Node Office conversion with independently packaged engines
独立打包引擎的 Node Office 转换
Binary Office and OOXML files require document layout before they can be previewed. Conversion must stay on the device without opening an Office application or adding source bytes to a model conversation. Native engines need platform-specific distribution, while browser conversion duplicates font transport, worker ownership, and resource limits across the Host and Client.
English
Problem
Binary Office and OOXML files require document layout before they can be previewed. Conversion must stay on the device without opening an Office application or adding source bytes to a model conversation. Native engines need platform-specific distribution, while browser conversion duplicates font transport, worker ownership, and resource limits across the Host and Client.
Decision
The office-to-pdf capability delegates conversion to the independently released @deepseek-ai/libreoffice-kit Node API. The kit ownership decision owns source maintenance, compatibility versions, and npm distribution. DSH owns Session file authorization, conversion concurrency, private scratch files, output limits, and Remote transport. The Web bundle declares the conversion service and shared Document Preview entry with stable IDs. The service owns authorized conversion Remote methods, whose descriptor is mounted by api/remotes; Office UI shares the Document Preview Loader lifetime.
The platform engine decision requires the kit’s declared native target engine, or WASM when no native target is declared. Missing or invalid required engines reject conversion. The shared bounded provider owns admission, conversion reuse, and cancellation through scratch cleanup. Preview consumes that provider without registering another converter or requiring Office authoring skills.
The service and Remote methods accept DOC, DOCX, XLS, XLSX, PPT, and PPTX. Client PDF previews use this path for Word and PowerPoint; browser spreadsheet preview reads XLSX, XLS, CSV, and TSV independently. The kit verifies bounded ZIP membership and content types for OOXML inputs and the OLE compound-file header for binary Office inputs before LibreOffice imports them. Renaming text to an Office suffix does not admit it. Binary formats return no missing-font diagnostics because the kit does not extract their font tables. It writes a fresh, exclusively created PDF in a caller-owned private directory. DSH reads and validates the complete output before removing scratch files. The service’s Remote method authorizes source access through Workspace Files, preserves the source path/version, and returns PDF bytes. Source read caps and generated PDF caps remain independent. One source path/version snapshot governs the access probe and deferred read, including size-failure rechecks, so conversion cannot publish bytes under another identity. Preview bytes do not enter Session storage or persistent caches.
A converter reuses font metadata obtained by its first conversion Worker; original font buffers and decoded glyph coverage remain conversion-local. Workers validate indexed files when reading them. Exact installed families precede configured alternatives, and complete family/style/weight/italic/width/pitch/language/code-point requests retain their distinct matches. WASM callbacks import original font files, including complete collections, into MEMFS. Native engines also retain their platform's font discovery. Neither path downloads or installs fonts; native OS-managed font memory is outside the explicit import budget. Recreating the converter refreshes its metadata after font changes.
The kit owns default serif, sans-serif, and monospace preference groups, including Chinese text families. Missing Chinese glyphs in Western text try the corresponding common text families before searching the remaining catalog, which avoids choosing a handwriting face solely because its file sorts first. The provider's optional fontFallbacks replaces these ordered groups without duplicating their defaults. Preferences preserve exact installed fonts and retain other covering fonts as a last resort; they are not a font whitelist. The native adapter writes missing-family choices into its private VCL profile. Installed metric-compatible fonts can resolve before that table, and platform glyph fallback remains available. Native platform selection and whole collection imports require inspecting the fonts actually used in exported PDFs.
DSH exports raster images at configurable resolution, defaulting to 192 DPI for the shared PDF canvas's 96 CSS DPI at device-pixel ratio 2. Text and vectors remain scalable; explicit bookmark export preserves the engine default when JSON filter options replace it. Node WASM downscales images with LibreOffice's CPU filter. Native conversion uses its separate platform engine.
The Office viewer lives under Document Preview’s client/office/ directory, alongside the loading lifecycle, PDF body, and reader types it uses. Keeping these components in one package removes an independent UI boot entry without creating cross-plugin runtime imports. Its bounded cache validates authorized source metadata, shares pending conversions between readers, cancels only when the final reader leaves, excludes failures, and clears on connection reset. Conversion starts when a user opens a preview. Missing declared font families accompany the PDF and are available from the document toolbar’s warning icon; font-table inventories and unrelated engine defaults are not warnings. The shared preview entry’s office cache settings use the existing page-global injection channel because the module boot graph carries package identities, not Loader configuration. Reloading the page adopts updated YAML values.
Office responses use Typert binary-result projection and Connection multipart framing. The Client receives one ArrayBuffer-backed Uint8Array without a base64 string or a separate decode buffer. The built browser scenario verifies the multipart attachment and PDF Worker together. Cache byte limits do not bound transient transport or viewer memory.
The kit ownership decision defines npm distribution and bundled offline conversion.
Desktop installs the kit through its existing target-Node pnpm dependency installation and retains the complete dependency tree. Worker paths and executable permissions remain ordinary package files. The Desktop build guide owns target selection and packaging; each signed application requires qualification on its target platform.
The Python executable distribution keeps the kit, target engine, and their dependency closure beside the executable. Its installed-wheel smoke relocates the payload, requires exactly the target backend, and converts DOCX once through that engine. Platform packaging and publication constraints belong to the platform engine decision.
The notices gate permits only the exact API and engine package names at MPL-2.0 and continues to reject unrelated MPL or changed non-permissive terms. Each recipient must retain access to the kit’s corresponding LibreOffice source pin, patches, build instructions, and license notices; the engine packages retain their third-party notices. MPL source availability applies when distributing covered executables outside the organization.
The browser-only preview replaces the kit entry with an unavailable converter and excludes its engine dependency tree from the VFS image. Keeping the Host provider loadable preserves the ordinary Office error presentation without shipping Node Workers, native helpers, or WASM conversion resources to the browser.
Alternatives considered
Keep conversion in browser Workers. This requires shipping the engine to the Client, browser font RPC, and shared-memory response headers. Node already owns authorized disk access and can serve the resulting PDF to every Client.
Use the installed soffice CLI or automate Microsoft Office. Executable discovery and ambient versions weaken reproducibility; Office GUI automation additionally changes focus and requires application permissions. The distributed helper owns a fixed engine without requiring either application installation.
Load a native addon in the Host process. A parser crash or synchronous stall would affect the Host. The separate native helper provides an independently terminable lifetime; its disk exchange also works with the WASM adapter.
Fall back after any native error. Retrying a damaged package or failed document with another engine hides release defects, duplicates work, and makes output depend on failure timing. Platform selection requires the kit’s declared native target engine, or WASM when no native target is declared; it does not retry native failures with WASM.
Return a temporary PDF path, or rasterize pages to PNG. The Client needs PDF bytes for its existing viewer and selectable text. Exposing temporary paths would add authorization and lease ownership; PNG would create another rendering pipeline and remove PDF controls.
Index every font for every render, or cache only family names. Repeated indexing dominates small-document work, while family-only keys lose style and glyph distinctions. Shared metadata and complete per-request keys remove repeated parsing without retaining font buffers or introducing a filesystem watcher.
Compile at install time, download engines or fonts at runtime, or use an online converter. These add toolchain or network requirements and may move private content off device. Prebuilt tarballs keep installation and conversion independent of those operations.
Keep an independent Office UI package. Its reader, cache, and font notice have the same preview lifecycle and PDF presentation consumers. A separate plugin adds a package, boot wiring, and a Loader switch without an independently evolving UI responsibility. Host conversion and Remote authorization retain their separate plugins.
Remove the shared Client conversion cache. Per-tab state cannot share pending conversion or retained PDFs across readers. A bounded cache avoids that repeated work while preserving per-reader cancellation and authorized version checks.
Add a model-facing rendering tool or durable preview events. Preview provides no model input. Such a tool would require logged facts and both SDK projections and remains a separate consumer.
Consequences
Native and WASM fidelity still depends on the build, source formatting, installed fonts, and platform font discovery. A missing glyph cannot be recovered without a covering font. Image resolution limits do not cap image decoding or total process memory. WASM retains bounded memory growth and checked stack space for large collections and CFF fonts; a fatal runtime abort prevents subsequent C++ cleanup calls. Macros and document-link updates are disabled by supported LOKit options and the pinned source patch; this does not constitute an OS sandbox.
The provider tests, Loader composition, and browser scenario own DSH lifecycle, authorization, and presentation evidence. Engine qualification additionally requires real DOC/DOCX/XLS/XLSX/PPT/PPTX conversion, external PDF text/font/page/image inspection, relocation and corrupted-package rejection, and same-input native/WASM performance samples. Mock helpers and microbenchmarks do not establish these outcomes. Each target's real builder and Desktop package need independent qualification; a successful local architecture does not prove the full matrix.
中文
问题
二进制 Office 和 OOXML 文件需要经过文档排版才能预览。转换必须留在设备上，不能打开 Office 应用，也不能将源字节加入模型对话。原生引擎需要按平台分发，而浏览器转换会在 Host 和 Client 两侧重复字体传输、worker 生命周期与资源限额管理。
决策
文档渲染能力将转换委托给独立发布的 @deepseek-ai/libreoffice-kit Node API。kit 归属决策负责源码维护、兼容版本和 npm 分发。DSH 负责 Session 文件授权、转换并发、私有临时文件、输出限制和 Remote 传输。Web bundle使用稳定 ID 声明转换服务与共享文档预览入口。服务负责带授权的转换 Remote 方法，其描述符由 api/remotes 挂载；Office UI 共享文档预览的 Loader 生命周期。
平台引擎决策要求使用 kit 已声明的原生目标引擎，未声明原生目标时使用 WASM。缺失或无效的必需引擎会拒绝转换。共享的有界提供方负责准入、转换复用以及持续到临时文件清理完成的取消。预览消费该提供方，不注册另一个转换器，也不依赖 Office 创作 skills。
服务和 Remote 方法接受 DOC、DOCX、XLS、XLSX、PPT 和 PPTX。Client PDF 预览通过此路径显示 Word 和 PowerPoint；浏览器表格预览独立读取 XLSX、XLS、CSV 和 TSV。LibreOffice 导入前，kit 校验 OOXML 输入的有界 ZIP 成员和内容类型，以及二进制 Office 输入的 OLE 复合文件头。将文本改为 Office 后缀不能通过校验。kit 不提取二进制格式的字体表，因此这些格式不返回缺失字体诊断。kit 在调用方拥有的私有目录中独占创建新的 PDF。DSH 读取并校验完整输出后才删除临时文件。服务的 Remote 方法通过 Workspace Files授权源文件访问，保留源路径和版本，并返回 PDF 字节。源文件读取上限与生成 PDF 上限相互独立。读取权限探测和延迟读取（包括超限失败后的复查）采用同一个源路径／版本快照，防止转换将字节发布到另一个源身份下。预览字节不会进入 Session 存储或持久缓存。
converter 复用首个转换 Worker 返回的字体元数据；原始字体缓冲区和解码后的字符覆盖范围仍只属于单次转换。Worker 读取字体时校验索引中的文件。已安装字体族的精确匹配优先于配置的替代字体，完整的字体族、样式、字重、斜体、宽度、字距、语言与码点请求保留各自的匹配结果。WASM 回调将包含完整字体集合的原始字体文件导入 MEMFS。原生引擎还保留各平台的字体发现能力。两条路径均不下载或安装字体；原生操作系统管理的字体内存不受显式导入预算约束。字体变化后，重新创建 converter 会刷新元数据。
kit 维护 serif、sans-serif 和 monospace 的默认优先组，其中包含中文正文字体。西文文本缺少中文字形时，先尝试同类的常用正文字体，再搜索其余字体目录，避免仅因文件排序靠前而选用手写体。provider 的可选 fontFallbacks 替换这些有序组，不重复维护默认值。优先规则保留已安装原字体的精确匹配，并以其他覆盖字体作为最后兜底；它们不是字体白名单。原生适配器将缺失字体的选择写入私有 VCL profile。已安装的度量兼容字体可能在查询该表前被选中，平台的字形回退仍然可用。原生平台的选择及完整字体集合的导入要求检查导出 PDF 实际使用的字体。
DSH 按可配置分辨率导出栅格图片，默认 192 DPI，对应共享 PDF 画布在设备像素比 2 时的 96 CSS DPI。文本与矢量仍可缩放；JSON 过滤选项替代隐式选项时，显式书签导出保留引擎默认行为。Node WASM 使用 LibreOffice 的 CPU 过滤器降采样图片。原生转换使用独立的平台引擎。
Office 查看器位于文档预览的 client/office/ 目录，与其使用的加载生命周期、PDF 正文和读取器类型同属一个包。这些组件放在同一包中，既减少一个独立 UI 启动入口，也无需跨插件运行时导入。其有界缓存校验已授权的源元数据，在读取方之间共享待完成转换，仅在最后一个读取方离开时取消，不缓存失败，并在连接重置时清空。用户打开预览时才开始转换。缺失的已声明字体族随 PDF 返回，可通过文档工具栏的警告图标查看；字体表清单与无关的引擎默认字体不构成警告。共享预览入口的 office 缓存设置复用页面全局注入通道，因为模块启动图携带包标识而不传递 Loader 配置。重新加载页面后采用更新的 YAML 值。
Office 响应使用 Typert 二进制结果投影和 Connection multipart 封装。Client 直接收到一个由 ArrayBuffer 支撑的 Uint8Array，不会生成 base64 字符串或单独的解码缓冲区。构建后的浏览器场景会同时验证 multipart 附件和 PDF Worker。缓存字节限制不约束临时传输或查看器内存。
kit 归属决策定义 npm 分发和随应用打包的离线转换。
Desktop 通过现有的目标 Node pnpm 依赖安装流程安装 kit，并保留完整依赖树。Worker 路径和可执行权限仍由普通包文件承载。Desktop 构建指南 负责目标选择与打包；每个签名应用仍需在目标平台验收。
Python 可执行分发将 kit、目标引擎及其依赖闭包保留在可执行文件旁。安装后的 wheel 冒烟测试会迁移载荷，要求仅存在目标后端，并通过该引擎转换一次 DOCX。各平台的打包与发布限制由平台引擎决策说明。
声明检查仅放行精确的 API 与引擎包名及 MPL-2.0 条款，继续拒绝无关 MPL 包或变更后的非宽松条款。每位接收者都必须保有访问 kit 对应 LibreOffice 源码版本、补丁、构建说明与许可证声明的权限；引擎包保留各自的第三方声明。向组织外部分发受覆盖的可执行文件时，须满足 MPL 源码可用性要求。
纯浏览器预览用返回不可用错误的转换器替代 kit 入口，并从 VFS 镜像排除其引擎依赖树。保留 Host provider 的可加载性，可以沿用正常的 Office 错误展示，而无需向浏览器分发 Node Worker、原生辅助程序或 WASM 转换资源。
考虑过的替代方案
保留浏览器 Worker 转换。 这需要向 Client 分发引擎、提供浏览器字体 RPC 和配置共享内存响应头。Node 已经拥有授权磁盘访问能力，可以向所有 Client 提供转换后的 PDF。
使用已安装的 soffice CLI 或自动化 Microsoft Office。 可执行文件发现和环境中的版本削弱可复现性；Office GUI 自动化还会改变焦点并需要应用权限。随包辅助进程拥有固定引擎，不要求安装这两类应用。
在 Host 进程中加载原生 addon。 解析器崩溃或同步阻塞会影响 Host。独立的原生辅助进程具有可单独终止的生命周期，其磁盘交换方式也适用于 WASM 适配层。
任何原生错误都触发回退。 使用另一引擎重试损坏的包或失败文档会隐藏发布缺陷、重复计算，并使输出取决于失败时机。平台选择要求使用 kit 已声明的原生目标引擎，未声明原生目标时使用 WASM；原生失败后不会使用 WASM 重试。
返回临时 PDF 路径，或将页面栅格化为 PNG。 Client 需要 PDF 字节来使用现有查看器和可选择文本。暴露临时路径会增加授权与租约管理；PNG 则会建立另一条渲染管线并丢失 PDF 控件。
每次渲染重新索引所有字体，或只按字体族缓存。 重复索引主导小文档成本，而仅以字体族为键会丢失样式和字形差异。共享元数据与完整请求键减少重复解析，无需保留字体缓冲区或引入文件系统监听器。
安装时编译、运行时下载引擎或字体，或使用在线转换器。 这些方案增加工具链或网络要求，并可能将私有内容移出设备。预构建 tarball 使安装与转换不依赖这些操作。
保留独立的 Office UI 包。 其读取器、缓存和字体提示共享预览生命周期与 PDF 展示消费者。独立插件增加包、启动接线和 Loader 开关，却没有独立演进的 UI 职责。Host 转换与 Remote 授权仍保留独立插件。
移除共享的 Client 转换缓存。 tab 内状态无法在读取方之间共享进行中的转换或已保留的 PDF。有界缓存减少这些重复工作，同时保留各读取方的取消和已授权版本检查。
添加面向模型的渲染工具或持久预览事件。 预览不会提供模型输入。这样的工具需要日志事实及两套 SDK 投影，仍属于独立消费者。
后果
原生与 WASM 的保真度仍取决于构建、源文件格式、已安装字体及平台字体发现。没有覆盖字体就无法恢复缺失字形。图片分辨率限额不限制图片解码或总进程内存。WASM 为大型字体集合和 CFF 字体保留有界内存增长与受检查的栈空间；致命运行时中止会阻止后续 C++ 清理调用。宏与文档链接更新由实际支持的 LOKit 选项和固定源码补丁禁用；这不构成操作系统沙箱。
提供方测试、Loader 组合和浏览器场景负责 DSH 生命周期、授权与展示证据。引擎验收还需要真实 DOC/DOCX/XLS/XLSX/PPT/PPTX 转换、外部 PDF 文本、字体、页数与图片检查、迁移安装和损坏包拒绝，以及同输入的原生/WASM 性能样本。模拟辅助进程和微基准不能证明这些结果。各目标的真实构建机与 Desktop 安装包需要独立验收；一个本地架构成功不能证明整个矩阵。