ⚠️ Chat Error Feedback Restored: Fixed an issue where various backend errors (tool server failures, API connection issues, malformed responses) would cause chats to load indefinitely without providing user feedback. The system now properly displays error messages when failures occur during chat generation, allowing users to understand issues and retry as needed instead of waiting indefinitely.
  • 🖼️ Image Generation Steps Setting Visibility Fixed: Fixed a UI issue where the “Set Steps” configuration option was incorrectly displayed for OpenAI and Gemini image generation engines that don’t support this parameter. The setting is now only visible for compatible engines like ComfyUI and Automatic1111, eliminating user confusion about non-functional configuration options.
  • 📄 Datalab Marker API Document Loader Fixed: Fixed broken Datalab Marker API document loader functionality by correcting URL path handling for both hosted Datalab services and self-hosted Marker servers. Removed hardcoded “/marker” paths from the loader code and restored proper default URL structure, ensuring PDF and document processing works correctly with both deployment types.
  • 📋 Citation Error Handling Improved: Fixed an issue where malformed citation or source objects from external tools, pipes, or filters would cause JavaScript errors and make the chat interface completely unresponsive. The system now gracefully handles missing or undefined citation properties, allowing conversations to load properly even when tools generate defective citation events.
  • 👥 Group User Add API Endpoint Fixed: Fixed an issue where the “/api/v1/groups/id/{group_id}/users/add” API endpoint would accept requests without errors but fail to actually add users to groups. The system now properly initializes and deduplicates user ID lists, ensuring users are correctly added to and removed from groups via API calls.
  • 🛠️ External Tool Server Error Handling Improved: Fixed an issue where unreachable or misconfigured external tool servers would cause JavaScript errors and prevent the interface from loading properly. The system now gracefully handles connection failures, displays appropriate error messages, and filters out inaccessible servers while maintaining full functionality for working connections.
  • 📋 Code Block Copy Button Content Fixed: Fixed an issue where the copy button in code blocks would copy the original AI-generated code instead of any user-edited content, ensuring the copy function always captures the currently displayed code as modified by users.
  • 📄 PDF Export Content Mismatch Fixed: Resolved an issue where exporting a PDF of one chat while viewing another chat would incorrectly generate the PDF using the currently viewed chat’s content instead of the intended chat’s content. Additionally optimized the PDF generation algorithm with improved canvas slicing, better memory management, and enhanced image quality, while removing the problematic PDF export option from individual chat menus to prevent further confusion.
  • 🖱️ Windows Sidebar Cursor Icon Corrected: Fixed confusing cursor icons on Windows systems where sidebar toggle buttons displayed resize cursors (ew-resize) instead of appropriate pointer cursors. The sidebar buttons now show standard pointer cursors on Windows, eliminating user confusion about whether the buttons expand/collapse the sidebar or resize it.
  • 📝 Safari IME Composition Bug Fixed: Resolved an issue where pressing Enter while composing Chinese text using Input Method Editors (IMEs) on Safari would prematurely send messages instead of completing text composition. The system now properly detects composition states and ignores keydown events that occur immediately after composition ends, ensuring smooth multilingual text input across all browsers.
  • 🔍 Hybrid Search Parameter Handling Fixed: Fixed an issue where the “hybrid” parameter in collection query requests was not being properly evaluated, causing the system to ignore user-specified hybrid search preferences and only check global configuration. Additionally resolved a division by zero error that occurred in hybrid search when BM25Retriever was called with empty document lists, ensuring robust search functionality across all collection states.
  • 💬 RTL Text Orientation in Messages Fixed: Fixed text alignment issues in user messages and AI responses for Right-to-Left languages, ensuring proper text direction based on user language settings. Code blocks now consistently use Left-to-Right orientation regardless of the user’s language preference, maintaining code readability across all supported languages.
  • 📁 File Content Preview in Modal Restored: Fixed an issue where clicking on uploaded files would display an empty preview modal, even when the files were successfully processed and available for AI context. File content now displays correctly in the preview modal, ensuring users can verify and review their uploaded documents as intended.
  • 🌐 Playwright Timeout Configuration Corrected: Fixed an issue where Playwright timeout values were incorrectly converted from milliseconds to seconds with an additional 1000x multiplier, causing excessively long web loading timeouts. The timeout parameter now correctly uses the configured millisecond values as intended, ensuring responsive web search and document loading operations.

Changed

  • 🔄 Follow-Up Question Language Constraint Removed: Follow-up question suggestions no longer strictly adhere to the chat’s primary language setting, allowing for more flexible and diverse suggestion generation that may include questions in different languages based on conversation context and relevance rather than enforced language matching.

下载链接