发布日期: 2026-02-23
版本号: v0.8.4

0.8.4版本（2026-02-23）带来了一系列功能增强、性能优化和错误修复。新增功能方面，连接配置表单现在为流行AI提供者提供URL建议，便于设置；系统能正确获取Anthropic模型列表，并在无模型时提示用户管理连接。管理员和用户可单独启用或禁用提供者连接与提示词。代理现在可以删除和列出特定记忆。管理员可通过环境变量控制自动pip安装。新增Anthropic消息API代理端点，支持Claude等工具通过Open WebUI进行认证。OAuth支持多设备同时登录会话，并可配置组的默认分享设置。网页内容导入支持追加而非覆盖，技能可从JSON文件导入。新增You.com作为网页搜索提供者。性能方面，消息发送、历史记录加载、共享聊天、归档和固定聊天列表、聊天标题查询、标签筛选、批量访问授权查询、笔记列表、工具列表加载等多项操作得到显著优化。管理界面新增分析面板排序、模型公开/私有徽章、全局模型默认设置、组描述显示与排序、管理设置页签导航支持浏览器新标签页打开等。API与安全性方面，修复了工具输出显示、函数调用输入掩码、JSON日志格式支持，并增强了UI无障碍访问（屏幕阅读器与键盘导航支持）。问题修复部分解决了管理员功能页面崩溃、循环聊天历史死锁、模型回退路由、默认模型选择、滚动跳跃、新聊天消息处理、RAG模板变异、iframe沙箱安全、注册竞态条件、LDAP可选字段处理、API工具覆盖、嵌入与代理超时、文本文件类型检测、文件访问控制、PostgreSQL工作区克隆、MCP SSL验证、模型功能权限、模型选择器输入、禁用模型克隆、SCIM参数处理、混合搜索结果、导入项显示、WebSocket竞态条件、拖放叠加层、组搜索可见性、用户菜单拖放、管理与工作区导航拖放、Playground导航拖放、下拉菜单拖放、文件夹菜单拖放、控制台日志刷屏、Firefox头像溢出、暗黑模式选择框背景色、提示词导入、Ollama推理努力级别、混合搜索去重以及SQLAlchemy警告等问题。更改部分将提示词建议和横幅配置分别移至模型设置和通用设置中。

更新内容 （中文）

[0.8.4] - 2026-02-23

新增

  • 🛜 供应商URL建议。 连接表单现在显示一个下拉菜单，提供热门AI供应商的建议URL，便于配置连接。提交
  • ☁️ Anthropic模型获取。 系统现在能从Anthropic API正确获取可用模型，确保所有Anthropic模型均可访问。提交
  • 💡 无模型提示。 当没有可用模型时，系统会显示一个有用的提示，引导用户管理供应商连接。提交
  • ⚙️ 连接启用/禁用开关。 单个供应商连接现在可以在管理员和用户设置中启用或禁用。提交
  • ⏸️ 提示启用/禁用开关。 用户现在可以直接从提示列表中使用开关启用或禁用提示，无需删除并重新创建。未激活的提示会显示“未激活”徽章，并在列表中保持可见。提交
  • 🗑️ 内存删除。 智能体现在可以删除不再相关、重复或不正确的特定记忆，从而更好地控制存储的记忆内容。提交
  • 📋 内存列表。 智能体现在可以列出所有存储的记忆，使其能够根据完整的内存清单识别需要管理或删除的记忆。提交
  • 📦 自动pip安装开关。 管理员现在可以使用ENABLE_PIP_INSTALL_FRONTMATTER_REQUIREMENTS环境变量禁用从函数frontmatter需求中自动安装pip包，从而更精细地控制函数依赖管理。提交
  • 🔗 Anthropic Messages API代理。 新增的API端点现在支持Anthropic Messages API格式，允许Claude Code等工具通过Open WebUI进行身份验证并访问已配置的模型。工具调用现在在流式响应中得到正确支持，并具有正确的多块索引，工具错误状态也能正确传播。该端点将请求内部转换为OpenAI格式，通过现有的聊天管道路由，并以Anthropic格式返回响应。#21390，提交，提交
  • 👥 多设备OAuth会话。 用户现在可以使用OAuth同时在多个设备上保持登录状态，因为重新登录不会终止现有会话。当会话数超出限制时，最旧的会话会被自动清除。#21647，提交
  • 🔐 OAuth组默认共享设置。 管理员现在可以使用OAUTH_GROUP_DEFAULT_SHARE环境变量配置OAuth创建组的默认共享设置，从而控制新组是默认为私有还是与成员共享。#21679，提交
  • 🔧 知识库导入行为。 网页内容导入端点现在支持可配置的覆盖标志，允许用户向同一知识库添加多个URL，而不是替换现有内容。#21613，#21336，提交
  • 🧩 技能JSON导入支持。 技能现在可以从JSON和Markdown文件导入。#21511
  • 🔍 You.com网页搜索供应商。 新增了You.com网页搜索供应商选项，为用户提供了另一种用于支持网页模型的搜索引擎选择。You.com供应商通过包含描述和片段来丰富搜索结果，以提供更好的上下文。#21599
  • 🚀 消息列表性能。 发送消息时加载对话历史记录的速度显著提高，改善了模型开始生成前的响应延迟。这也加快了聊天搜索和RAG上下文构建。#21588
  • 🎯 并发嵌入请求控制。 管理员现在可以使用RAG_EMBEDDING_CONCURRENT_REQUESTS环境变量控制并发嵌入API请求的最大数量，有助于在保持嵌入性能的同时管理API速率限制。#21662，提交
  • ⚡ 消息更新优化。 通过消除每次消息更新（在流式响应期间多次发生）时进行的冗余数据库调用，消息保存期间加载聊天数据的速度显著提高。#21592
  • ⚡ 消息发送优化。 通过消除不必要的完整对话历史加载，消息发送期间加载聊天数据的速度显著提高。系统现在使用有针对性的查询，仅获取所需数据，而不是加载包含所有消息历史的完整聊天对象。#21596
  • 🚀 标签过滤优化。 带标签过滤的聊天搜索现在使用更高效的数据库查询，使过滤搜索的速度显著提高。提交
  • ⚡ 共享聊天加载优化。 共享聊天端点现在仅加载所需的列，而不是完整对话历史记录，使共享聊天列表的速度显著提高。#21614
  • 🗂️ 归档和置顶聊天加载。 通过仅加载所需的列而不是完整对话数据，加载归档和置顶聊天列表的速度显著提高。#21591
  • 💨 聊天标题查询优化。 检索聊天标题现在只查询标题列，而不是整个对话历史记录，使标题查找速度显著提高并减少了数据库负载。#21590
  • 🗄️ 多资源批量访问授权。 加载频道、知识库、模型、笔记、提示、技能和工具时，现在对访问授权使用批量数据库查询，而不是逐项查询，显著降低了数据库负载。对于30个项目，这将大约31个查询减少到仅3个。#21616
  • 📋 笔记列表负载优化。 笔记列表和搜索端点现在仅返回200字符的预览，而不是完整的笔记内容，将60个笔记的响应负载从约167 MB减少到约10 KB，并消除了访问授权的N+1查询。笔记选项卡现在可以在几秒内加载，而不是几十秒。#21549
  • ⚡ 工具列表性能。 通过延迟数据库查询中的内容和规范字段，并使用缓存的工具模块而不是为每个请求重新加载它们，加载工具列表的速度显著提高。提交
  • 📝 组描述显示。 管理员组列表现在显示每个组的描述，使管理员更容易一目了然地识别组。
  • 🏷️ 排序下拉菜单。 管理员现在可以使用带有“名称”或“成员”选项的下拉菜单对组进行排序，替代了以前的可点击列标题。
  • 📶 管理员组列表排序。 管理员组列表中的“组”和“用户”列现在可点击进行排序，允许管理员按名称字母顺序或按成员数量数字顺序对组进行排序。#21692
  • 🔽 富UI自动滚动。 视图现在会在富UI内容渲染完成后自动滚动到操作生成的富UI内容，确保用户无需手动滚动即可看到结果。#21698，#21482
  • 📊 管理员分析开关。 管理员现在可以使用ENABLE_ADMIN_ANALYTICS环境变量启用或禁用分析功能，从而更精细地控制可用的管理员功能。#21651，提交
  • 📊 分析排序增强。 分析仪表板现在支持按“模型使用情况”和“用户使用情况”表中的“令牌”列进行排序，“分享/百分比”列现在也可点击进行排序。管理员可以更容易地识别令牌消耗最多的模型和用户。提交
  • 📑 获取URL引用来源。 当模型在工具调用期间获取URL时，获取的URL现在会作为可点击的引用源显示在UI中，并带有内容预览，与现有的网页搜索和知识文件工具行为相匹配。#21669
  • 🔗 管理员设置选项卡导航。 管理员设置侧边栏现在支持原生浏览器选项卡打开，允许用户通过中键单击或右键单击在新选项卡中打开设置页面。导航已从基于按钮转换为基于锚点元素。#21721
  • 🏷️ 模型可见性徽章。 管理员设置模型页面现在直接在每个模型上显示“公开”或“私有”徽章，无需打开编辑屏幕即可轻松一目了然地识别模型访问级别。#21732，提交
  • 🛠️ 全局模型默认值。 管理员现在可以配置默认元数据和参数，这些配置将自动应用于所有模型，减少新发现模型的手动配置工作。默认功能（如视觉、网页搜索、代码解释器）和参数（如温度、max_tokens）可以在管理员设置中全局设置，仍允许每个模型覆盖。#20658，提交
  • 💬 纯文本工具输出显示。 纯字符串的工具输出现在自然显示在等宽字体块中，而不是引号/转义格式，使多行字符串输出更易于阅读。#21553，提交
  • 🔐 事件调用输入掩码。 函数现在可以在确认对话框中请求掩码密码输入，允许隐藏敏感数据的输入。这扩展了现有的用户阀门掩码功能到事件调用。#21540，提交
  • 🗂️ JSON日志支持。 管理员现在可以通过将LOG_FORMAT环境变量设置为“json”来启用JSON格式的日志记录，使日志适用于Loki、Fluentd、CloudWatch和Datadog等日志聚合器。#21747
  • ♿ UI无障碍性改进。 屏幕阅读器用户现在可以更轻松地导航界面，对话框中的键盘导航得到改进，所有交互元素都有了正确的ARIA标签。在各个组件中的关闭、返回和操作按钮上添加了aria-labels，并改进了身份验证、侧边栏、聊天和通知组件中的语义HTML和屏幕阅读器支持，解决了WCAG合规性问题。在管理员和用户设置中的搜索输入框、选择字段和模态框上添加了aria-labels，并改进了文本输入、评级组件、引用和网页搜索结果的无障碍性。在工作区组件（包括知识库、模型、提示、技能和工具页面）上添加了aria-labels，以改善屏幕阅读器支持。#21706，#21705，#21710，#21709，#21717，#21715，#21708，#21719
  • 🔄 一般性改进。 在整个应用程序中实施了各种改进，以增强性能、稳定性和安全性。
  • 🌐 芬兰语、法语、葡萄牙语（巴西）、简体中文和繁体中文的翻译得到增强和扩展。

修复

  • 💥 管理员函数页面崩溃修复。 管理员“函数”选项卡在点击时不再崩溃，修复了函数列表加载期间发生的空引用错误。#21661，提交
  • 💀 循环聊天历史死锁修复。 具有循环父子消息引用的聊天历史在同步使用统计信息时不再导致后端冻结。系统现在会检测并在遇到循环消息引用时安全地中止。#21681
  • 🔀 模型回退路由修复。 自定义模型回退现在在所有模型类型上都能正常工作，防止当回退模型使用不同后端（管道、Ollama或OpenAI）时出现“未找到模型”错误。#21736
  • 🐛 默认模型选择修复。 管理员配置的默认模型现在在开始新聊天时得到正确遵守，而不是被第一个可用模型覆盖。#21736
  • 👁️ 滚动跳转修复。 在停止生成后删除消息对不再导致聊天在视觉上跳来跳去，使消息删除更流畅。#21743，提交
  • 💬 新聊天消息处理修复。 修复了在发送消息后点击“新聊天”会静默丢弃后续消息的错误。系统现在在开始新对话时正确清除待处理的消息队列。#21731
  • 🔍 RAG模板变异修复。 修复了在多次连续工具调用期间，RAG模板文本被递归注入用户消息中，导致消息内容呈指数级增长并可能混淆模型的错误。系统现在在工具调用循环之前保留原始用户消息，并正确累积引用源。#21663，#21668，提交
  • 🔒 iframe沙盒安全。 嵌入式工具默认情况下无法再提交表单或访问同源内容，从而提高了用户的安全性。#21529
  • 🔐 注册竞争条件修复。 修复了一个安全漏洞，该漏洞在首次部署时运行多个uvicorn工作线程可能导致创建多个管理员帐户。注册处理程序现在在首次用户注册期间正确处理并发请求，防止未经授权的管理员权限提升。#21631
  • 🔐 LDAP可选字段修复。 LDAP配置现在正确接受空的应用程序DN和密码值，允许在不提供这些可选字段的情况下进行LDAP身份验证。以前，空值会导致身份验证失败。提交
  • 🛠️ API工具修复。 /api/v1/chat/completions端点现在正确遵守调用方提供的工具，而不是用服务器端工具覆盖它们，修复了Claude Code或Cursor等外部代理接收到意外工具广告的问题。#21557，#21555
  • ⏱️ 嵌入和代理超时修复。 嵌入和OpenAI代理端点现在正确遵守AIOHTTP_CLIENT_TIMEOUT环境变量，而不是使用可能导致请求挂起的默认超时。#21558
  • 📄 文本文件类型检测修复。 TypeScript和其他根据扩展名被错误检测为视频文件的文本文件，现在被正确识别和处理为文本文件，修复了.ts文件的上传拒绝问题。#21454，提交
  • 🗄️ 文件访问控制遵守。 文件列表和搜索端点现在正确遵守BYPASS_ADMIN_ACCESS_CONTROL设置，确保管理员在禁用该设置时只能看到自己的文件，与其他端点保持一致。#21595，#21589
  • 🗄️ PostgreSQL工作区克隆。 通过为访问授权生成正确的唯一ID，而不是使用可能重复或无效的ID，克隆工作区模型现在在PostgreSQL数据库上正常工作。提交
  • 🔓 MCP SSL验证修复。 MCP工具连接现在正确遵守AIOHTTP_CLIENT_SESSION_TOOL_SERVER_SSL环境变量以禁用SSL验证，而不是始终验证SSL证书。提交，#21481
  • 🔒 模型默认功能权限。 模型默认功能（如代码解释器、网页搜索和图像生成）现在遵守全局配置和用户权限设置，防止已禁用的功能出现在聊天输入中。#21690
  • 🔍 模型选择器输入修复。 由于改进了虚拟滚动处理，在搜索字段中快速输入时，模型选择器列表不再消失或变成灰色。#21659
  • ⛔ 禁用模型克隆防止。 禁用模型无法再作为工作区模型克隆，防止创建无效的空配置。“克隆”选项现在对不活动的模型隐藏。#21724
  • 🔧 SCIM参数处理。 SCIM用户和组端点现在接受超出范围的startIndex和count值，通过将其限制在有效范围内而不是返回错误，符合RFC 7644。#21577
  • 🔍 混合搜索结果修复。 修复了查询结果解包顺序不匹配导致搜索结果显示为空的bug后，混合搜索现在返回正确的结果。#21562
  • 🛠️ 导入项目显示。 导入的函数和工具现在在导入后立即出现在列表中，无需重新加载页面。#21593
  • 🔄 WebSocket竞争条件修复。 协作笔记保存不再在用户断开连接且待处理保存未完成时因错误而崩溃，防止了AttributeError异常和过多的日志记录。#21601，提交
  • ✋ 拖放叠加层修复。 在Mozilla Firefox中将文件拖出聊天窗口后，“添加文件”叠加层不再停留在屏幕上。#21664
  • 👁️ 组搜索可见性修复。 即使搜索未匹配任何用户，组现在也会在访问控制搜索结果中正确显示。#21691
  • 🖱️ 用户菜单拖动和点击修复。 修复了拖动菜单项时出现的可拖动幽灵图像，并消除了在下拉菜单外拖动时出现的幻影链接点击。#21699
  • 🧭 管理员和工作区导航拖动修复。 通过添加适当的拖动约束和防止文本选择，修复了在管理员和工作区面板中拖动顶部导航选项卡时出现的幽灵拖动图像。#21701
  • 🎮 Playground导航拖动修复。 通过添加适当的拖动约束和防止文本选择，修复了在Playground面板中拖动顶部导航选项卡时出现的幽灵拖动图像。#21704
  • ✋ 下拉菜单拖动修复。 下拉菜单项现在无法在突出显示文本时被意外拖动为幽灵图像，使菜单交互更流畅。#21713
  • 🗂️ 文件夹菜单拖动修复。 文件夹下拉菜单项现在无法被意外突出显示或拖动为幽灵图像，使文件夹选项的行为类似于标准菜单。#21753
  • 📝 控制台日志垃圾修复。 请求已删除或缺失的文件不再用Python回溯日志淹没后端控制台，这要归功于对预期404错误的正确异常处理。#21687
  • 🐛 Firefox头像溢出修复。 修复了Firefox中的一个视觉bug，即损坏的模型或用户头像图像会显示溢出的alt文本，覆盖分析和排行榜页面上的相邻标签。失败的头像图像现在正确显示回退图标。#21730
  • 🎨 暗色模式选择背景修复。 通过移除冲突的暗色主题覆盖，修复了选择输入和下拉菜单在暗色模式下背景颜色不一致的问题，确保了协调的透明外观。#21728
  • 💾 提示导入修复。 导入先前导出的提示不再因“[object Object]”错误提示而失败，使提示备份和恢复正常工作。#21594
  • 🔧 Ollama推理努力修复。 推理努力现在与需要字符串值（“low”、“medium”、“high”）而不是布尔值的Ollama模型正常工作，修复了使用GPT-OSS等模型时出现的“提供了无效选项”错误。#20921，#20928，提交
  • 🔍 混合搜索去重修复。 混合搜索现在使用内容哈希正确对结果进行去重，防止在使用富文本进行BM25搜索时出现重复块。提交
  • 📋 SQLAlchemy警告修复。 修复了删除共享聊天时日志中出现的SQLAlchemy警告，提高了日志清晰度。提交

变更

  • 🎯 提示建议位置调整。 提示建议已从管理员面板 - 设置 - 界面移至管理员面板 - 设置 - 模型，现在可以按模型或通过新的模型默认值进行全局配置。
  • 📢 横幅位置调整。 横幅配置已从管理员面板 - 设置 - 界面移至管理员面板 - 设置 - 常规。

更新内容 （原始）

[0.8.4] - 2026-02-23

Added

  • 🛜 Provider URL suggestions. The connection form now displays a dropdown with suggested URLs for popular AI providers, making it easier to configure connections. Commit
  • ☁️ Anthropic model fetching. The system now properly fetches available models from the Anthropic API, ensuring all Anthropic models are accessible. Commit
  • 💡 No models prompt. When no models are available, a helpful prompt now guides users to manage their provider connections. Commit
  • ⚙️ Connection enable/disable toggles. Individual provider connections can now be enabled or disabled from both admin and user settings. Commit
  • ⏸️ Prompt enable/disable toggle. Users can now enable or disable prompts directly from the prompts list using a toggle switch, without needing to delete and recreate them. Inactive prompts display an “Inactive” badge and are still visible in the list. Commit
  • 🗑️ Memory deletion. Agents can now delete specific memories that are no longer relevant, duplicated, or incorrect, giving better control over stored memory content. Commit
  • 📋 Memory listing. Agents can now list all stored memories, enabling them to identify which memories to manage or delete based on the complete memory inventory. Commit
  • 📦 Auto pip install toggle. Administrators can now disable automatic pip package installation from function frontmatter requirements using the ENABLE_PIP_INSTALL_FRONTMATTER_REQUIREMENTS environment variable, providing more control over function dependency management. Commit
  • 🔗 Anthropic Messages API proxy. A new API endpoint now supports the Anthropic Messages API format, allowing tools like Claude Code to authenticate through Open WebUI and access configured models. Tool calls are now properly supported in streaming responses with correct multi-block indexing, and error status from tools is propagated correctly. The endpoint converts requests to OpenAI format internally, routes them through the existing chat pipeline, and returns responses in Anthropic format. #21390, Commit, Commit
  • 👥 Multi-device OAuth sessions. Users can now stay logged in on multiple devices simultaneously with OAuth, as re-logging in no longer terminates existing sessions. The oldest sessions are automatically pruned when the session limit is exceeded. #21647, Commit
  • 🔐 OAuth group default share setting. Administrators can now configure the default sharing setting for OAuth-created groups using the OAUTH_GROUP_DEFAULT_SHARE environment variable, allowing control over whether new groups default to private or shared with members. #21679, Commit
  • 🔧 Knowledge base import behavior. The web content import endpoint now supports a configurable overwrite flag, allowing users to add multiple URLs to the same knowledge base instead of replacing existing content. #21613, #21336, Commit
  • 🧩 Skill JSON import support. Skills can now be imported from both JSON and Markdown files. #21511
  • 🔍 You.com web search provider. A new web search provider option for You.com is now available, giving users another search engine choice for web-enabled models. The You.com provider enriches search results by including both descriptions and snippets for better context. #21599
  • 🚀 Message list performance. Loading conversation history when sending messages is now significantly faster, improving response latency before the model starts generating. This also speeds up chat search and RAG context building. #21588
  • 🎯 Concurrent embedding request control. Administrators can now control the maximum number of concurrent embedding API requests using the RAG_EMBEDDING_CONCURRENT_REQUESTS environment variable, helping manage API rate limits while maintaining embedding performance. #21662, Commit
  • ⚡ Message upsert optimization. Loading chat data during message saving is now significantly faster by eliminating a redundant database call that occurred on every message upsert, which happens many times during streaming responses. #21592
  • ⚡ Message send optimization. Loading chat data during message sending is now significantly faster by eliminating unnecessary full conversation history loads. The system now uses targeted queries that fetch only the needed data instead of loading entire chat objects with all message history. #21596
  • 🚀 Tag filtering optimization. Chat search with tag filtering now uses more efficient database queries, making filtered searches significantly faster. Commit
  • ⚡ Shared chat loading optimization. The shared chats endpoint now loads only the needed columns instead of the full conversation history, making shared chat listings significantly faster. #21614
  • 🗂️ Archived and pinned chat loading. Loading archived and pinned chat lists is now significantly faster by loading only the needed columns instead of full conversation data. #21591
  • 💨 Chat title query optimization. Retrieving chat titles now queries only the title column instead of the entire conversation history, making title lookups significantly faster and reducing database load. #21590
  • 🗄️ Batch access grants for multiple resources. Loading channels, knowledge bases, models, notes, prompts, skills, and tools now uses batch database queries for access grants instead of individual queries per item, significantly reducing database load. For 30 items, this reduces approximately 31 queries to just 3. #21616
  • 📋 Notes list payload optimization. Notes list and search endpoints now return only a 200-character preview instead of the full note content, reducing response payload from ~167 MB to ~10 KB for 60 notes and eliminating N+1 queries for access grants. The Notes tab now loads in seconds instead of tens of seconds. #21549
  • ⚡ Tools list performance. Loading the tools list is now significantly faster by deferring content and specs fields from database queries, and using cached tool modules instead of reloading them for each request. Commit
  • 📝 Group description display. The admin groups list now shows each group’s description, making it easier for administrators to identify groups at a glance.
  • 🏷️ Sort by dropdown. Administrators can now sort groups using a dropdown menu with options for Name or Members, replacing the previous clickable column headers.
  • 📶 Admin groups list sorting. The Group and Users columns in the admin groups list are now clickable for sorting, allowing administrators to sort groups alphabetically by name or numerically by member count. #21692
  • 🔽 Rich UI auto-scroll. The view now automatically scrolls to action-generated Rich UI content once it renders, ensuring users can see the results without manually scrolling. #21698, #21482
  • 📊 Admin analytics toggle. Administrators can now enable or disable the analytics feature using the ENABLE_ADMIN_ANALYTICS environment variable, giving more control over available admin features. #21651, Commit
  • 📊 Analytics sorting enhancement. The Analytics dashboard now supports sorting by Tokens column for both Model Usage and User Usage tables, and the Share/Percentage columns are now clickable for sorting. Administrators can more easily identify the most token-consuming models and users. Commit
  • 📑 Fetch URL citation sources. When models fetch URLs during tool calling, the fetched URLs now appear as clickable citation sources in the UI with content previews, matching the existing behavior of web search and knowledge file tools. #21669
  • 🔗 Admin settings tab navigation. The admin settings sidebar now supports native browser tab opening, allowing users to middle-click or right-click to open settings pages in new tabs. The navigation was converted from button-based to anchor-based elements. #21721
  • 🏷️ Model visibility badges. The Admin Settings Models page now displays Public or Private badges directly on each model, making it easy to identify model access levels at a glance without opening the edit screen. #21732, Commit
  • 🛠️ Global model defaults. Administrators can now configure default metadata and parameters that automatically apply to all models, reducing manual configuration for newly discovered models. Default capabilities (like vision, web search, code interpreter) and parameters (like temperature, max_tokens) can be set globally in Admin Settings, with per-model overrides still available. #20658, Commit
  • 💬 Plaintext tool output display. Tool outputs that are plain strings now display naturally in a monospace block instead of quoted/escaped format, making multi-line string outputs easier to read. #21553, Commit
  • 🔐 Event call input masking. Functions can now request masked password input in confirmation dialogs, allowing sensitive data entry to be hidden from view. This extends the existing masking feature from user valves to event calls. #21540, Commit
  • 🗂️ JSON logging support. Administrators can now enable JSON-formatted logging by setting the LOG_FORMAT environment variable to “json”, making logs suitable for log aggregators like Loki, Fluentd, CloudWatch, and Datadog. #21747
  • ♿ UI accessibility improvements. Screen reader users can now navigate the interface more easily with improved keyboard navigation in dialogs and proper ARIA labels on all interactive elements. Added aria-labels to close, back, and action buttons across various components, and improved semantic HTML and screen reader support across auth, sidebar, chat, and notification components, addressing WCAG compliance. Added aria-labels to search inputs, select fields, and modals in admin and user settings, and improved accessibility for text inputs, rating components, citations, and web search results. Added aria-labels to workspace components including Knowledge, Models, Prompts, Skills, and Tools pages for improved screen reader support. #21706, #21705, #21710, #21709, #21717, #21715, #21708, #21719
  • 🔄 General improvements. Various improvements were implemented across the application to enhance performance, stability, and security.
  • 🌐 Translations for Finnish, French, Portuguese (Brazil), Simplified Chinese, and Traditional Chinese were enhanced and expanded.

Fixed

  • 💥 Admin functions page crash fix. The admin Functions tab no longer crashes when clicked, fixing a null reference error that occurred while the functions list was loading. #21661, Commit
  • 💀 Cyclic chat history deadlock fix. Chat histories with circular parent-child message references no longer cause the backend to freeze when syncing usage stats. The system now detects and safely aborts when encountering cyclic message references. #21681
  • 🔀 Model fallback routing fix. Custom model fallback now works correctly across all model types, preventing “Model not found” errors when the fallback model uses a different backend (pipe, Ollama, or OpenAI). #21736
  • 🐛 Default model selection fix. Admin-configured default models are now properly respected when starting new chats instead of being overwritten by the first available model. #21736
  • 👁️ Scroll jumping fix. Deleting a message pair after stopping generation no longer causes the chat to visually jump around, making message deletion smoother. #21743, Commit
  • 💬 New chat message handling fix. Fixed a bug where clicking “New Chat” after sending a message would silently drop subsequent messages. The system now properly clears pending message queues when starting a new conversation. #21731
  • 🔍 RAG template mutation fix. Fixed a bug where RAG template text was recursively injected into user messages during multiple sequential tool calls, causing message content to grow exponentially and potentially confuse the model. The system now preserves the original user message before tool-calling loops and correctly accumulates citation sources. #21663, #21668, Commit
  • 🔒 Iframe sandbox security. Embedded tools can no longer submit forms or access same-origin content by default, improving security for users. #21529
  • 🔐 Signup race condition fix. Fixed a security vulnerability where multiple admin accounts could be created on fresh deployments when running multiple uvicorn workers. The signup handler now properly handles concurrent requests during first-user registration, preventing unauthorized admin privilege escalation. #21631
  • 🔐 LDAP optional fields fix. LDAP configuration now properly accepts empty Application DN and password values, allowing LDAP authentication to work without these optional fields. Previously, empty values caused authentication failures. Commit
  • 🛠️ API tools fix. The /api/v1/chat/completions endpoint now properly respects caller-provided tools instead of overriding them with server-side tools, fixing issues where external agents like Claude Code or Cursor would receive unexpected tool advertisements. #21557, #21555
  • ⏱️ Embeddings and proxy timeout fix. The embeddings and OpenAI proxy endpoints now properly honor the AIOHTTP_CLIENT_TIMEOUT environment variable, instead of using default timeouts that could cause requests to hang. #21558
  • 📄 Text file type detection fix. TypeScript and other text files that were mis-detected as video files based on their extension are now correctly identified and processed as text files, fixing upload rejections for .ts files. #21454, Commit
  • 🗄️ File access control respect. The files list and search endpoints now properly respect the BYPASS_ADMIN_ACCESS_CONTROL setting, ensuring admins only see their own files when the setting is disabled, consistent with other endpoints. #21595, #21589
  • 🗄️ PostgreSQL workspace cloning. Cloning workspace models now works correctly on PostgreSQL databases by generating proper unique IDs for access grants instead of using potentially duplicate or invalid IDs. Commit
  • 🔓 MCP SSL verification fix. MCP tool connections now properly respect the AIOHTTP_CLIENT_SESSION_TOOL_SERVER_SSL environment variable to disable SSL verification, instead of always verifying SSL certificates. Commit, #21481
  • 🔒 Model default feature permissions. Model default features like code interpreter, web search, and image generation now respect global configuration and user permission settings, preventing disabled features from appearing in the chat input. #21690
  • 🔍 Model selector typing fix. The model selector list no longer disappears or becomes grayed out when typing quickly in the search field, thanks to improved virtual scroll handling. #21659
  • ⛔ Disabled model cloning prevention. Disabled models can no longer be cloned as workspace models, preventing invalid empty configurations from being created. The Clone option is now hidden for inactive models. #21724
  • 🔧 SCIM parameter handling. The SCIM Users and Groups endpoints now accept out-of-range startIndex and count values by clamping them to valid ranges instead of returning errors, in compliance with RFC 7644. #21577
  • 🔍 Hybrid search result fix. Hybrid search now returns correct results after fixing a bug where query result unpacking order was mismatched, causing search results to appear empty. #21562
  • 🛠️ Imported items display. Imported functions and tools now appear immediately in the list after import, without requiring a page reload. #21593
  • 🔄 WebSocket race condition fix. Collaborative note saves no longer crash with errors when users disconnect before pending saves complete, preventing AttributeError exceptions and excessive logging. #21601, Commit
  • ✋ Drag-and-drop overlay fix. The “Add Files” overlay no longer remains stuck on screen when dragging files back out of the chat window in Mozilla Firefox. #21664
  • 👁️ Group search visibility fix. Groups now appear correctly in access control search results, even when the search doesn’t match any users. #21691
  • 🖱️ User menu drag and click fixes. Fixed draggable ghost images when dragging menu items and eliminated phantom link clicks that occurred when dragging outside dropdown menus. #21699
  • 🧭 Admin and workspace nav drag fix. Fixed ghost drag images when dragging top navigation tabs in the Admin and Workspace panels by adding proper drag constraints and text selection prevention. #21701
  • 🎮 Playground nav drag fix. Fixed ghost drag images when dragging top navigation tabs in the Playground panel by adding proper drag constraints and text selection prevention. #21704
  • ✋ Dropdown menu drag fix. Dropdown menu items can no longer be accidentally dragged as ghost images when highlighting text, making menu interactions smoother. #21713
  • 🗂️ Folder menu drag fix. Folder dropdown menu items can no longer be accidentally highlighted or dragged as ghost images, making folder options behave like standard menus. #21753
  • 📝 Console log spam fix. Requesting deleted or missing files no longer floods the backend console with Python traceback logs, thanks to proper exception handling for expected 404 errors. #21687
  • 🐛 Firefox avatar overflow fix. Fixed a visual bug in Firefox where broken model or user avatar images would display overflowing alt text that overlapped adjacent labels on the Analytics and Leaderboard pages. Failed avatar images now properly show fallback icons instead. #21730
  • 🎨 Dark mode select background fix. Fixed an issue where select inputs and dropdown menus had inconsistent lighter background colors in dark mode by removing conflicting dark theme overrides, ensuring a cohesive transparent look. #21728
  • 💾 Prompt import fix. Importing prompts that were previously exported no longer fails with a “[object Object]” error toast, making prompt backup and restore work correctly. #21594
  • 🔧 Ollama reasoning effort fix. Reasoning effort now works correctly with Ollama models that require string values (“low”, “medium”, “high”) instead of boolean, fixing “invalid option provided” errors when using models like GPT-OSS. #20921, #20928, Commit
  • 🔍 Hybrid search deduplication fix. Hybrid search now correctly deduplicates results using content hashes, preventing duplicate chunks from appearing when using enriched text for BM25 search. Commit
  • 📋 SQLAlchemy warning fix. Fixed a SQLAlchemy warning that appeared in logs when deleting shared chats, improving log clarity. Commit

Changed

  • 🎯 Prompt suggestions relocated. Prompt suggestions have been moved from Admin Panel - Settings - Interface to Admin Panel - Settings - Models, where they can now be configured per-model or globally via the new model defaults.
  • 📢 Banners relocated. Banners configuration has been moved from Admin Panel - Settings - Interface to Admin Panel - Settings - General.

下载链接