github.com/Tencent/WeKnora

New Features

• NEW: Local Browser (BrowserSkill) — agents in smart-reasoning chats can drive the user's own Chrome or Edge through the open-source BrowserSkill extension (v0.3.1+). Pair once from Toolbox → Browser Connection, turn on "Local Browser" in the composer, and the new local_browser tool opens pages, clicks, fills forms and reads content in a dedicated task window. The chat shows a live (picture-in-picture) task preview with pause / resume / end controls; logins and CAPTCHAs are handed to the user through request_help, and interrupted tasks can be resumed. The app image bundles the bsk daemon, the matching extension ZIP and licenses; multi-replica deployments set BROWSERSKILL_INTERNAL_URL / BROWSERSKILL_CLUSTER_SECRET (#3180, #3193, #3206, #3563).
• NEW: Built-in MCP Server — each workspace can publish MCP endpoints at /mcp/<endpoint_id> (Streamable HTTP) from Settings → Integrations → MCP Server, with its own token, knowledge-base scope, per-minute rate limit (default 60) and tool groups: retrieval & reading, ask (runs the endpoint's default agent), Wiki, and write tools (off by default). The page generates mcpServers config for Cursor / VS Code / Claude Desktop and a one-line Claude Code command; tokens can be rotated. Endpoints are managed through /api/v1/mcp-endpoints (API key capability manage_channels). The Python mcp-server/ is now deprecated (#3371).
• NEW: Toolbox — skills, MCP services and the browser connection move from Settings into a sidebar Toolbox (old settings links redirect); the sidebar shows the browser connection status.
• NEW: Conversation control — append requirements to a running agent turn (queued until the step ends, or injected immediately) (#3123); fork a conversation from any earlier question (#3304); rewind in place, resetting the sandbox workspace to the matching checkpoint (#3469); choose the reasoning effort per session.
• NEW: Artifacts library — a sidebar "Artifacts" page lists every generated file across conversations with type filters, search, date grouping and version history, backed by a message_artifacts table; generated files can be deleted (#3384, #3447).
• NEW: Sandbox terminal & graphical desktop — the sandbox panel beside the chat gains an interactive terminal (Cube / E2B) that reconnects to running shells (#3110, #3128) and a browser-based XFCE desktop for sandboxes built from a desktop template (desktop_enabled) (#3245). Shell and skill-install commands stream their output live; idle terminal/desktop sessions disconnect after terminal_idle_disconnect_sec (default 900 s).
• NEW: Lite host sandbox & project folders — the desktop app (macOS) runs sessions without a pinned sandbox inside an OS sandbox (Seatbelt): no network, writes limited to the chosen project folder, .git read-only, credential directories blocked. New chats can bind a host project directory or use a dated temporary workspace (#3472, #3549, #3554).
• NEW: Model catalog — model integration is rebuilt on protocol, vendor and catalog layers (#3470, #3589): 27 built-in vendors, a generated catalog that fills context window, max output, reasoning levels and vision support, a "resolved call" panel in the model editor, per-model protocol override and protocol-compat JSON, GET/POST /models/catalog/resolve, and a deployment overlay file (config/models.json, MODELS_CONFIG). Speech-to-text now covers every vendor that documents it (#3503).
• NEW: Confluence & DingTalk Docs data sources — Confluence Cloud and Server/Data Center spaces sync published pages incrementally with deletion protection (#3287); DingTalk Docs sync with localized setup (#2313). IMA sync traverses subfolders.
• NEW: Bocha and Serply web search providers (#3036); country / freshness filters work with Brave and Serply.
• NEW: Japanese UI (ja-JP), selectable at runtime or via DEFAULT_LOCALE (#3080).
• NEW: Per-channel IM reply language — each IM channel can set its reply locale; unset channels follow WEKNORA_LANGUAGE (#3455).
• NEW: Knowledge base workflow — batch download of original files as a ZIP (#3300); configurable sorting for document and resource lists (#3318); documents that stop making progress are flagged as queued or stalled with a TASK_STALLED code (#3557); per-upload option to skip summary generation; knowledge base descriptions generated from document profiles (#3370); an upload panel with overall progress, cancel and retry (#3399); FAQ list filter by enabled status (#3157); activity records show the API key name (#3405).
• NEW: Whitelist-only outbound mode — SSRF_DNS_WHITELIST_ONLY=true refuses any host outside SSRF_WHITELIST before DNS resolution, closing DNS-based exfiltration from offline deployments (#3393).
• NEW: Admin user creation UI — system administrators create accounts from Settings with a one-time password and a copyable account summary (#3034, #3039).
• NEW: CLI setup page under Integrations with a three-step guide (#3081).

Improvements

• IMPROVED: Agent knowledge-retrieval tools consolidated — knowledge_search and grep_chunks are replaced by a single search_knowledge tool (query, mode = hybrid | semantic | keyword, optional knowledge_base_ids, limit default 10 / max 30); keyword mode is served by the keyword index (BM25 / engine keyword retrieval) instead of an unindexed regular-expression scan of the chunks table. list_knowledge_chunks, get_document_info and wiki_read_source_doc are replaced by read_document (id = a dN document or cN chunk handle, offset / limit paging, in-document query with optional regex, context neighbours around a chunk), which always returns the document metadata header together with the chunks. New list_documents (knowledge_base_id, keyword, page, page_size) pages through one knowledge base. wiki_search takes a single query (a case-insensitive regular expression, as before) plus optional regex, knowledge_base_ids and limit; the old queries / knowledge_base_id parameters are still accepted. query_knowledge_graph is only offered when a graph-enabled knowledge base is in scope. New agents default to search_knowledge, read_document, list_documents and search_conversations. The built-in MCP server keeps its tool names: search_knowledge gains mode and limit, grep_chunks keeps its regular-expression contract on index-backed retrieval and gains limit, read_document gains query, and wiki_search gains regex.
• IMPROVED: Retrieval tool compatibility fixes — search_knowledge no longer refuses a mode that some bases cannot serve: FAQ and vector-only bases answer keyword requests semantically and keyword-only bases answer semantic requests by keyword, with requested_mode / mode_fallbacks in the result and <mode_fallback> in the model rendering, which now shows the mode actually used instead of always semantic. read_document and list_documents are offered on wiki-only scopes (chunks are always stored) without widening RAG agents' derived KB filter; list_documents applies the @file / @tag scope before paging so totals and pages only count readable documents; chunk context is resolved by chunk index so parent, summary and image chunks no longer shift the neighbourhood. wiki_search (agent and MCP) keeps its regular-expression semantics by default and only falls back to a literal match for text that is not a valid expression; MCP grep_chunks keeps grep semantics by fetching candidates with the pattern's literal terms and verifying each chunk against the pattern.
• IMPROVED: Backward-compatible tool aliasing — old tool names stored in agent configurations, presets and API calls are mapped to their successors at run time (legacyToolSuccessors / NormalizeAllowedTools); no data migration is needed, and stored chat history that references the old tool names still renders.
• IMPROVED: Agent context — history is sized by the model's context window with calibrated token estimates (#3413); compaction checkpoints persist across turns (#3407); search_knowledge reranks with document titles and allows a keyword retry after a rerank rejection; read_document queries are keyword-based and bounded.
• IMPROVED: MCP client — tool catalogs are persisted per service and caller with a guided two-step setup and AI-drafted usage notes (#3118); tools can be enabled or disabled individually (#2989); agents discover and call MCP tools on demand through discover_mcp_tools / call_mcp_tool instead of loading every tool; advanced_config.timeout extends the call window; multi-line SSE events are assembled correctly.
• IMPROVED: Skills — outdated sandbox skills upgrade without rolling back the catalog or taking the skill offline (#3406); runtime prerequisites are verified after install and installs accept extra guidance; manifests with a UTF-8 BOM load.
• IMPROVED: Chat UX — rebuilt Markdown editor for saving answers to a knowledge base (#3467); in-progress status and API-session owner badges in the sidebar; full dates on older message timestamps (#3576); truncated answers are marked instead of silently cut; stored streams replay one segment at a time; the chat stream recovers from an expired access token (#3028).
• IMPROVED: Wiki — whole-page rewrites continue when cut by the completion budget and are refused if still incomplete (#3582); same-title items reuse the existing page across types (#3547); cited chunks include image captions (#3292); ambiguous single-Han auto-links are skipped (#2860); taxonomy embeddings are batched.
• IMPROVED: Retrieval & graph — Milvus supports multilingual BM25 (#3235); keyword-only BM25 scores are bounded before rerank (#3343); Neo4j node search caps seeds and rows (#3550); entity extraction gets a larger output budget (#2859) and uses OCR text of scanned pages (#3092).
• IMPROVED: Parsing — PaddleOCR-VL request timeout is configurable via WEKNORA_PADDLEOCR_VL_TIMEOUT (#3537); timed-out converter process groups are killed (#2990); PDFs stay on the builtin parser even when anydoc is linked; inline HTML tables no longer produce oversized chunks (#3404); the OCR prompt no longer includes the knowledge base's image-caption instructions (#3291).
• IMPROVED: Performance — faster chunk ingestion and batch updates (#3369); fewer redundant sandbox remote calls.
• IMPROVED: Storage — OBS uses virtual-hosted addressing (#3273); object-storage transfers may exceed 30 seconds (#3368); the KS3 client keeps the SSRF redirect policy.
• IMPROVED: Frontend — style consistency pass with shared shells, tokens and unified confirmations (#3376); list title icons and resource cards refined; embedded pages enforce host origin allowlists.

Bug Fixes

• FIXED: Documents no longer get stranded in processing or deleting: pipeline errors and panics fail the row, never-run stages are not reported as failed, stranded Wiki queues release their documents, and housekeeping recovers stuck deletions (#3445, #3367).
• FIXED: Cross-workspace access paths in shared spaces are closed (#3408, #3416); path traversal in local prefixes, task IDs and Wiki sort is rejected.
• FIXED: Wiki tasks of soft-deleted tenants stop issuing model requests (#3602); deleting a knowledge base cleans up its Wiki data (#3109).
• FIXED: The agent ends the turn when the completion cap truncates an answer (#3449); failed tool executions are reported as tool_result rather than error; MCP answers no longer cite unrelated knowledge-base sources (#3144); @Skill / @MCP pickers list only the resources of the agent that will run (#3475).
• FIXED: Updating an agent's config no longer erases its avatar, and driver errors stay server-side (#3230).
• FIXED: Remote embedders, rerankers and four vendor chat request shapes now match the vendors' documentation (#3484, #3488, #3489, #3500, #3513).
• FIXED: Queue claims order by failure count so retries cannot starve new work (#3444).
• FIXED: Invalid QA queries are rejected before the session lookup (#3477); suggested follow-ups no longer echo the current question (#2582); images from organization-shared knowledge bases load through the message file proxy (#3069).
• FIXED: Long-term memory ranks recalled items by the query and preserves extraction progress (#3224).
• FIXED: IM replies finalize when agent QA returns; Yunzhijia recovers stale WebSocket connections; Feishu keeps images in post messages and keeps streams alive during long answers; clearing an IM channel's knowledge base is persisted.
• FIXED: Multipart upload temp files are cleaned up; XMind files are accepted before parser discovery; the upload timeout scales with payload size; a non-ASCII base64 image no longer fails the whole parse.

Breaking Changes

• BREAKING: DingTalk channels support Stream mode only. Migration 000096 switches existing webhook channels to websocket; enable Stream for the app in the DingTalk developer console.
• BREAKING: Sandbox commands run as root by default (previously uid 1000) (#3067). Each session owns its sandbox, so isolation stays at the container boundary.
• BREAKING: MCP service updates that include usage_instructions must send non-blank text of at most 16,000 characters; omitting the field keeps the existing instructions.
• BREAKING: .env.example no longer ships default JWT_SECRET / SYSTEM_AES_KEY values and adds SYSTEM_SIGNING_KEY. New deployments must generate them; existing deployments must keep their current SYSTEM_AES_KEY, or stored credentials cannot be decrypted. An unset JWT_SECRET falls back to a random per-process value, so logins do not survive restarts and replicas must share it.
• BREAKING: Organization invites resolve an exact workspace ID (/organizations/{id}/search-tenants); searching workspaces by name is removed.
• BREAKING: Go SDK UpdateAgentRequest.Avatar is now *string: nil keeps the avatar, an empty string clears it.

Infrastructure & Build

• BUILD: Migrations 000091–000110 (MCP tool toggles and catalogs, browser authorization, memory consistency and vector search, DingTalk Stream-only, session fork and snapshot leases, ParadeDB pg_search 0.22.6, chunk index diet, knowledge profiles, MCP endpoints, message artifacts, skill served version, context checkpoints, message index, session sandbox tenant and host workspace dir, IM channel locale); SQLite migrations 000013–000030. Migration 000106 builds an index CONCURRENTLY; if interrupted, drop the invalid index and rerun.
• BUILD: ParadeDB image upgraded to run on CPUs without AVX2; the MinIO image is pulled from quay.io.
• BUILD: The frontend image builds in a multi-stage Dockerfile without host npm (#3074); the app image bundles BrowserSkill (scripts/build_browserskill.sh, pinned by scripts/browserskill-release.json); sandbox image targets sandbox, cube, desktop and desktop-cube.
• BUILD: Model catalog generation and checks (make model-catalog-generate, model-catalog-check, model-catalog-diff); per-vendor golden request tests.
• BUILD: The GPL document converter and vendored archives are removed; third-party license sources are bundled; Helm exposes global.maxFileSizeMB.

Documentation

• DOC: The official site and documentation are unified under website-docs/ — a product homepage plus the docs, built and deployed together as static files or a Docker image (#3297). Valid content from the old docs/ is migrated and the outdated pages are removed (#3580).
• DOC: Documentation synced for v0.8.2: local browser, built-in MCP Server, toolbox, conversation control, artifacts, sandbox terminal/desktop and Lite host sandbox, model catalog and protocol-compat fields, Confluence / DingTalk sources, new configuration variables, and migrations through 000110.
• DOC: Removed unsupported Homebrew installation guidance and the claim that MySQL is a supported DB_DRIVER.

发布日期：2026-09-24

v0.8.2 让智能体走出对话框：它可以操作你电脑上的浏览器；知识库可以作为 MCP Server 发布给 Claude、Cursor 等其他 AI 工具；对话过程也更可控，可以中途补充要求、分叉或回滚。沙箱新增交互终端和图形桌面，模型接入改为按厂商目录自动配置，另有 Confluence、钉钉文档数据源和日语界面等更新。

从 v0.8.0 升级前，请先阅读 升级须知 。完整的变更列表见 CHANGELOG 。

本机浏览器 ​

智能推理对话可以直接操作你电脑上的 Chrome 或 Edge。浏览器通过腾讯开源的 BrowserSkill 扩展连接到 WeKnora，智能体在独立的任务窗口中打开网页、点击、填写表单、读取内容，再把结果汇报到对话里。

截图待补充

智能体查询 GitHub 最新 Release 后填写订单表单，对话右下角是任务预览 website-docs/public/screenshots/local-browser-task.png 智能体查询 GitHub 最新 Release 后填写订单表单，对话右下角是任务预览

• 对话中可控 ：任务预览显示在对话里，可以随时暂停、继续或结束，也可以跳到浏览器查看对应页面。

• 需要人时交给人 ：遇到登录、验证码或需要授权的步骤，智能体会发起人工协助；你在浏览器里处理完，它再接着执行。

• 一次配对，多处可用 ：在「工具箱 → 浏览器连接」复制配对链接，粘贴到扩展里即可。同一空间内的多个对话共用这台浏览器的授权，开启输入框的「本机浏览器」开关后才会使用。

截图待补充

工具箱 → 浏览器连接：扩展版本、连接状态与智能体可执行的网页操作 website-docs/public/screenshots/browser-connection.png 工具箱 → 浏览器连接：扩展版本、连接状态与智能体可执行的网页操作 Docker 部署的 app 镜像已内置 BrowserSkill 服务和配套扩展，通常无需额外配置；多副本部署需要配置节点互通地址和共享密钥。详见 本机浏览器 。

内置 MCP Server ​

每个空间都可以把知识库发布为 MCP 服务，供 Claude Desktop、Cursor、Claude Code、VS Code 等支持 MCP 的客户端直接检索和提问，无需再单独部署 MCP 进程。

截图待补充

设置 → 发布集成 → MCP Server：端点地址与客户端配置 website-docs/public/screenshots/mcp-server-endpoint.png 设置 → 发布集成 → MCP Server：端点地址与客户端配置

• 在「设置 → 发布集成 → MCP Server」新建端点，选择可访问的知识库和要开放的工具。

• 工具分四组：检索与阅读、问答（运行端点指定的智能体）、Wiki、写入；写入工具默认关闭。

• 每个端点有独立的令牌和调用频率上限（默认每分钟 60 次），令牌可以随时轮换。

• 页面直接给出 Cursor、VS Code、Claude Desktop 的配置片段和 Claude Code 的一行命令。

原来仓库中的 Python 版 mcp-server/ 已标记为弃用。详见 MCP 集成 。

对话控制 ​

截图待补充

回答进行中，输入框上方排队等待的补充要求 website-docs/public/screenshots/chat-steer-queue.png 回答进行中，输入框上方排队等待的补充要求

• 中途补充要求 ：智能体还在执行时就可以继续输入。按 Enter 排队，当前步骤结束后处理；按 ⌘Enter / Alt+Enter 立即插入当前任务。

• 分叉与回滚 ：可以从任意一次历史提问分叉出新对话，也可以原地回滚到某次提问；使用沙箱的对话会同时恢复对应时刻的工作区。

• 思考强度 ：每个对话可以单独选择模型的思考强度。

• 截断提示 ：回答因达到输出上限被截断时会明确标出，不再悄悄截断。

截图待补充

侧栏「产物」页：跨对话汇总智能体生成的文件 website-docs/public/screenshots/artifacts-library.png 侧栏「产物」页：跨对话汇总智能体生成的文件 智能体生成的文件统一收在侧栏的「产物」页，可以按类型筛选、搜索、查看历史版本，也可以删除不需要的文件。把回答存入知识库时，编辑器改为分屏的 Markdown 源码与预览，并支持暂存草稿。

截图待补充

把回答添加到知识库：分屏编辑与预览 website-docs/public/screenshots/chat-save-answer-editor.png 把回答添加到知识库：分屏编辑与预览 详见 会话体验 。

沙箱：图形桌面、交互终端与本机沙箱 ​

对话右侧的「沙箱可视化」面板新增「桌面」和「终端」两个页签，可以直接进入智能体正在使用的同一个沙箱。

截图待补充

图形桌面：通过浏览器操作沙箱内的桌面环境 website-docs/public/screenshots/sandbox-desktop.png 图形桌面：通过浏览器操作沙箱内的桌面环境

截图待补充

交互终端：查看工作区文件与命令结果 website-docs/public/screenshots/sandbox-panel-terminal.png 交互终端：查看工作区文件与命令结果

• 图形桌面 ：需要 Cube 或 E2B 的桌面模板，并在沙箱配置中开启桌面。

• 交互终端 ：Cube 和 E2B 沙箱可用，刷新页面后可以重新接上仍在运行的终端。

• 打开面板不会创建或唤醒沙箱；长时间无操作会自动断开，默认 900 秒。

• 智能体执行命令、安装技能时，工具卡片会实时显示输出。

Lite 桌面版本机沙箱（macOS） ：没有选择远程沙箱的对话，会在系统沙箱中执行命令。新对话可以绑定电脑上的一个项目文件夹，智能体只能写入这个文件夹；命令不能联网， .git 目录只读，凭据目录始终不可访问。

工具箱 ：技能、MCP 服务和浏览器连接从「设置」移到侧栏的「工具箱」，旧的设置入口会自动跳转。

截图待补充

侧栏工具箱：技能管理、MCP 服务与浏览器连接 website-docs/public/screenshots/toolbox.png 侧栏工具箱：技能管理、MCP 服务与浏览器连接 详见 技能目录与沙箱 。

模型接入 ​

模型接入改为按「协议、厂商、模型目录」三层组织，内置 27 个厂商。添加模型时：

• 从厂商目录中选择模型后，会自动填入上下文窗口、最大输出、思考强度和视觉能力。

• 「实际调用方式」面板直接显示这个模型将使用的请求协议、请求地址和可选的思考强度。

• 个别接口与默认值不一致时，可以在高级选项中覆盖协议，或填写协议兼容 JSON。

• 语音识别覆盖所有在文档中提供该能力的厂商。

截图待补充

添加模型：从厂商目录选择模型并查看实际调用方式 website-docs/public/screenshots/model-editor-catalog.png 添加模型：从厂商目录选择模型并查看实际调用方式 详见 模型管理 。

知识库与数据源 ​

• Confluence 数据源 ：支持 Cloud 和 Server / Data Center，按空间同步已发布页面，增量更新，并带删除保护。

• 钉钉文档数据源 ；腾讯 IMA 同步时会遍历子文件夹。

• 上传与解析 ：上传面板显示整体进度，可以取消、重试；解析时间线展示各阶段耗时和失败原因；长时间没有进展的文档会标记为「排队中」或「疑似卡住」。

• 列表与下载 ：文档和资源列表支持排序；可以批量打包下载原始文件。

• 其他 ：上传时可以关闭文档摘要；知识库描述可以根据文档画像自动生成；FAQ 可以按启用状态筛选；活动记录会显示操作所用的 API Key 名称。

截图待补充

上传任务面板：整体进度与单个文件的取消、重试 website-docs/public/screenshots/kb-upload-tasks.png 上传任务面板：整体进度与单个文件的取消、重试

截图待补充

解析时间线：各阶段耗时、失败原因与重试 website-docs/public/screenshots/kb-parse-timeline.png 解析时间线：各阶段耗时、失败原因与重试

截图待补充

数据源：配置 Confluence 并选择要同步的空间 website-docs/public/screenshots/datasource-confluence.png 数据源：配置 Confluence 并选择要同步的空间 详见 知识库 与 数据源导入 。

集成、界面与安全 ​

• 联网搜索 ：新增博查和 Serply；Brave 与 Serply 支持按地区和时效过滤。

• IM ：每个渠道可以单独设置回复语言。

• 日语界面 ：界面新增日语，可在界面中切换，也可以用 DEFAULT_LOCALE 设为默认。

• CLI 配置页 ：「设置 → 发布集成 → CLI」按三步引导完成安装和登录。

• 管理员创建用户 ：系统管理员可以在控制台直接创建账号，生成一次性密码并复制账号信息。

• 仅白名单出站 ：设置 SSRF_DNS_WHITELIST_ONLY=true 后，不在 SSRF_WHITELIST 中的域名在 DNS 解析之前就会被拒绝，适合离线或严格管控的私有化部署。

截图待补充

发布集成 → CLI：三步完成安装与登录 website-docs/public/screenshots/integration-cli.png 发布集成 → CLI：三步完成安装与登录

升级须知 ​

变化 需要做什么

.env.example 不再提供 JWT_SECRET 、 SYSTEM_AES_KEY 的默认值，并新增 SYSTEM_SIGNING_KEY 新部署需要自行生成。 已有部署必须保留原来的 SYSTEM_AES_KEY ，否则已保存的凭据无法解密。未设置 JWT_SECRET 时每个进程会随机生成，重启后登录失效，多副本之间也无法共用

钉钉 IM 渠道只支持 Stream 模式 迁移 000096 会把 webhook 渠道自动改为 Stream；需要在钉钉开发者后台为应用开启 Stream 模式

沙箱内命令默认以 root 执行（v0.8.0 为 uid 1000） 每个对话独占一个沙箱，隔离边界仍是容器；依赖普通用户身份的脚本需要自行调整

更新 MCP 服务时如果传了 usage_instructions ，必须是非空文本且不超过 16000 字符 不传该字段则保留原有说明

组织邀请只能按工作空间 ID 精确查找 不再支持按名称搜索工作空间

Go SDK 的 UpdateAgentRequest.Avatar 改为 *string nil 表示保留头像，空字符串表示清除

数据库迁移：PostgreSQL 为 000091 – 000110 ，SQLite 为 000013 – 000030 。其中 000106 以 CONCURRENTLY 方式创建索引，如果中途中断，需要先删除无效索引再重新执行迁移，参见 常见问题与升级排障 。ParadeDB 镜像已升级，可以运行在不支持 AVX2 的 CPU 上；MinIO 镜像改为从 quay.io 拉取。

其他改进与修复 ​

• 智能体 ：检索工具合并为 search_knowledge 、 read_document 、 list_documents ，旧工具名自动映射，无需迁移配置；对话历史按模型上下文窗口裁剪，压缩检查点跨轮保留；MCP 工具改为按需发现和调用，服务的工具目录会持久保存，并可以单独启用或停用某个工具。

• 技能 ：旧版本技能升级时不再回滚目录，也不会中途下线；安装后会校验运行时依赖。

• Wiki ：整页重写被截断时会续写；同名条目跨类型复用页面；引用的分块会附带图片说明。

• 检索与解析 ：Milvus 支持多语言 BM25；Neo4j 查询增加上限保护；PaddleOCR-VL 请求超时可配置；PDF 在链接 anydoc 时仍默认使用内置解析器；内联 HTML 表格不再产生超长分块。

• 稳定性 ：修复文档卡在「处理中」或「删除中」的多种情况；关闭共享空间中跨工作区访问的路径；已删除租户的 Wiki 任务不再调用模型；向量、重排和部分对话厂商的请求格式已与其官方文档对齐。

完整列表见 CHANGELOG 。