发布日期: 2025-10-07
版本号: v0.6.33

[0.6.33]版本于2025年10月8日发布，主要包含以下更新：对工作区界面（模型、知识、提示词和工具）及函数管理界面进行了全面重新设计，优化了控制布局、视图过滤和组织方式。通过并行数据加载和动态加载等技术显著提升了页面初始化性能。新增了文件夹附加至聊天、安卓系统分享支持、Redis实验性会话存储、Vega/Vega-Lite图表渲染、自定义HTTP头、OpenID Connect无邮箱认证、Ollama模型管理增强、OneDrive搜索功能、TTS参数配置、工具服务器信息展示、文件夹记忆所选模型、Ollama向量维度参数、笔记PDF导出优化、Docling文档提取参数支持以及服务器启动脚本参数传递等功能。修复了系统提示重复、OAuth认证流程（包括PKCE、令牌刷新、元数据发现等）、文件夹数据完整性、图表渲染错误处理、模型选择器搜索、任务函数调用性能、任务取消状态反馈、临时聊天建议生成、Azure TTS特殊字符处理、OpenAPI工具调用、MCP模式验证、第三方工具响应处理、工作区过滤器渲染、键盘快捷键冲突、粘贴格式保留、文本选择复制以及日志级别等多项问题。同时，更新了Milvus向量数据库依赖，并完善了多语言翻译。

更新内容 （中文）

[0.6.33] - 2025-10-08

新增

  • 🎨 工作区界面在模型、知识库、提示词和工具部分进行了全面重新设计，包括重组的控件、针对“已创建”与“已共享”项目的视图过滤器、标签选择器、改进的视觉层次结构以及简化的导入/导出功能。提交，提交，提交，提交，提交，提交
  • 🛠️ 函数管理界面进行了全面重新设计，增加了创建者归属显示、针对“已创建”与“已共享”项目的所有权过滤器、改进的组织结构和精炼的样式。提交
  • ⚡ 通过并行数据加载和优化的文件夹 API 调用，显著提升了页面初始化性能，减少了初始页面加载时间。#17559，#17889
  • ⚡ 聊天概览组件现在按需动态加载，将初始页面包大小减少了约 470KB，提升了首屏加载速度。#17595
  • 📁 现在可以使用“#”命令将文件夹附加到聊天中，该命令会自动展开以包含文件夹内的所有文件，简化了知识库集成。提交
  • 📱 渐进式 Web 应用现在支持 Android 分享目标功能，允许用户通过系统分享菜单将网页、YouTube 视频和文本直接分享到 Open WebUI。#17633，#17125
  • 🗄️ Redis 会话存储现在可通过 ENABLE_STAR_SESSIONS_MIDDLEWARE 环境变量作为 OAuth 认证流程的实验性选项使用，为多副本部署提供共享会话状态以解决 CSRF 错误，但目前仅支持基本的 Redis 设置。#17223，#15373，文档：提交
  • 📊 代码块中现在支持 Vega 和 Vega-Lite 图表可视化渲染器，能够以内联方式渲染数据可视化，并自动编译 Vega-Lite 规范。#18033，#18040，#18022
  • 🔗 OpenAI 连接现在支持自定义 HTTP 头，允许用户为特定部署需求配置认证和路由头。#18021，#9732
  • 🔐 OpenID Connect 认证现在支持通过 ENABLE_OAUTH_WITHOUT_EMAIL 环境变量启用无邮箱作用域的 OIDC 提供商，从而兼容不公开电子邮件地址的身份提供商。#18047，#18045
  • 🤖 Ollama 模型管理模态框现在具有单独的模型更新取消功能、所有按钮的综合工具提示以及简化的通知行为，以减少吐司提示轰炸。#16863
  • ☁️ OneDrive 文件选择器现在包括搜索功能和企业账户的“我的组织”透视表，便于在组织内容中查找文件。#17930，#17929
  • 📊 聊天概览流程图现在支持在垂直和水平布局方向之间切换，以提高可视化灵活性。#17941
  • 🔊 OpenAI 文字转语音引擎现在支持额外参数，允许用户通过 JSON 配置使用提供商特定选项来自定义 TTS 行为。#17985，#17188
  • 🛠️ 工具服务器列表现在显示服务器名称、URL 和类型（OpenAPI 或 MCP），便于识别和管理。#18062
  • 📁 文件夹现在会记住上次选择的模型，并在该文件夹内启动新聊天时自动应用。#17836
  • 🔢 Ollama 嵌入端点现在支持可选的 dimensions 参数以控制嵌入输出大小，兼容 Ollama v0.11.11 及更高版本。#17942
  • ⚡ 通过移除冗余 API 调用，改进了工作区知识页面的加载时间，提升了整体响应速度。#18057
  • ⚡ 通过仅选择相关列而非检索整个记录，提升了文件元数据查询性能，减少了数据库开销。#18013
  • 📄 笔记 PDF 导出现在包含标题，并能在深色模式下正确渲染，具有适当的背景颜色。提交
  • 📄 Docling 文档提取现在支持用于 VLM 流水线配置的额外参数，能够进行自定义视觉模型设置。#17363
  • ⚙️ 服务器启动脚本现在支持向 uvicorn 传递任意参数，实现自定义服务器配置选项。#17919，#17918
  • 🔄 在前端和后端实施了多项改进，以增强性能、稳定性和安全性。
  • 🌐 增强并扩展了德语、丹麦语、西班牙语、韩语、葡萄牙语（巴西）、简体中文和繁体中文的翻译。

修复

  • 💬 系统提示词不再在聊天请求中重复，消除了因向模型发送重复指令而导致的混淆和过多的令牌使用。#17198，#16855
  • 🔐 MCP OAuth 2.1 认证现在符合标准，实施了 S256 代码挑战方法的 PKCE，并在令牌授权期间显式传递客户端凭据，解决了连接到受 OAuth 保护的 MCP 服务器时出现的“code_challenge: Field required”和“client_id: Field required”错误。提交，#18010，#18087
  • 🔐 OAuth 注册流程现在通过从 passlib 迁移到原生 bcrypt 来正确处理密码哈希，防止密码超过 72 字节时失败。#17917
  • 🔐 通过正确注册和存储 OAuth 客户端，解决了 OAuth 令牌刷新错误，修复了 Google、Microsoft 和 OIDC 提供商的“Constructor parameter should be str”异常。#17829
  • 🔐 OAuth 服务器元数据 URL 现在通过正确的属性访问，修复了当未设置 OPENID_PROVIDER_URL 时，Microsoft OAuth 提供商的自动令牌刷新和登出功能。#18065
  • 🔐 OAuth 凭据解密失败现在允许应用程序以清晰的错误消息优雅启动，而不是崩溃，在数据库迁移或环境更改期间发生 WEBUI_SECRET_KEY 不匹配时可防止完全服务中断。#18094，#18092
  • 🔐 OAuth 2.1 服务器发现现在正确地按顺序尝试所有配置的发现 URL，而不是只尝试第一个 URL。#17906，#17904，#18026
  • 🔐 登录重定向现在在认证后正确遵守重定向查询参数，确保用户被返回到其预期的目的地并保留查询参数，而不是默认返回主页。#18071
  • ☁️ 修复了 OneDrive 商业集成认证回退问题，确保连接到 OneDrive 帐户时弹出窗口能正常触发。#17902，#17825，#17816
  • 👥 默认群组设置现在在页面导航后正确持久化，确保配置更改被正确保存和保留。#17899，#18003
  • 📁 检索时现在会验证文件夹数据完整性，自动修复具有无效父引用的孤立文件夹，并确保正确级联删除嵌套的文件夹结构。提交
  • 🗄️ Redis Sentinel 和 Redis Cluster 配置与实验性 ENABLE_STAR_SESSIONS_MIDDLEWARE 功能现在通过仅将该功能设为可选来正确隔离，防止在多节点 Redis 部署中连接到只读副本时出现 ReadOnlyError 失败。#18073
  • 📊 Mermaid 和 Vega 图表渲染现在在检测到语法错误时显示错误吐司通知，帮助用户识别和修复图表问题，而不是静默失败。#18068
  • 🤖 返回 reasoning_content 而非 content 的推理模型在聊天标题生成、后续建议和标签生成期间不再导致 NoneType 错误。#18080
  • 📚 引用渲染现在能正确处理单个方括号中的多个来源引用，将 [1,2] 和 [1, 2] 等格式解析为单独的可点击引用链接。#18120
  • 🔍 网页搜索现在可以优雅地处理单个来源失败，当某个 URL 无法访问或返回错误时，继续处理剩余来源而不是完全失败。提交，提交
  • 🔍 带重排序的混合搜索现在可以优雅地处理空结果集，而不是在因相关性阈值过滤掉所有结果时因 ValueError 而崩溃。#18096
  • 🔍 没有定义填充令牌的重排序模型现在可以通过自动回退到 eos_token_id 作为 pad_token_id 来正常工作，修复了 Qwen3-Reranker 等模型的“Cannot handle batch sizes > 1”错误。#18108，#16027
  • 🔍 模型选择器搜索现在可以通过在模型列表更改时动态更新搜索索引，为非管理员用户正确返回结果，修复了导致空搜索结果的竞争条件。#17996，#17960
  • ⚡ 通过从有效载荷中排除 base64 图像数据，提升了任务模型函数调用性能，当对话中包含图像时显著减少了令牌数量和内存使用。#17897
  • 🤖 文本选择“提问”操作现在可以正确识别并使用通过直接连接配置的本地模型，而不是仅显示外部提供商模型。#17896
  • 🛑 任务取消 API 现在返回准确的响应状态，正确报告成功的取消操作，而不是错误地指示失败。#17920
  • 💬 后续查询建议现在在临时聊天中生成和显示，与已保存聊天的行为匹配。#14987
  • 🔊 Azure 文字转语音现在可以正确转义 SSML 中的特殊字符（如 & 符号），防止 HTTP 400 错误，确保所有文本内容的音频生成成功。#17962
  • 🛠️ 带可选参数的 OpenAPI 工具服务器调用现在即使在未提供参数时也能成功执行，移除了对请求正文的不正确要求。#18036
  • 🛠️ MCP 模式工具服务器连接现在不再错误地验证 OpenAPI 路径字段，允许在 OpenAPI 和 MCP 连接类型之间无缝切换。#17989，#17988
  • 🛠️ 包含非 UTF8 或无效字节序列的第三方工具响应现在可以优雅处理，不会导致请求失败。#17882
  • 🎨 工作区过滤器下拉菜单现在可以将模型标签正确渲染为字符串，而不是显示单个字符，修复了模型具有多个标签时过滤界面损坏的问题。#18034
  • ⌨️ 在 Chrome 的移动和窄浏览器视图中，Ctrl+Enter 快捷键现在可以正确发送消息，而不是插入换行符。#17975
  • ⌨️ 在纯文本模式下将代码或格式化文本粘贴到聊天输入框时，Tab 字符现在会被保留。#17958
  • 📋 从聊天输入框复制文本选择现在会正确复制仅选中的文本，而不是整个文本框内容。#17911
  • 🔍 网页搜索查询日志现在使用调试级别而非信息级别，防止用户搜索查询出现在生产日志中。#17888
  • 📝 移除了中间件中的调试打印语句，以防止过度的日志污染并尊重配置的日志记录级别。#17943

变更

  • 🗄️ Milvus 向量数据库依赖从 pymilvus 2.5.0 更新到 2.6.2，确保与更新版本的 Milvus 兼容，但要求使用旧版 Milvus 实例的用户升级其数据库或手动降级 pymilvus 包。#18066

更新内容 （原始）

[0.6.33] - 2025-10-08

Added

  • 🎨 Workspace interface received a comprehensive redesign across Models, Knowledge, Prompts, and Tools sections, featuring reorganized controls, view filters for created vs shared items, tag selectors, improved visual hierarchy, and streamlined import/export functionality. Commit, Commit, Commit, Commit, Commit, Commit
  • 🛠️ Functions admin interface received a comprehensive redesign with creator attribution display, ownership filters for created vs shared items, improved organization, and refined styling. Commit
  • ⚡ Page initialization performance is significantly improved through parallel data loading and optimized folder API calls, reducing initial page load time. #17559, #17889
  • ⚡ Chat overview component is now dynamically loaded on demand, reducing initial page bundle size by approximately 470KB and improving first-screen loading speed. #17595
  • 📁 Folders can now be attached to chats using the “#” command, automatically expanding to include all files within the folder for streamlined knowledge base integration. Commit
  • 📱 Progressive Web App now supports Android share target functionality, allowing users to share web pages, YouTube videos, and text directly to Open WebUI from the system share menu. #17633, #17125
  • 🗄️ Redis session storage is now available as an experimental option for OAuth authentication flows via the ENABLE_STAR_SESSIONS_MIDDLEWARE environment variable, providing shared session state across multi-replica deployments to address CSRF errors, though currently only basic Redis setups are supported. #17223, #15373, Docs:Commit
  • 📊 Vega and Vega-Lite chart visualization renderers are now supported in code blocks, enabling inline rendering of data visualizations with automatic compilation of Vega-Lite specifications. #18033, #18040, #18022
  • 🔗 OpenAI connections now support custom HTTP headers, enabling users to configure authentication and routing headers for specific deployment requirements. #18021, #9732
  • 🔐 OpenID Connect authentication now supports OIDC providers without email scope via the ENABLE_OAUTH_WITHOUT_EMAIL environment variable, enabling compatibility with identity providers that don’t expose email addresses. #18047, #18045
  • 🤖 Ollama model management modal now features individual model update cancellation, comprehensive tooltips for all buttons, and streamlined notification behavior to reduce toast spam. #16863
  • ☁️ OneDrive file picker now includes search functionality and “My Organization” pivot for business accounts, enabling easier file discovery across organizational content. #17930, #17929
  • 📊 Chat overview flow diagram now supports toggling between vertical and horizontal layout orientations for improved visualization flexibility. #17941
  • 🔊 OpenAI Text-to-Speech engine now supports additional parameters, allowing users to customize TTS behavior with provider-specific options via JSON configuration. #17985, #17188
  • 🛠️ Tool server list now displays server name, URL, and type (OpenAPI or MCP) for easier identification and management. #18062
  • 📁 Folders now remember the last selected model, automatically applying it when starting new chats within that folder. #17836
  • 🔢 Ollama embedding endpoint now supports the optional dimensions parameter for controlling embedding output size, compatible with Ollama v0.11.11 and later. #17942
  • ⚡ Workspace knowledge page load time is improved by removing redundant API calls, enhancing overall responsiveness. #18057
  • ⚡ File metadata query performance is enhanced by selecting only relevant columns instead of retrieving entire records, reducing database overhead. #18013
  • 📄 Note PDF exports now include titles and properly render in dark mode with appropriate background colors. Commit
  • 📄 Docling document extraction now supports additional parameters for VLM pipeline configuration, enabling customized vision model settings. #17363
  • ⚙️ Server startup script now supports passing arbitrary arguments to uvicorn, enabling custom server configuration options. #17919, #17918
  • 🔄 Various improvements were implemented across the frontend and backend to enhance performance, stability, and security.
  • 🌐 Translations for German, Danish, Spanish, Korean, Portuguese (Brazil), Simplified Chinese, and Traditional Chinese were enhanced and expanded.

Fixed

  • 💬 System prompts are no longer duplicated in chat requests, eliminating confusion and excessive token usage caused by repeated instructions being sent to models. #17198, #16855
  • 🔐 MCP OAuth 2.1 authentication now complies with the standard by implementing PKCE with S256 code challenge method and explicitly passing client credentials during token authorization, resolving “code_challenge: Field required” and “client_id: Field required” errors when connecting to OAuth-secured MCP servers. Commit, #18010, #18087
  • 🔐 OAuth signup flow now handles password hashing correctly by migrating from passlib to native bcrypt, preventing failures when passwords exceed 72 bytes. #17917
  • 🔐 OAuth token refresh errors are resolved by properly registering and storing OAuth clients, fixing “Constructor parameter should be str” exceptions for Google, Microsoft, and OIDC providers. #17829
  • 🔐 OAuth server metadata URL is now correctly accessed via the proper attribute, fixing automatic token refresh and logout functionality for Microsoft OAuth provider when OPENID_PROVIDER_URL is not set. #18065
  • 🔐 OAuth credential decryption failures now allow the application to start gracefully with clear error messages instead of crashing, preventing complete service outages when WEBUI_SECRET_KEY mismatches occur during database migrations or environment changes. #18094, #18092
  • 🔐 OAuth 2.1 server discovery now correctly attempts all configured discovery URLs in sequence instead of only trying the first URL. #17906, #17904, #18026
  • 🔐 Login redirect now correctly honors the redirect query parameter after authentication, ensuring users are returned to their intended destination with query parameters intact instead of defaulting to the homepage. #18071
  • ☁️ OneDrive Business integration authentication regression is resolved, ensuring the popup now properly triggers when connecting to OneDrive accounts. #17902, #17825, #17816
  • 👥 Default group settings now persist correctly after page navigation, ensuring configuration changes are properly saved and retained. #17899, #18003
  • 📁 Folder data integrity is now verified on retrieval, automatically fixing orphaned folders with invalid parent references and ensuring proper cascading deletion of nested folder structures. Commit
  • 🗄️ Redis Sentinel and Redis Cluster configurations with the experimental ENABLE_STAR_SESSIONS_MIDDLEWARE feature are now properly isolated by making the feature opt-in only, preventing ReadOnlyError failures when connecting to read replicas in multi-node Redis deployments. #18073
  • 📊 Mermaid and Vega diagram rendering now displays error toast notifications when syntax errors are detected, helping users identify and fix diagram issues instead of silently failing. #18068
  • 🤖 Reasoning models that return reasoning_content instead of content no longer cause NoneType errors during chat title generation, follow-up suggestions, and tag generation. #18080
  • 📚 Citation rendering now correctly handles multiple source references in a single bracket, parsing formats like [1,2] and [1, 2] into separate clickable citation links. #18120
  • 🔍 Web search now handles individual source failures gracefully, continuing to process remaining sources instead of failing entirely when a single URL is unreachable or returns an error. Commit, Commit
  • 🔍 Hybrid search with reranking now handles empty result sets gracefully instead of crashing with ValueError when all results are filtered out due to relevance thresholds. #18096
  • 🔍 Reranking models without defined padding tokens now work correctly by automatically falling back to eos_token_id as pad_token_id, fixing “Cannot handle batch sizes > 1” errors for models like Qwen3-Reranker. #18108, #16027
  • 🔍 Model selector search now correctly returns results for non-admin users by dynamically updating the search index when the model list changes, fixing a race condition that caused empty search results. #17996, #17960
  • ⚡ Task model function calling performance is improved by excluding base64 image data from payloads, significantly reducing token count and memory usage when images are present in conversations. #17897
  • 🤖 Text selection “Ask” action now correctly recognizes and uses local models configured via direct connections instead of only showing external provider models. #17896
  • 🛑 Task cancellation API now returns accurate response status, correctly reporting successful cancellations instead of incorrectly indicating failures. #17920
  • 💬 Follow-up query suggestions are now generated and displayed in temporary chats, matching the behavior of saved chats. #14987
  • 🔊 Azure Text-to-Speech now properly escapes special characters like ampersands in SSML, preventing HTTP 400 errors and ensuring audio generation succeeds for all text content. #17962
  • 🛠️ OpenAPI tool server calls with optional parameters now execute successfully even when no arguments are provided, removing the incorrect requirement for a request body. #18036
  • 🛠️ MCP mode tool server connections no longer incorrectly validate the OpenAPI path field, allowing seamless switching between OpenAPI and MCP connection types. #17989, #17988
  • 🛠️ Third-party tool responses containing non-UTF8 or invalid byte sequences are now handled gracefully without causing request failures. #17882
  • 🎨 Workspace filter dropdown now correctly renders model tags as strings instead of displaying individual characters, fixing broken filtering interface when models have multiple tags. #18034
  • ⌨️ Ctrl+Enter keyboard shortcut now correctly sends messages in mobile and narrow browser views on Chrome instead of inserting newlines. #17975
  • ⌨️ Tab characters are now preserved when pasting code or formatted text into the chat input box in plain text mode. #17958
  • 📋 Text selection copying from the chat input box now correctly copies only the selected text instead of the entire textbox content. #17911
  • 🔍 Web search query logging now uses debug level instead of info level, preventing user search queries from appearing in production logs. #17888
  • 📝 Debug print statements in middleware were removed to prevent excessive log pollution and respect configured logging levels. #17943

Changed

  • 🗄️ Milvus vector database dependency is updated from pymilvus 2.5.0 to 2.6.2, ensuring compatibility with newer Milvus versions but requiring users on older Milvus instances to either upgrade their database or manually downgrade the pymilvus package. #18066

下载链接