发布日期: 2026-02-12
版本号: v0.8.0

Open WebUI 发布了 0.8.0 版本，这是一个包含大量新功能、优化和重要修复的重大更新。主要新增了分析仪表盘、对技能标准和 Open Responses 协议的实验性支持，并重新设计了访问控制界面，支持将资源单独共享给用户。聊天功能增强了，包括消息队列、任务侧边栏指示器和提示词版本控制。性能方面进行了全面优化，提升了认证、聊天补全、数据库查询等多方面的速度和效率。安全性和稳定性也得到加强，修复了数据库连接池耗尽、权限控制、各类API错误处理等多个关键问题。

⚠️ 升级重要警告：本次更新包含数据库结构变更。强烈建议在升级前备份所有数据。在多服务器或负载均衡环境中，必须同时更新所有实例，不支持滚动更新，否则将导致应用故障。同时，聊天消息表的迁移可能耗时较长，需规划充足的停机时间。

更新内容 （中文）

[0.8.0] - 2026-02-12

[!CAUTION] ⚠️ 数据库迁移：此版本包含数据库架构变更；强烈建议在生产环境升级前备份数据库及所有相关数据。如果您运行的是多工作进程、多服务器或负载均衡部署，则必须同时更新所有实例，滚动更新不受支持，且会因架构不兼容导致应用程序故障。

[!WARNING] ⚠️ 长时间运行的数据库迁移：此版本包含一个针对 chat_message 表的迁移，该迁移可能需要较长时间才能完成，尤其是在具有大量聊天记录的较大规模部署中。请规划足够的停机时间，并让迁移完全完成，期间不要中断。

新增

  • 📊 分析仪表板。管理员现在可以访问分析仪表板，查看模型使用统计、按模型和用户统计的令牌消耗量、用户活跃度排名，以及按小时或天粒度的时间序列图表；点击任何模型可打开详细视图，包含反馈历史、关联标签和聊天浏览器，结果可按用户群组过滤。#21106、提交、提交
  • 🎯 实验性支持技能。Open WebUI 现在支持技能标准——允许用户创建和管理具有详细说明的可重用 AI 技能，在聊天中使用 “$” 命令引用它们，或将其附加到特定模型以在对话中自动提供上下文。#21312
  • 🧪 实验性支持 Open Responses 协议。连接现在可以配置为使用实验性的 Open Responses 协议替代聊天补全（Chat Completions），为兼容的提供商提供对扩展思维、流式推理令牌和更丰富工具调用处理的原生支持。提交、提交、提交、提交、提交、提交、提交、提交、提交
  • 👥 重新设计的访问控制 UI。访问控制 UI 重新设计，界面更直观，便于一次添加多个群组。#21277
  • 👤 按用户资源共享。现在可以直接与现有的按群组共享功能一起，将资源（包括知识库、提示词、模型、工具、频道和基础模型）共享给单个用户。#21277
  • 📨 消息排队。在生成响应期间，消息现在可以排队而非被阻塞，允许您继续您的思路；排队的消息会在生成完成时自动合并并发送，可以从输入区域编辑、删除或立即发送。提交、提交
  • 💡 活跃任务侧边栏指示器。用户现在可以直接在侧边栏中看到哪些聊天有活跃任务正在运行。提交
  • 📝 提示词版本控制。提示词现在包含版本控制功能，可进行完整历史跟踪，允许用户使用消息提交更改、查看历史版本、比较版本之间的差异，并在需要时回滚到以前的版本。#20945
  • 🏷️ 提示词标签。提示词现在可以使用标签进行组织，用户可以通过标签过滤提示词工作区，在大量集合中快速找到相关的提示词。#20945
  • 🐍 原生函数调用代码执行。代码执行现在可以在原生函数调用模式下工作，允许模型自主运行 Python 代码进行计算、数据分析和可视化，无需默认模式。#20592、文档:#998
  • 🚀 异步网页搜索。网页搜索操作现在在后台异步运行，允许用户在搜索完成时继续与应用程序交互。#20630
  • ⚡ 搜索防抖。应用程序中的搜索操作现在使用防抖功能进行更高效的响应，在输入时减少不必要的服务器请求，提高了搜索用户、群组、函数、工具、提示词、知识库、笔记，以及在聊天中使用知识和提示词命令时的响应速度。#20982、提交、提交、提交
  • 🤝 共享聊天管理。用户现在可以从设置中查看和管理他们所有的共享聊天，并可选择复制共享链接或取消共享他们不再想公开的对话。提交
  • 📁 用户文件管理。用户现在可以从设置中查看、搜索和删除他们上传的所有文件，在一个地方提供集中的文件管理。提交、#21047
  • 🗑️ Shift-点击快速删除。文件管理器中的文件现在可以通过按住 Shift 并点击删除按钮快速删除，绕过确认对话框进行更快的批量清理。#21044
  • ⌨️ 模型选择器快捷方式。现在可以使用 Ctrl+Shift+M 键盘快捷键打开模型选择器。#21130
  • 🧠 更智能的知识库与网页搜索。模型现在能更智能地在知识库搜索和网页搜索之间进行选择，而不是总是首先尝试知识库。#21115
  • 🌍 社区模型评论。用户现在可以直接从模型选择器菜单访问模型的社区评论，并在评分响应后被提示留下评论，管理员可以通过“社区共享”设置禁用此功能。提交
  • 📄 提示词工作区分页。提示词工作区现在为大型提示词集合包含分页功能，一次加载30个提示词，并具有搜索、过滤和排序功能，以提高性能和导航性。提交
  • 🎨 动作函数 HTML 渲染。动作函数现在可以直接在聊天中以嵌入式 iframe 的形式渲染丰富的 HTML 内容，匹配工具已有的功能，消除了动作作者注入代码块的需要。#21294、提交
  • 🔒 密码掩码字段。工具和函数开发人员现在可以将敏感字段标记为密码，这些字段在设置 UI 中会自动掩码，以防止窥屏和意外泄露。#20852、提交
  • 📋 提示词快速复制。工作区中的提示词现在包含一个快速复制按钮，可轻松将提示词内容复制到剪贴板。提交
  • 🔔 可关闭的通知提示。新消息和其他事件的通知提示现在包含一个悬停时出现的关闭按钮，允许用户立即关闭它们，而无需等待自动消失。#21056、提交
  • 🔔 临时聊天通知隐私。临时聊天的通知现在仅出现在运行该聊天的设备上，防止跨登录会话的隐私泄露。#21292
  • 💡 空聊天标题回退。没有聊天标题的通知现在显示“新聊天”，而不是显示 null。#21292
  • 🖼️ 并发图像编辑。涉及多张图像的图像编辑操作现在通过并发加载所有图像而非顺序加载来更快完成。#20911
  • 📧 USER_EMAIL 模板变量。用户现在可以在提示词和系统消息中使用“{{USER_EMAIL}}”模板变量引用他们的电子邮件地址。#20881
  • 🔤 字母顺序工具排序。聊天控制侧边栏中的工具和函数现在按字母顺序排列，在使用多个集成时更容易找到特定工具。#20871
  • 👁️ 模型列表状态过滤。管理员现在可以按状态（启用、禁用、可见、隐藏）过滤模型列表，并一次性批量启用或禁用所有过滤后的模型。#20553、#20774、提交
  • ⚙️ 按模型内置工具切换。管理员现在可以为每个模型启用或禁用单个内置工具，包括时间实用程序、内存、聊天历史、笔记、知识库和频道。#20641、提交
  • 📑 PDF 加载模式。管理员现在可以在“页面”和“单个”PDF 加载模式之间进行选择，允许将文档作为单个页面或完整文档进行处理，以便在页面边界处更好地分块。提交
  • 📑 模型设置分页。管理员现在可以在模型设置中更高效地浏览大型模型列表，分页显示每页30个模型，实现更流畅的导航。提交
  • 📌 固定只读模型。用户现在可以从工作区固定只读模型。#21308、提交
  • 🔍 Yandex 搜索提供商。管理员现在可以将 Yandex 配置为网页搜索提供商，为检索增强生成扩展搜索引擎选项。#20922
  • 🔐 自定义密码提示。管理员现在可以通过“PASSWORD_VALIDATION_HINT”环境变量向用户提供自定义密码要求提示，使注册或更改密码时必须满足的密码标准更加清晰。#20647、#20650
  • 🔑 OAuth 令牌交换。管理员现在可以通过“ENABLE_OAUTH_TOKEN_EXCHANGE”启用 OAuth 令牌交换，允许外部应用程序通过交换 OAuth 提供商令牌为 Open WebUI 会话令牌来对用户进行身份验证。提交
  • 🗄️ Weaviate 自定义端点。管理员现在可以通过新的环境变量连接到具有单独 HTTP 和 gRPC 端点的自托管 Weaviate 部署。#20620
  • 🛡️ MCP 自定义 SSL 证书。管理员现在可以通过“AIOHTTP_CLIENT_SESSION_TOOL_SERVER_SSL”环境变量使用自签名或自定义 SSL 证书连接到 MCP 服务器。#20875、提交
  • 🗃️ Redis Sentinel 重连延迟。使用 Redis Sentinel 的管理员现在可以通过“REDIS_RECONNECT_DELAY”配置重连延迟，以防止在故障转移选举期间重试耗尽。#21021
  • 📡 自定义用户信息头。管理员现在可以自定义将用户信息转发给外部服务时使用的头信息名称，从而与需要特定头前缀的服务（如 AWS Bedrock AgentCore）兼容。提交
  • 🔗 将用户信息转发给工具服务器。当启用“ENABLE_FORWARD_USER_INFO_HEADERS”时，用户身份和聊天上下文现在可以转发给 MCP 服务器和外部工具服务器，允许工具提供商实现按用户授权、审计和速率限制。#21092、提交
  • 📬 外部工具事件发射器。外部工具（OpenAPI/MCP）现在可以使用事件发射器端点将工具事件发回 Open WebUI，因为当启用“ENABLE_FORWARD_USER_INFO_HEADERS”时，消息 ID 现在会与聊天 ID 一起转发。#21214
  • 📥 Playground 聊天导出。管理员现在可以将 playground 聊天导出为 JSON 或纯文本文件，允许他们保存对话以进行备份或在平台外共享。提交
  • 🖼️ 图像 Playground。管理员现在可以在新的图像 Playground 中直接测试图像生成和编辑，支持上传源图像进行编辑和下载结果。提交
  • 🛠️ 动态下拉字段。工具和函数开发人员现在可以创建具有动态生成选项的下拉字段，这些选项会根据运行时上下文（如可用模型或用户权限）进行更新。提交
  • 🏎️ 更快的配置文件更新。通过消除冗余的数据库查询，用户配置文件更新和角色更改现在更快。#21011
  • 🔑 更快的身份验证。通过将数据库查询合并为单个查询，用户身份验证现在快了34%。#21010
  • 🔋 更快的聊天补全。通过在获取模型列表之前检查模型缓存，聊天补全和嵌入现在响应更快，将大多数请求的首令牌时间（Time To First Token）从几秒减少到亚秒级。#20886、#20069
  • 🏎️ 更快的 Redis 模型列表加载。在使用 Redis 和多个模型时，模型列表加载现在显著加快，通过本地缓存配置值而非每次模型迭代时进行重复的 Redis 查询，减少了 API 响应延迟。#21306
  • 💨 更快的知识库文件批量添加。批量向知识库添加文件现在更快，通过单个数据库查询而非每个文件一个查询。#21006
  • ⚡ 更流畅的模型选择器下拉菜单。模型选择器下拉菜单现在即使在有数百个模型的情况下也能流畅渲染，消除了在大型模型列表时打开下拉菜单出现的滞后和冻结。提交
  • 🚗 更快的模型可见性切换。通过优化数据库访问，在管理面板中切换模型可见性现在更快。#21009
  • 💾 更快的模型访问控制检查。通过预先批量获取模型信息和群组成员关系，而非为每个模型单独查询，模型访问控制检查现在更快。#21008
  • ⚙️ 更快的模型列表和导入。通过消除冗余的数据库查询，模型列表加载和模型导入现在更快。#21004
  • 🏃 更快的 SCIM 群组成员查找。通过批量处理用户查询而非单独获取每个成员，SCIM 群组成员查找现在快了多达13倍。#21005
  • 💨 批量群组成员计数。加载群组列表时，群组成员计数现在通过单个批量查询获取，消除了冗余的数据库查找。提交
  • 💨 更快的批量操作。通过优化的批量操作，批量反馈删除和群组成员移除现在快了4-5倍。#21019
  • 🧠 更快的内存更新。通过消除冗余的数据库查询，内存更新现在快了多达39%。#21013
  • ⚙️ 更快的过滤函数加载。通过批量处理数据库查询而非单独获取每个函数，过滤函数加载现在更快。#21018
  • 🖼️ 图像模型正则配置。管理员现在可以通过新的正则表达式环境变量配置哪些图像生成模型支持自动调整大小和 URL 响应，提高了与使用带前缀模型名称的 LiteLLM 和其他代理的兼容性。#21126、提交
  • 🎁 彩蛋功能切换。管理员现在可以通过“ENABLE_EASTER_EGGS”环境变量控制彩蛋功能的可见性。提交
  • 🔌 独立的访问控制更新。API 端点现在支持对模型、工具、知识库和笔记进行独立的访问控制更新，实现更细粒度的权限管理。提交、提交、提交
  • ♿ 屏幕阅读器可访问性。屏幕阅读器用户现在在登录页面只会听到一次密码字段标签，改善了辅助技术用户的表单导航。提交
  • 🔄 总体改进。在应用程序中实施了各种改进，以增强性能、稳定性和安全性。
  • 🌐 翻译更新。增强了加泰罗尼亚语、芬兰语、爱尔兰语、法语、德语、日语、拉脱维亚语、波兰语、葡萄牙语（巴西）、简体中文、斯洛伐克语、西班牙语和繁体中文的翻译。

修复

  • ⚡ 连接池耗尽修复。通过在聊天补全请求和用于内存与知识库处理的嵌入操作之前释放连接，解决了并发使用期间的数据库连接池耗尽和超时错误。#20569、#20570、#20571、#20572、#20573、#20574、#20575、#20576、#20577、#20578、#20579、#20580、#20581、提交、提交
  • 🚫 LDAP 认证挂起修复。LDAP 认证在登录不存在的账户时不再会冻结整个服务，防止应用程序挂起。提交、#21300
  • ✅ 受信任头自动注册修复。受信任头认证现在在首次登录后正确自动注册新用户，分配配置的默认角色，而不是对尚未在数据库中的用户失败。提交
  • 🛡️ 图像加载的 SSRF 保护。外部图像加载现在会在获取之前验证 URL，以防止针对本地和私有网络地址的 SSRF 攻击。提交
  • 🛡️ 格式错误的 Authorization 头修复。格式错误的 Authorization 头不再导致服务器崩溃；请求现在被优雅地处理，而不是返回 HTTP 500 错误。#20938、提交
  • 🚪 频道通知访问控制。没有频道权限的用户无法再通过通知访问频道，正确执行所有频道入口点的访问控制。#20883、#20789
  • 🐛 Ollama 模型名称后缀修复。不使用“:latest”作为模型名称的 Ollama 兼容提供商现在可以成功聊天，修复了模型名称被错误地附加“:latest”后缀的错误。#21331、提交
  • ♻️ 流式连接清理。流式响应现在在中断时正确清理网络连接，防止“未关闭客户端会话”错误随时间累积。#20889、#17058
  • 💾 内联图像上下文耗尽修复。内联图像现在不再通过将完整的 base64 数据包含在聊天元数据中来耗尽模型的上下文窗口，防止了图像密集型对话中上下文的过早耗尽。#20916
  • 🚀 状态指示器 GPU 使用修复。由用户在线状态指示器动画引起的高 GPU 使用率问题已解决，在具有硬件加速的浏览器中将消耗从35-40%降低到接近零。#21062、提交
  • 🔧 异步管道操作。管道操作现在异步运行，而不是阻塞 FastAPI 事件循环，允许服务器在等待外部管道 API 调用时处理其他请求。#20910
  • 🔌 MCP 工具回归修复。MCP 工具在 v0.7.2 的回归后再次可靠工作，该回归在尝试在聊天中使用 MCP 服务器时导致“无法序列化 ‘_asyncio.Future’ 对象”错误。#20629、#20500、提交
  • 🔗 函数聊天 ID 传播修复。函数现在在内部任务调用（如网页搜索查询生成、RAG 查询生成和图像提示生成）期间可靠地接收聊天标识符，使有状态函数能够维护一致的单个聊天状态，避免碎片化。#20563、#20585
  • 💻 Markdown 围栏代码执行修复。当模型将代码包装在 Markdown 围栏中时，代码执行现在可靠地工作，在执行前自动去除反引号，防止影响大多数非 GPT 模型的语法错误。#20941、提交、提交
  • 💻 ANSI 代码执行修复。当大型语言模型在输出中包含 ANSI 终端颜色代码时，代码执行现在可靠，防止了以前导致语法错误的随机故障。#21091、提交
  • 🗨️ 不完整模型元数据崩溃修复。使用元数据信息不完整的模型开始聊天不再导致应用程序崩溃。#20565、提交
  • 💬 不可用模型崩溃修复。使用 Ctrl+Shift+Enter 添加消息对时，如果聊天模型不可用，不再崩溃，而是显示有用的错误消息。#20663
  • 📚 知识库文件上传修复。现在可以正确上传文件到知识库，修复了阻止文件上传的数据库映射错误。#20925、#20931
  • 🧠 知识库查询类型修复。当模型将工具调用参数作为字符串而非预期类型发送时，知识库查询不再间歇性失败。#20705
  • 📚 知识库重新索引修复。重新索引知识库文件现在可以正确工作，而不是因重复内容错误而失败。#20854、#20857
  • 🔧 多工作进程知识库超时修复。在多工作进程部署中，上传超大文档到知识库不再会导致工作进程因健康检查超时而被终止，管理员现在可以通过“RAG_EMBEDDING_TIMEOUT”配置自定义嵌入超时。#21158、讨论、提交
  • 🌅 暗模式图标反转修复。聊天和操作菜单中的图标现在在暗模式下正确显示，修复了 base64 编码中包含“svg”的 PNG 图标被随机反转的问题。#21272、提交
  • 🛠️ 管理员模型写入权限修复。修复了管理面板允许为模型分配写入权限的问题，因为具有写入权限的用户不是管理员且无法写入。提交
  • 🛠️ 提示词访问控制保存修复。修改资源权限时，提示词访问控制设置现在可以正确保存。提交
  • ✏️ 知识库文件编辑修复。编辑知识库中的文件现在可以正确保存并可用于检索，修复了保存看似成功但文件无法被搜索的静默失败问题。提交
  • 🖼️ 推理部分伪像渲染修复。模型推理部分中的代码块不再错误地渲染为交互式伪像，确保只有预期的输出显示为预览。#20801、#20877、提交
  • 🔐 群组资源共享修复。与群组共享资源现在可以正确工作，修复了数据库错误以及以只读访问权限共享的模型对群组成员不可见的问题。#20666、#21043、提交
  • 🔑 Docling API 密钥修复。Docling API 密钥身份验证现在可以正确工作，通过使用正确的“X-Api-Key”头格式而非错误的“Bearer”授权前缀。#20652
  • 🔌 MCP OAuth 2.1 修复。MCP OAuth 2.1 身份验证现在可以正确工作，解决了授权回调期间的连接验证失败和 401 错误。#20808、#20828、提交
  • 💻 MATLAB 语法高亮。MATLAB 代码块现在在聊天消息中显示适当的语法高亮。#20719、#20773
  • 📊 CSV 导出 HTML 实体解码。将表格导出为 CSV 时现在可以正确解码 HTML 实体，确保特殊字符在导出的文件中正确显示。#20688
  • 📄 Markdown Header Text Splitter 持久性。“Markdown Header Text Splitter”文档设置现在在禁用时可以正确持久化，防止在页面刷新后恢复为启用状态。#20929、#20930
  • 🔌 音频服务超时处理。音频转录和文本转语音请求现在具有适当的超时，防止当外部服务无响应时 UI 冻结。#21055
  • 💬 参考聊天可见性修复。“参考聊天”选项现在即使在侧边栏折叠时也出现在消息输入菜单中，修复了在移动设备和首次加载时隐藏的问题。#20827、提交
  • 🔍 聊天搜索自排除。“search_chats”内置工具现在从搜索结果中排除当前对话，防止冗余匹配。#20718、提交
  • 📚 知识库分页修复。在知识库文件中分页不再显示重复项或在多个文档共享相同更新时间戳时跳过文件。#20846、提交
  • 📋 批量文件错误报告。批量文件处理操作现在在发生故障时返回结构正确的错误信息，使多文件操作中哪里出错更清晰。#20795、提交
  • ⚙️ 使用 Redis 的持久配置修复。当使用 Redis 时，配置值现在遵循“ENABLE_PERSISTENT_CONFIG”设置，确保在禁用持久配置时环境变量在重新启动时重新加载。#20830、提交
  • 🔧 Engine.IO 日志记录修复。“WEBSOCKET_SERVER_ENGINEIO_LOGGING”环境变量现在可以正确工作，允许管理员独立于常规 Websocket 日志配置 Engine.IO 日志记录。#20727、提交
  • 🌐 法语默认修复。请求法语的浏览器现在默认为法语（法国）而非法语（加拿大），符合标准语言偏好期望。#20603、提交
  • 🔘 Firefox 删除按钮修复。在 Firefox 中点击删除按钮后按 Enter 键不再错误地重新触发确认模态框。提交
  • 🌍 RTL 表格渲染修复。聊天 Markdown 表格现在在包含从右到左（RTL）语言内容（阿拉伯语、希伯来语、波斯语等）时正确显示从右到左，匹配“自动”方向设置行为。#21160、提交
  • 🔒 工具的写入权限执行。没有写入权限的用户现在被正确阻止编辑工具，在尝试未经授权编辑时显示清晰的错误消息。提交、提交
  • 🛡️ 聊天阀门权限执行。“允许聊天阀门”权限现在在集成菜单中正确执行，防止用户绕过访问限制。#20691
  • 📝 审计日志浏览器会话修复。审计日志现在正确捕获所有用户活动，包括基于浏览器的会话，而不仅仅是 API 密钥请求。#20651、提交、提交
  • 🎨 长模型名称截断。管理面板中的长模型名称和 ID 现在正确截断以防止视觉溢出，悬停时显示完整名称。#20696
  • 👥 管理员用户过滤分页修复。在管理面板中过滤用户现在自动重置到第1页，防止从第1页之后的页面搜索时出现空结果。#20723、提交
  • 🔎 工作区页面上的用户名搜索。现在可以在模型、知识和函数工作区页面上按用户名正确搜索用户，更容易找到特定用户拥有的资源。#20780
  • 🗑️ 文件删除孤立嵌入修复。删除文件现在正确移除关联的知识库嵌入，防止孤立数据累积。提交
  • 🧹 事件监听器内存泄漏修复。导航期间未清理事件监听器导致的内存泄漏已解决。#20913
  • 🐳 Docker Ollama 更新修复。添加缺失的 zstd 依赖后，Ollama 现在可以在 Docker 容器内更新。#20994、#21052
  • 📝 工作区重复 API 请求修复。提示词、知识和模型工作区在加载时不再发出重复的 API 请求。提交、提交
  • 📡 OpenTelemetry Redis 集群修复。OpenTelemetry 插桩现在可以正确与 Redis 集群模式部署一起工作。#21129
  • 🐳 气隙 NLTK 分词器修复。通过在 Docker 镜像中打包 NLTK 分词器数据，文档提取现在在容器重启后的气隙环境中可靠工作。#21165、#21150
  • 💬 频道模型提及崩溃修复。在频道中提及模型不再在较旧的线程消息数据缺失时崩溃。#21112
  • 🔧 OpenAPI 工具导入修复。导入 OpenAPI 工具规范不再在参数缺少显式名称字段时崩溃，修复了与复杂请求体定义的兼容性。#21121、提交
  • 🌐 网页附件内容修复。将网页附加到聊天现在检索完整内容而不仅仅是元数据，修复了 SSL 证书验证中未等待的协程。#21166、提交
  • 💾 文件上传设置持久性。文件上传设置（最大上传大小、最大文件数、图像压缩尺寸）现在正确持久化，在更新其他 RAG 配置设置时不再被擦除。#21057、提交
  • 📦 工具调用展开/折叠修复。最近的重构禁用此行为后，聊天中的工具调用结果现在可以再次展开和折叠。#21205
  • 🪛 禁用 API 端点绕过修复。修复了当提供“url_idx”参数时，Ollama/OpenAI API 端点绕过“ENABLE_OLLAMA_API”和“ENABLE_OPENAI_API”标志的问题。端点现在正确返回 503 错误并带有清晰的“API 已禁用”消息，而不是尝试连接并记录令人困惑的连接错误。
  • 🛠️ OpenSearch 3.0 兼容性修复。当使用 opensearch-py >= 3.0.0 的 OpenSearch 后端时，向知识库上传文档现在可以正确工作，修复了以前导致失败的 TypeError。#21248、#20649
  • 📱 Gboard 多行粘贴修复。从 Android 上的 Gboard 粘贴的多行文本现在正确插入，而不是被替换为单个换行符，修复了键盘剪贴板建议条通过“insertText”事件而非标准粘贴事件发送文本的错误。#21265
  • 🔧 批量嵌入端点修复。当处理对 Ollama 提供商的批量请求时，“/api/embeddings”端点现在正确为每个输入字符串返回单独的嵌入。提交、#21279
  • 🗝️ 嵌入的 SSL 验证。SSL 证书验证现在遵循“AIOHTTP_CLIENT_SESSION_SSL”设置进行 OpenAI 和 Azure OpenAI 嵌入请求，允许在禁用时连接到自签名证书端点。提交
  • 🔧 工具调用 HTML 实体修复。模型现在在多轮对话中接收格式正确的工具调用结果，修复了 HTML 实体导致内容格式错误且难以解析的问题。#20755
  • 💾 重复内联图像上下文修复。内联图像现在不再通过将完整的 base64 数据包含在聊天元数据中来耗尽模型的上下文窗口，防止了图像密集型对话中上下文的过早耗尽。#20916
  • 🐛 OpenAI 模型缓存查找修复。OpenAI API 路由器模型查找已更正为在检查缓存时使用正确的模型标识符，确保聊天补全期间模型检索一致且正确。#21327
  • 🐛 Ollama latest 后缀修复。不使用“:latest”作为模型名称的 Ollama 兼容提供商现在可以成功聊天，修复了模型名称被错误地附加“:latest”后缀的错误。#21331、提交
  • ⛔ OpenAI 端点检测修复。OpenAI API 端点检测已更正为使用精确主机名匹配而非子字符串匹配，防止具有类似 URL 模式的第三方提供商被错误过滤。提交
  • 🛠️ RedisCluster 任务停止修复。任务停止现在在 RedisCluster 部署中正常工作，修复了任务在取消尝试后仍保持活动状态的问题。#20803、提交
  • 📎 引用解析错误修复。当内置工具返回错误响应时，引用解析不再崩溃，修复了 search_web 等工具失败时的 AttributeError 问题。#21071

变更

  • ‼️ 需要数据库迁移 — 此版本包含数据库架构变更；多工作进程、多服务器或负载均衡部署必须同时更新所有实例，而不是执行滚动更新，因为运行混合版本将因新旧实例之间的架构不兼容而导致应用程序故障。
  • ⚠️ 聊天消息表迁移 — 此版本包含新的聊天消息表迁移，在具有大量聊天记录的较大规模部署中可能需要较长时间才能完成。管理员应规划足够的维护窗口，并让迁移在完全不中断的情况下完成。在时间或资源不足的情况下运行迁移可能导致数据完整性问题。
  • 🔗 基于提示词 ID 的 URL。提示词现在使用基于唯一 ID 的 URL 替代基于命令的 URL，允许更灵活的命令重命名，而不会破坏已保存的链接或集成。#20945

[!TIP]
寻找 企业计划? – 立即与我们的销售团队联系！

获取增强功能，包括自定义主题和品牌、服务水平协议（SLA）支持等！

👏 非常感谢我们出色的贡献者们，是他们的辛勤工作和奉献使此版本成为可能： @aindriu80, @Edgaras0x4E, @tomasloksa, @ctolon, @ShirasawaSama, @sjhddh, @PeterDaveHello, @schneider-f, @joaoback, @silentoplayz, @tremlin, @aleixdorca, @veeceey, @Algorithm5838, @jannikstdl, @Th3Un1q3, @EntropyYue, @Classic298, @pelmennoteam, @mickeytheseal, @hsienz

赞助商 🙌

🚀 我们要向我们了不起的赞助商们表示衷心的感谢，感谢他们的慷慨支持（注意：我们已将私人赞助商从此列表中排除。如果您想在此处展示，请随时联系我们！）

@amd, @roosi-gmbh, @rndmcnlly, @feddersen-group, @SRKConsulting, @mushmax, @icanbwell, @Classic298, @loitragg, @haka4githu6, @silenceroom, @walsermedia, @crizCraig, @GenieDev101, @CubicleGangster, @TOWaB, @comet-ml, @catalyst-net-nz, @JrdnHans, @hkosm, @joaoback, @LuxTronic

更新内容 （原始）

[0.8.0] - 2026-02-12

[!CAUTION] ⚠️ Database Migrations: This release includes database schema changes; we strongly recommend backing up your database and all associated data before upgrading in production environments. If you are running a multi-worker, multi-server, or load-balanced deployment, all instances must be updated simultaneously, rolling updates are not supported and will cause application failures due to schema incompatibility.

[!WARNING] ⚠️ Long-Running Database Migration: This release includes a migration for the chat_message table that can take a significant amount of time to complete, especially in larger deployments with extensive chat histories. Please plan for adequate downtime and allow the migration to finish completely without interruption.

Added

  • 📊 Analytics dashboard. Administrators now have access to an Analytics dashboard showing model usage statistics, token consumption by model and user, user activity rankings, and time-series charts with hourly or daily granularity; clicking any model opens a detail view with feedback history, associated tags, and chat browser, and results can be filtered by user group. #21106, Commit, Commit
  • 🎯 Experimental support for Skills. Open WebUI now supports the Skill standard — allowing users to create and manage reusable AI skills with detailed instructions, reference them in chats using the “$” command, or attach them to specific models for automatic context in conversations. #21312
  • 🧪 Experimental support for Open Responses protocol. Connections can now be configured to use the experimental Open Responses protocol instead of Chat Completions, enabling native support for extended thinking, streaming reasoning tokens, and richer tool call handling for compatible providers. Commit, Commit, Commit, Commit, Commit, Commit, Commit, Commit, Commit
  • 👥 Redesigned access control UI. The access control UI was redesigned with a more intuitive interface that makes it easier to add multiple groups at once. #21277
  • 👤 Per-user resource sharing. Resources including knowledge bases, prompts, models, tools, channels, and base models can now be shared directly to individual users alongside the existing per-group sharing capability. #21277
  • 📨 Message queuing. Messages can now be queued while a response is generating rather than being blocked, allowing you to continue your train of thought; queued messages are automatically combined and sent when generation completes, and can be edited, deleted, or sent immediately from the input area. Commit, Commit
  • 💡 Active task sidebar indicator. Users can now see which chats have active tasks running directly in the sidebar. Commit
  • 📝 Prompt version control. Prompts now include version control with full history tracking, allowing users to commit changes with messages, view past versions, compare differences between versions, and roll back to previous versions when needed. #20945
  • 🏷️ Prompt tags. Prompts can now be organized with tags, and users can filter the prompt workspace by tag to quickly find related prompts across large collections. #20945
  • 🐍 Native function calling code execution. Code execution now works with Native function calling mode, allowing models to autonomously run Python code for calculations, data analysis, and visualizations without requiring Default mode. #20592, Docs:#998
  • 🚀 Async web search. Web search operations now run asynchronously in the background, allowing users to continue interacting with the application while searches complete. #20630
  • ⚡ Search debouncing. Search operations across the application now respond more efficiently with debouncing that reduces unnecessary server requests while typing, improving responsiveness when searching users, groups, functions, tools, prompts, knowledge bases, notes, and when using the knowledge and prompts commands in chat. #20982, Commit, Commit, Commit
  • 🤝 Shared chats management. Users can now view and manage all their shared chats from Settings, with options to copy share links or unshare conversations they no longer want public. Commit
  • 📁 User file management. Users can now view, search, and delete all their uploaded files from Settings, providing centralized file management in one place. Commit, #21047
  • 🗑️ Shift-click quick delete. Files in the File Manager can now be quickly deleted by holding Shift and clicking the delete button, bypassing the confirmation dialog for faster bulk cleanup. #21044
  • ⌨️ Model selector shortcut. The model selector can now be opened with Ctrl+Shift+M keyboard shortcut. #21130
  • 🧠 Smarter knowledge vs web search. Models now choose more intelligently between knowledge base search and web search rather than always trying knowledge first. #21115
  • 🌍 Community model reviews. Users can now access community reviews for models directly from the model selector menu and are prompted to leave reviews after rating responses, with administrators able to disable this via the “Community Sharing” setting. Commit
  • 📄 Prompts workspace pagination. The prompts workspace now includes pagination for large prompt collections, loading 30 prompts at a time with search, filtering, and sorting capabilities for improved performance and navigation. Commit
  • 🎨 Action function HTML rendering. Action functions can now render rich HTML content directly in chat as embedded iframes, matching the capabilities that tools already had and eliminating the need for action authors to inject codeblocks. #21294, Commit
  • 🔒 Password-masked valve fields. Tool and function developers can now mark sensitive fields as passwords, which are automatically masked in the settings UI to prevent shoulder surfing and accidental exposure. #20852, Commit
  • 📋 Prompt quick copy. Prompts in the workspace now include a quick copy button for easily copying prompt content to the clipboard. Commit
  • 🔔 Dismissible notification toasts. Notification toasts for new messages and other events now include a close button that appears on hover, allowing users to dismiss them immediately instead of waiting for auto-dismissal. #21056, Commit
  • 🔔 Temporary chat notification privacy. Notifications from temporary chats now only appear on the device where the chat is running, preventing privacy leaks across logged-in sessions. #21292
  • 💡 Null chat title fallback. Notifications without chat titles now display “New Chat” instead of showing null. #21292
  • 🖼️ Concurrent image editing. Image editing operations with multiple images now complete faster by loading all images concurrently instead of sequentially. #20911
  • 📧 USER_EMAIL template variable. Users can now reference their email address in prompts and system messages using the “{{USER_EMAIL}}” template variable. #20881
  • 🔤 Alphabetical tool ordering. Tools and Functions in the Chat Controls sidebar now appear in alphabetical order, making it easier to locate specific tools when working with multiple integrations. #20871
  • 👁️ Model list status filtering. Administrators can now filter the model list by status (enabled, disabled, visible, hidden) and bulk enable or disable all filtered models at once. #20553, #20774, Commit
  • ⚙️ Per-model built-in tool toggles. Administrators can now enable or disable individual built-in tools for each model, including time utilities, memory, chat history, notes, knowledge base, and channels. #20641, Commit
  • 📑 PDF loading modes. Administrators can now choose between “page” and “single” PDF loading modes, allowing documents to be processed as individual pages or as complete documents for better chunking across page boundaries. Commit
  • 📑 Model Settings pagination. Administrators can now navigate large model lists more efficiently in Model Settings, with pagination displaying 30 models per page for smoother navigation. Commit
  • 📌 Pin read-only models. Users can now pin read-only models from the workspace. #21308, Commit
  • 🔍 Yandex search provider. Administrators can now configure Yandex as a web search provider, expanding search engine options for retrieval-augmented generation. #20922
  • 🔐 Custom password hints. Administrators can now provide custom password requirement hints to users via the “PASSWORD_VALIDATION_HINT” environment variable, making it clearer what password criteria must be met during signup or password changes. #20647, #20650
  • 🔑 OAuth token exchange. Administrators can now enable OAuth token exchange via “ENABLE_OAUTH_TOKEN_EXCHANGE”, allowing external applications to authenticate users by exchanging OAuth provider tokens for Open WebUI session tokens. Commit
  • 🗄️ Weaviate custom endpoints. Administrators can now connect to self-hosted Weaviate deployments with separate HTTP and gRPC endpoints via new environment variables. #20620
  • 🛡️ MCP custom SSL certificates. Administrators can now connect to MCP servers with self-signed or custom SSL certificates via the “AIOHTTP_CLIENT_SESSION_TOOL_SERVER_SSL” environment variable. #20875, Commit
  • 🗃️ Redis Sentinel reconnection delay. Administrators using Redis Sentinel can now configure a reconnection delay via “REDIS_RECONNECT_DELAY” to prevent retry exhaustion during failover elections. #21021
  • 📡 Custom user info headers. Administrators can now customize the header names used when forwarding user information to external services, enabling compatibility with services like AWS Bedrock AgentCore that require specific header prefixes. Commit
  • 🔗 Forward user info to tool servers. User identity and chat context can now be forwarded to MCP servers and external tool servers when “ENABLE_FORWARD_USER_INFO_HEADERS” is enabled, allowing tool providers to implement per-user authorization, auditing, and rate limiting. #21092, Commit
  • 📬 External tool event emitters. External tools (OpenAPI/MCP) can now send tool events back to Open WebUI using the event emitter endpoint, as message ID is now forwarded alongside chat ID when “ENABLE_FORWARD_USER_INFO_HEADERS” is enabled. #21214
  • 📥 Playground chat export. Administrators can now export playground chats as JSON or plain text files, allowing them to save their conversations for backup or sharing outside the platform. Commit
  • 🖼️ Images playground. Administrators can now test image generation and editing directly in a new Images playground, with support for uploading source images for edits and downloading results. Commit
  • 🛠️ Dynamic dropdown valve fields. Tool and function developers can now create dropdown fields with dynamically-generated options that update based on runtime context, such as available models or user permissions. Commit
  • 🏎️ Faster profile updates. User profile updates and role changes are now faster by eliminating redundant database queries. #21011
  • 🔑 Faster authentication. User authentication is now 34% faster by combining database lookups into a single query. #21010
  • 🔋 Faster chat completions. Chat completions and embeddings now respond much faster by checking the model cache before fetching model lists, reducing Time To First Token from several seconds to subsecond for most requests. #20886, #20069
  • 🏎️ Faster Redis model list loading. Model list loading is now significantly faster when using Redis with many models, reducing API response latency by caching configuration values locally instead of making repeated Redis lookups on every model iteration. #21306
  • 💨 Faster knowledge base file batch-add. Batch-adding files to knowledge bases is now faster with a single database query instead of one query per file. #21006
  • ⚡ Smoother model selector dropdown. The model selector dropdown now renders smoothly even with hundreds of models, eliminating the lag and freezing that occurred when opening the dropdown with large model lists. Commit
  • 🚗 Faster model visibility toggling. Toggling model visibility in the admin panel is now faster with optimized database access. #21009
  • 💾 Faster model access control checks. Model access control checks are now faster by batch-fetching model info and group memberships upfront instead of querying for each model. #21008
  • ⚙️ Faster model list and imports. Model list loading and model imports are now faster by eliminating redundant database queries. #21004
  • 🏃 Faster SCIM group member lookups. SCIM group member lookups are now up to 13x faster by batching user queries instead of fetching each member individually. #21005
  • 💨 Batched group member counts. Group member counts are now fetched in a single batch query when loading group lists, eliminating redundant database lookups. Commit
  • 💨 Faster bulk operations. Bulk feedback deletion and group member removal are now 4-5x faster with optimized batch operations. #21019
  • 🧠 Faster memory updates. Memory updates are now up to 39% faster by eliminating redundant database queries. #21013
  • ⚙️ Faster filter function loading. Filter function loading is now faster by batching database queries instead of fetching each function individually. #21018
  • 🖼️ Image model regex configuration. Administrators can now configure which image generation models support auto-sizing and URL responses via new regex environment variables, improving compatibility with LiteLLM and other proxies that use prefixed model names. #21126, Commit
  • 🎁 Easter eggs toggle. Administrators can now control the visibility of easter egg features via the “ENABLE_EASTER_EGGS” environment variable. Commit
  • 🔌 Independent access control updates. API endpoints now support independent access control updates for models, tools, knowledge bases, and notes, enabling finer-grained permission management. Commit, Commit, Commit
  • ♿ Screen reader accessibility. Screen reader users now hear the password field label only once on the login page, improving form navigation for assistive technology users. Commit
  • 🔄 General improvements. Various improvements were implemented across the application to enhance performance, stability, and security.
  • 🌐 Translation updates. Translations for Catalan, Finnish, Irish, French, German, Japanese, Latvian, Polish, Portuguese (Brazil), Simplified Chinese, Slovak, Spanish, and Traditional Chinese were enhanced and expanded.

Fixed

  • ⚡ Connection pool exhaustion fix. Database connection pool exhaustion and timeout errors during concurrent usage have been resolved by releasing connections before chat completion requests and embedding operations for memory and knowledge base processing. #20569, #20570, #20571, #20572, #20573, #20574, #20575, #20576, #20577, #20578, #20579, #20580, #20581, Commit, Commit
  • 🚫 LDAP authentication hang fix. LDAP authentication no longer freezes the entire service when logging in with non-existent accounts, preventing application hangs. Commit, #21300
  • ✅ Trusted Header auto-registration fix. Trusted Header Authentication now properly auto-registers new users after the first login, assigning the configured default role instead of failing for users not yet in the database. Commit
  • 🛡️ SSRF protection for image loading. External image loading now validates URLs before fetching to prevent SSRF attacks against local and private network addresses. Commit
  • 🛡️ Malformed Authorization header fix. Malformed Authorization headers no longer cause server crashes; requests are now handled gracefully instead of returning HTTP 500 errors. #20938, Commit
  • 🚪 Channel notification access control. Users without channel permissions can no longer access channels through notifications, properly enforcing access controls across all channel entry points. #20883, #20789
  • 🐛 Ollama model name suffix fix. Ollama-compatible providers that do not use “:latest” in model names can now successfully chat, fixing errors where model names were incorrectly appended with “:latest” suffixes. #21331, Commit
  • ♻️ Streaming connection cleanup. Streaming responses now properly clean up network connections when interrupted, preventing “Unclosed client session” errors from accumulating over time. #20889, #17058
  • 💾 Inline image context exhaustion fix. Inline images no longer exhaust the model’s context window by including their full base64 data in chat metadata, preventing premature context exhaustion with image-heavy conversations. #20916
  • 🚀 Status indicator GPU usage fix. High GPU usage caused by the user online status indicator animation has been resolved, reducing consumption from 35-40% to near-zero in browsers with hardware acceleration. #21062, Commit
  • 🔧 Async pipeline operations. Pipeline operations now run asynchronously instead of blocking the FastAPI event loop, allowing the server to handle other requests while waiting for external pipeline API calls. #20910
  • 🔌 MCP tools regression fix. MCP tools now work reliably again after a regression in v0.7.2 that caused “cannot pickle ‘_asyncio.Future’ object” errors when attempting to use MCP servers in chat. #20629, #20500, Commit
  • 🔗 Function chat ID propagation fix. Functions now reliably receive the chat identifier during internal task invocations like web search query generation, RAG query generation, and image prompt generation, enabling stateful functions to maintain consistent per-chat state without fragmentation. #20563, #20585
  • 💻 Markdown fence code execution fix. Code execution now works reliably when models wrap code in markdown fences, automatically stripping the backticks before execution to prevent syntax errors that affected most non-GPT models. #20941, Commit, Commit
  • 💻 ANSI code execution fix. Code execution is now reliable when LLMs include ANSI terminal color codes in their output, preventing random failures that previously caused syntax errors. #21091, Commit
  • 🗨️ Incomplete model metadata crash fix. Starting chats with models that have incomplete metadata information no longer crashes the application. #20565, Commit
  • 💬 Unavailable model crash fix. Adding message pairs with Ctrl+Shift+Enter no longer crashes when the chat’s model is unavailable, showing a helpful error message instead. #20663
  • 📚 Knowledge base file upload fix. Uploading files to knowledge bases now works correctly, fixing database mapping errors that prevented file uploads. #20925, #20931
  • 🧠 Knowledge base query type fix. Knowledge base queries no longer fail intermittently when models send tool call parameters as strings instead of their expected types. #20705
  • 📚 Knowledge base reindex fix. Reindexing knowledge base files now works correctly instead of failing with duplicate content errors. #20854, #20857
  • 🔧 Multi-worker knowledge base timeout fix. In multi-worker deployments, uploading very large documents to knowledge bases no longer causes workers to be killed by health check timeouts, and administrators can now configure a custom embedding timeout via “RAG_EMBEDDING_TIMEOUT”. #21158, Discussion, Commit
  • 🌅 Dark mode icon inversion fix. Icons in chat and action menus are now displayed correctly in dark mode, fixing an issue where PNG icons with “svg” in their base64 encoding were randomly inverted. #21272, Commit
  • 🛠️ Admin model write permission fix. Fixed the admin panel allowing models to be assigned write permissions, since users with write permission are not admins and cannot write. Commit
  • 🛠️ Prompt access control save fix. Prompt access control settings are now saved correctly when modifying resource permissions. Commit
  • ✏️ Knowledge base file edit fix. Editing files within knowledge bases now saves correctly and can be used for retrieval, fixing a silent failure where the save appeared successful but the file could not be searched. Commit
  • 🖼️ Reasoning section artifact rendering fix. Code blocks within model reasoning sections no longer incorrectly render as interactive artifacts, ensuring only intended output displays as previews. #20801, #20877, Commit
  • 🔐 Group resource sharing fix. Sharing resources with groups now works correctly, fixing database errors and an issue where models shared with read-only access were not visible to group members. #20666, #21043, Commit
  • 🔑 Docling API key fix. Docling API key authentication now works correctly by using the proper “X-Api-Key” header format instead of the incorrect “Bearer” authorization prefix. #20652
  • 🔌 MCP OAuth 2.1 fix. MCP OAuth 2.1 authentication now works correctly, resolving connection verification failures and 401 errors during the authorization callback. #20808, #20828, Commit
  • 💻 MATLAB syntax highlighting. MATLAB code blocks now display with proper syntax highlighting in chat messages. #20719, #20773
  • 📊 CSV export HTML entity decoding. Exporting tables to CSV now properly decodes HTML entities, ensuring special characters display correctly in the exported file. #20688
  • 📄 Markdown Header Text Splitter persistence. The “Markdown Header Text Splitter” document setting now persists correctly when disabled, preventing it from reverting to enabled after page refresh. #20929, #20930
  • 🔌 Audio service timeout handling. Audio transcription and text-to-speech requests now have proper timeouts, preventing the UI from freezing when external services don’t respond. #21055
  • 💬 Reference Chats visibility fix. The “Reference Chats” option now appears in the message input menu even when the sidebar is collapsed, fixing the issue where it was hidden on mobile devices and at first load. #20827, Commit
  • 🔍 Chat search self-exclusion. The “search_chats” builtin tool now excludes the current conversation from search results, preventing redundant matches. #20718, Commit
  • 📚 Knowledge base pagination fix. Paginating through knowledge base files no longer shows duplicates or skips files when multiple documents share the same update timestamp. #20846, Commit
  • 📋 Batch file error reporting. Batch file processing operations now return properly structured error information when failures occur, making it clearer what went wrong during multi-file operations. #20795, Commit
  • ⚙️ Persistent config with Redis fix. Configuration values now respect the “ENABLE_PERSISTENT_CONFIG” setting when Redis is used, ensuring environment variables are reloaded on restart when persistent config is disabled. #20830, Commit
  • 🔧 Engine.IO logging fix. The “WEBSOCKET_SERVER_ENGINEIO_LOGGING” environment variable now works correctly, allowing administrators to configure Engine.IO logging independently from general websocket logging. #20727, Commit
  • 🌐 French language default fix. Browsers requesting French language now default to French (France) instead of French (Canada), matching standard language preference expectations. #20603, Commit
  • 🔘 Firefox delete button fix. Pressing Enter after clicking delete buttons no longer incorrectly retriggers confirmation modals in Firefox. Commit
  • 🌍 RTL table rendering fix. Chat markdown tables now correctly display right-to-left when containing RTL language content (Arabic, Hebrew, Farsi, etc.), matching the “Auto” direction setting behavior. #21160, Commit
  • 🔒 Write permission enforcement for tools. Users without write permissions are now properly prevented from editing tools, with a clear error message displayed when attempting unauthorized edits. Commit, Commit
  • 🛡️ Chat Valves permission enforcement. The “Allow Chat Valves” permission is now properly enforced in the integrations menu, preventing users from bypassing access restrictions. #20691
  • 📝 Audit log browser session fix. Audit logs now properly capture all user activity including browser-based sessions, not just API key requests. #20651, Commit, Commit
  • 🎨 Long model name truncation. Long model names and IDs in the admin panel now truncate properly to prevent visual overflow, with full names visible on hover. #20696
  • 👥 Admin user filter pagination fix. Filtering users in the admin panel now automatically resets to page 1, preventing empty results when searching from pages beyond the first. #20723, Commit
  • 🔎 Username search on workspace pages. Searching for users by username now works correctly on Models, Knowledge, and Functions workspace pages, making it easier to find resources owned by specific users. #20780
  • 🗑️ File deletion orphaned embeddings fix. Deleting files now properly removes associated knowledge base embeddings, preventing orphaned data from accumulating. Commit
  • 🧹 Event listener memory leak fix. Memory leaks caused by event listeners not being cleaned up during navigation have been resolved. #20913
  • 🐳 Docker Ollama update fix. Ollama can now be updated within Docker containers after adding a missing zstd dependency. #20994, #21052
  • 📝 Workspace duplicate API request fix. The prompts, knowledge, and models workspaces no longer make duplicate API requests when loading. Commit, Commit
  • 📡 OpenTelemetry Redis cluster fix. OpenTelemetry instrumentation now works correctly with Redis cluster mode deployments. #21129
  • 🐳 Airgapped NLTK tokenizer fix. Document extraction now works reliably in airgapped environments after container restarts by bundling NLTK tokenizer data in the Docker image. #21165, #21150
  • 💬 Channel model mention crash fix. Mentioning a model in channels no longer crashes when older thread messages have missing data. #21112
  • 🔧 OpenAPI tool import fix. Importing OpenAPI tool specifications no longer crashes when parameters lack explicit name fields, fixing compatibility with complex request body definitions. #21121, Commit
  • 🌐 Webpage attachment content fix. Attaching webpages to chats now retrieves full content instead of only metadata, fixing an unawaited coroutine in SSL certificate verification. #21166, Commit
  • 💾 File upload settings persistence. File upload settings (Max Upload Size, Max File Count, Image Compression dimensions) now persist correctly and are no longer erased when updating other RAG configuration settings. #21057, Commit
  • 📦 Tool call expand/collapse fix. Tool call results in chat can now be expanded and collapsed again after a recent refactor disabled this behavior. #21205
  • 🪛 Disabled API endpoint bypass fix. Fixed Ollama/OpenAI API endpoints bypassing ‘ENABLE_OLLAMA_API’ and ‘ENABLE_OPENAI_API’ flags when the ‘url_idx’ parameter was provided. Endpoints now properly return a 503 error with a clear “API is disabled” message instead of attempting to connect and logging confusing connection errors.
  • 🛠️ OpenSearch 3.0 compatibility fix. Document uploads to knowledge bases now work correctly when using OpenSearch backend with opensearch-py >= 3.0.0, fixing a TypeError that previously caused failures. #21248, #20649
  • 📱 Gboard multi-line paste fix. Multi-line text pasted from Gboard on Android now inserts correctly instead of being replaced with a single newline, fixing a bug where the keyboard’s clipboard suggestion strip sent text via ‘insertText’ events instead of standard paste events. #21265
  • 🔧 Batch embeddings endpoint fix. The ‘/api/embeddings’ endpoint now correctly returns separate embeddings for each input string when processing batch requests to Ollama providers. Commit, #21279
  • 🗝️ SSL verification for embeddings. SSL certificate verification now respects the “AIOHTTP_CLIENT_SESSION_SSL” setting for OpenAI and Azure OpenAI embedding requests, allowing connections to self-signed certificate endpoints when disabled. Commit
  • 🔧 Tool call HTML entity fix. Models now receive properly formatted tool call results in multi-turn conversations, fixing an issue where HTML entities caused malformed content that was hard to parse. #20755
  • 💾 Duplicate inline image context fix. Inline images no longer exhaust the model’s context window by including their full base64 data in chat metadata, preventing premature context exhaustion with image-heavy conversations. #20916
  • 🐛 OpenAI model cache lookup fix. The OpenAI API router model lookup was corrected to use the proper model identifier when checking the cache, ensuring consistent and correct model retrieval during chat completions. #21327
  • 🐛 Ollama latest suffix fix. Ollama-compatible providers that don’t use “:latest” in model names can now successfully chat, fixing errors where model names were incorrectly appended with “:latest” suffixes. #21331, Commit
  • ⛔ OpenAI endpoint detection fix. OpenAI API endpoint detection was corrected to use exact hostname matching instead of substring matching, preventing third-party providers with similar URL patterns from being incorrectly filtered. Commit
  • 🛠️ RedisCluster task stopping fix. Task stopping now works correctly in RedisCluster deployments, fixing an issue where tasks would remain active after cancellation attempts. #20803, Commit
  • 📎 Citation parsing error fix. Citation parsing no longer crashes when builtin tools return error responses, fixing AttributeError issues when tools like search_web fail. #21071

Changed

  • ‼️ Database Migration Required — This release includes database schema changes; multi-worker, multi-server, or load-balanced deployments must update all instances simultaneously rather than performing rolling updates, as running mixed versions will cause application failures due to schema incompatibility between old and new instances.
  • ⚠️ Chat Message Table Migration — This release includes a new chat message table migration that can take a significant amount of time to complete in larger deployments with extensive chat histories. Administrators should plan for adequate maintenance windows and allow the migration to complete fully without interruption. Running the migration with insufficient time or resources may result in data integrity issues.
  • 🔗 Prompt ID-based URLs. Prompts now use unique ID-based URLs instead of command-based URLs, allowing more flexible command renaming without breaking saved links or integrations. #20945

[!TIP]
Looking for an Enterprise Plan? – Speak with Our Sales Team Today!

Get enhanced capabilities, including custom theming and branding, Service Level Agreement (SLA) support, and more!

👏 Massive thanks to our incredible contributors for their hard work and dedication to making this release possible: @aindriu80, @Edgaras0x4E, @tomasloksa, @ctolon, @ShirasawaSama, @sjhddh, @PeterDaveHello, @schneider-f, @joaoback, @silentoplayz, @tremlin, @aleixdorca, @veeceey, @Algorithm5838, @jannikstdl, @Th3Un1q3, @EntropyYue, @Classic298, @pelmennoteam, @mickeytheseal, @hsienz

Sponsors 🙌

🚀 We’d like to extend a heartfelt thank you to our amazing sponsors for their generous support (Note: We’ve excluded private sponsors from this list. If you’d like to get featured here, feel free to reach out to us!)

@amd, @roosi-gmbh, @rndmcnlly, @feddersen-group, @SRKConsulting, @mushmax, @icanbwell, @Classic298, @loitragg, @haka4githu6, @silenceroom, @walsermedia, @crizCraig, @GenieDev101, @CubicleGangster, @TOWaB, @comet-ml, @catalyst-net-nz, @JrdnHans, @hkosm, @joaoback, @LuxTronic

下载链接