Repository navigation
Releases
BabelDOC Desktop plugin v0.2.4
English
This release adds conservative list-item boundaries and makes PDF object protection more resilient. It updates the optional BabelDOC plugin and Python package, not the Wenyi Desktop application itself.
New features
- Structured list items: Repeated unknown leading glyphs can establish explicit item rows when marker identity, alignment, spacing and continuation-line evidence agree. The bridge preserves the original marker glyphs and returns row boundaries for independent translation and fillback; it does not guess missing Unicode mappings or substitute bullets.
Fixes
- Whole-glyph protection: Mathematical protection follows complete original glyph spans, including characters expanded by Unicode normalization. Protection boundaries no longer split a glyph between frozen objects and translatable source text.
- Recoverable protection failures: If source mapping or object protection is ambiguous, extraction retains that paragraph's original glyphs and layout, excludes it from translation and continues with other paragraphs without renumbering their IDs. The same paragraph remains in the source language after fillback and session recovery. Diagnostics record paragraph ID, stage and reason in the service log and session
extraction_warnings.json, without logging paragraph text.
Upgrade notes
- Update Wenyi and the optional bridge together, then extract affected PDFs into fresh state to enable the new list boundaries and protection behavior. Existing frozen sessions are not migrated in place.
- Detection requires at least three matching unknown markers, aligned body starts and indented continuation evidence in multiple items. It is not general list, table or bibliography reconstruction. Translated continuation lines do not yet gain hanging indentation, and shorter translations retain the original item regions rather than compacting gaps.
- Inspect extraction warnings and final PDFs: paragraphs retained after a protection failure are intentionally untranslated. Synthetic PDF regressions do not establish model translation quality or pixel-identical layout preservation.
- Desktop users update the optional plugin in Settings. Python users upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip checkand restart the service. Publication does not update an installed plugin or running service. - BabelDOC remains pinned to 0.7.1, with the independent AGPL-3.0-only HTTP process and protocol 1 unchanged.
中文
本版新增保守的列表项边界识别,并增强 PDF 对象保护的容错能力。发布范围是可选 BabelDOC 插件与 Python 包,不是 Wenyi Desktop 主程序更新。
新增功能
- 结构化列表项: 当重复未知前导字形的标记身份、对齐、间距和续行证据一致时,bridge 可建立明确的条目行边界;保留原始标记字形,供逐项翻译与回填,不猜测缺失的 Unicode 映射,也不替换为项目符号。
问题修复
- 完整字形保护: 数学保护范围按完整原始字形确定,包括 Unicode 归一化后展开为多个字符的字形,不再将同一字形拆分到冻结对象和可翻译原文两侧。
- 保护失败可恢复: 原文映射或对象保护存在歧义时,抽取保留该段原始字形与版式,不将其送入翻译,并继续处理其他段落,不重新编号段落 ID。回填及恢复会话后,该段仍保留源语言内容。服务日志及会话内
extraction_warnings.json记录段落 ID、阶段和原因,不记录段落正文。
升级说明
- 同时更新 Wenyi 和可选 bridge,再将受影响 PDF 抽取到新状态中,才能启用新的列表边界与保护行为;已有冻结会话不原地迁移。
- 检测要求至少三个相同未知标记、对齐的正文起点,以及多个条目的缩进续行证据;不是通用列表、表格或参考文献重建。译文续行暂不增加悬挂缩进;短译文仍保留原条目区域,不自动压缩条目间空隙。
- 请检查抽取警告与最终 PDF:保护失败后保留原样的段落会有意不翻译。合成 PDF 回归不代表模型翻译质量评估,也不保证逐像素保留版面。
- Desktop 用户在设置中更新可选插件;Python 用户在专用环境中升级
wenyi-babeldoc-bridge,运行uv pip check后重启服务。发布本版不会自动更新已安装插件或正在运行的服务。 - BabelDOC 仍固定为 0.7.1;独立 AGPL-3.0-only HTTP 进程及 protocol 1 保持不变。
BabelDOC Desktop plugin v0.2.3
English
This release keeps detached mathematical accents and scripts with their expressions and preserves non-translatable numeric paragraphs. It updates the optional BabelDOC plugin and Python package, not the Wenyi Desktop application itself.
Improvements
- Explicit best-effort omission: Updated clients can explicitly omit an unresolved protected object when best-effort export is requested. Strict validation remains the default; incomplete placements are not silently accepted.
Fixes
- Detached mathematical marks: New extraction conservatively joins overhead accents to a unique mathematical base and small scripts in all four corners inside mathematical sequences. Expressions move as protected objects while retaining original glyphs, fonts and relative geometry. Ambiguity, prose, spatial gaps and drawing-context boundaries stop reassignment or expansion.
- Duplicate or displaced scripts: A page-level script transferred into a protected expression is no longer drawn separately. Detached accent runs and inferred wide spaces no longer incorrectly bind distant expressions together.
- Numeric paragraphs: Numeric-only, separator-only and vertical paragraphs remain in the original PDF instead of becoming translation inputs. The same selection rule applies when restoring frozen sessions.
Upgrade notes
- Update Wenyi and the optional BabelDOC plugin together, then import the PDF into fresh state to enable repaired mathematical grouping and extraction boundaries. Existing translation state and frozen sessions are not migrated in place.
- Desktop users update the plugin in Settings; Python users upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip checkand restart the service. Publication does not upgrade an installed plugin. - Recognition depends on native formula detection, mathematical character/font evidence and grouping boundaries. Check formula coverage and final layout; explicitly omitted objects can remove mathematical meaning. Synthetic PDF regressions do not establish real-model translation quality or pixel-identical preservation.
- BabelDOC remains pinned to 0.7.1. The service remains an independent AGPL-3.0-only process with protocol 1, Python 3.10–3.13 and PyMuPDF 1.x. Linux x64, Windows x64, macOS Intel and Apple Silicon retain separate runtimes and a shared model/font archive. Packages are not Windows code-signed or macOS notarized.
中文
本版修正游离数学重音与角标的归属,并保留不应翻译的数字段落。发布范围是可选 BabelDOC 插件和 Python 包,不是 Wenyi Desktop 主程序更新。
改进
- 显式尽力省略: 更新后的客户端可在请求尽力导出时,显式省略无法定位的受保护对象。默认仍执行严格校验,不会静默接受缺少对象定位的请求。
问题修复
- 游离数学符号: 新抽取保守地将正上方重音绑定到唯一的数学基字,并归并数学序列内部左上、左下、右上、右下的小号角标。表达式作为保护对象一起移动,保留原字形、字体和相对几何。遇到歧义、正文、空间间隔或绘制上下文边界时停止归属或扩展。
- 重复或错位角标: 从页面级转入保护表达式的角标不再单独绘制;游离重音组和推断出的宽空格不再将远处表达式错误绑定在一起。
- 数字段落: 纯数字、纯分隔符及竖排段落保留在原 PDF 中,不再成为翻译输入;恢复冻结会话时也使用同一筛选规则。
升级说明
- 同时更新文译和可选 BabelDOC 插件,再将 PDF 导入到新状态中,才能启用修正后的数学分组和抽取边界;已有译文状态及冻结会话不原地迁移。
- Desktop 用户在设置中更新插件;Python 用户在专用环境中升级
wenyi-babeldoc-bridge,运行uv pip check后重启服务。发布本版不会自动升级已安装的插件。 - 识别覆盖率仍取决于原生公式检测、数学字符/字体证据及合并边界,需检查公式覆盖和最终版面;显式省略对象可能损失数学含义。合成 PDF 回归不代表真实模型翻译质量评估,也不保证逐像素保留版面。
- BabelDOC 仍固定为 0.7.1;服务保持独立 AGPL-3.0-only 进程、protocol 1、Python 3.10–3.13 和 PyMuPDF 1.x。Linux x64、Windows x64、macOS Intel 和 Apple Silicon 继续提供独立运行包及共用模型/字体包。运行包未做 Windows 代码签名或 macOS 公证。
BabelDOC Desktop plugin v0.2.2
English
This release adds read-only formula reference text and corrects protected inline-formula baselines. It updates the optional BabelDOC plugin and Python package, not the Wenyi Desktop application itself.
Improvements
- Read-only formula references: Newly extracted protected formula annotations include
object_source, the object's extracted reading-order text. Wenyi v1.5.7 can use it as context across translation, polishing, Review and Autofix. Formula text stays separate from editable prose; frozen original glyphs, fonts and geometry remain the source for PDF output.
Fixes
- Inline-formula alignment: Before fillback, protected mathematical objects recover their baseline from neighboring prose in the frozen snapshot. Subscripts and tall operators no longer make the object's bounding-box bottom incorrectly act as the prose baseline. The correction preserves individual glyph geometry and also applies to existing protected-object snapshots.
Upgrade notes
- Update Wenyi to v1.5.7 and the optional bridge to v0.2.2, then reimport the PDF into fresh state to obtain formula reference text. Existing projects and frozen sessions are not migrated in place; older snapshots without
object_sourceretain object protection but cannot provide the new references. - Desktop users update the plugin in Settings. Python users upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip checkand restart the service. Publishing this release does not update an installed plugin automatically. - BabelDOC remains pinned to 0.7.1 and the service remains an independent AGPL-3.0-only process. Formula recognition and extracted text can be incomplete; check formula coverage and final layout. Read-only reference text is not used to reconstruct original glyphs and does not guarantee translation quality or pixel-identical output.
中文
本版增加公式只读参考文本,并修正受保护行内公式的基线。发布范围为可选 BabelDOC 插件与 Python 包,不是 Wenyi Desktop 主程序更新。
改进
- 公式只读参考: 新抽取的受保护公式标注提供
object_source,内容为按阅读顺序抽取的对象文本。Wenyi v1.5.7 可在翻译、润色、Review 与 Autofix 中将其作为上下文。公式文本与可编辑正文分离,PDF 输出仍以冻结的原始字形、字体和几何信息为准。
问题修复
- 行内公式对齐: 回填前从冻结快照中的相邻正文恢复受保护数学对象的基线,避免下标及高大运算符使对象包围框底部被误当作正文基线。修正保留对象内部字形几何,也适用于已有的受保护对象快照。
升级说明
- 将文译更新到 v1.5.7、可选 bridge 更新到 v0.2.2,再将 PDF 重新导入到新状态中,才能获得公式参考文本。已有项目与冻结会话不原地迁移;缺少
object_source的旧快照仍保留对象保护,但无法提供新增参考。 - Desktop 用户在设置中更新插件。Python 用户在专用环境中升级
wenyi-babeldoc-bridge,运行uv pip check后重启服务。发布本版不会自动更新已安装的插件。 - BabelDOC 仍固定为 0.7.1,服务保持独立 AGPL-3.0-only 进程。公式识别与抽取文本可能不完整,仍需检查公式覆盖及最终版面。只读参考文本不用于重建原始字形,也不保证翻译质量或逐像素一致的输出。
Full changelog / 完整变更: v0.2.1...v0.2.2
BabelDOC Desktop plugin v0.2.1
English
This release adds protected mathematical objects and improves PDF link styling
and directory-row preservation. It updates the optional BabelDOC plugin and Python
package, not the Wenyi Desktop application itself.
Improvements
- Original mathematical objects: New extraction groups confirmed formulas,
mathematical-font glyphs and operators with adjacent variables. The bridge
preserves their original glyphs, fonts and relative geometry; clients receive
opaque identities and insertion points rather than lossy formula transcriptions. - Validated PDF fillback: The
protected-objects-v1capability requires complete,
valid protected-object placements, including explicit order for coincident points.
Ordinary styles, links and legacy unprotected formulas remain best effort.
Fixes
- Link appearance: Precisely aligned links retain their original color and font
when an overlapping body-style range is too broad. Links are not uniformly
recolored blue. - Adjacent CJK prose: Mathematical protection keeps neighboring Chinese,
Japanese and Korean prose translatable even without intervening spaces. - Directories containing mathematics: Row offsets follow the same clean-source
mapping as annotations, preserving the original row boxes and order through
fillback and session restoration.
Upgrade notes
- Update Wenyi and the optional BabelDOC plugin together, then import the PDF into
fresh state to enable protected objects and corrected extraction boundaries.
Existing translation state and frozen sessions are not migrated in place. - Desktop users update the plugin in Settings; Python users upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip checkand
restart the service. Publication does not upgrade an installed plugin. - Recognition depends on native formula detection, mathematical character/font
evidence and grouping boundaries. Check formula coverage and final layout.
Synthetic PDF regressions do not establish real-model translation quality or
pixel-identical preservation. - BabelDOC remains pinned to 0.7.1. The service remains an independent AGPL-3.0-only
process with protocol 1, Python 3.10–3.13 and PyMuPDF 1.x. Linux x64, Windows x64,
macOS Intel and Apple Silicon retain separate runtimes and a shared model/font
archive. Packages are not Windows code-signed or macOS notarized.
中文
本版增加数学对象保护,并改进 PDF 链接样式和目录行保留。
发布范围是可选 BabelDOC 插件和 Python 包,不是 Wenyi Desktop 主程序更新。
改进
- 原始数学对象: 新抽取将确认的公式、数学字体字形及运算符与相邻变量组成完整对象。
bridge 保留原字形、字体和相对几何;客户端只接收不透明身份和插入点,
不接收有损的公式转写。 - 校验 PDF 回填:
protected-objects-v1能力要求受保护对象的位置完整有效,
同一插入点的对象须明确排序。普通样式、链接及旧会话中未受保护的公式仍采用尽力恢复。
问题修复
- 链接外观: 正文样式范围过宽并与链接重叠时,精确对齐的链接仍保留原始颜色和字体,
不将所有链接统一染蓝。 - 相邻中日韩正文: 数学保护不再吞掉没有空格分隔的相邻中文、日文和韩文正文,
这些文字仍可正常翻译。 - 含数学内容的目录: 行偏移与标注采用同一份清理后原文映射,
在回填和会话恢复后保留原始行框与顺序。
升级说明
- 同时更新文译和可选 BabelDOC 插件,再将 PDF 导入到新状态中,
才能启用对象保护和修正后的抽取边界;已有译文状态及冻结会话不原地迁移。 - Desktop 用户在设置中更新插件;Python 用户在专用环境中升级
wenyi-babeldoc-bridge,运行uv pip check后重启服务。
发布本版不会自动升级已安装的插件。 - 识别覆盖率仍取决于原生公式检测、数学字符/字体证据及合并边界,
需检查公式覆盖和最终版面。合成 PDF 回归不代表真实模型翻译质量评估,
也不保证逐像素保留版面。 - BabelDOC 仍固定为 0.7.1;服务保持独立 AGPL-3.0-only 进程、protocol 1、
Python 3.10–3.13 和 PyMuPDF 1.x。Linux x64、Windows x64、macOS Intel
和 Apple Silicon 继续提供独立运行包及共用模型/字体包。
运行包未做 Windows 代码签名或 macOS 公证。
BabelDOC Desktop plugin v0.2.0
English
This release upgrades the independent bridge to BabelDOC 0.7.1 and introduces
precompiled Python wheels. It includes the optional plugin and Python package,
not an update to the Wenyi Desktop application itself.
Native engine and Python wheels
- Pinned Rust stages: PDF parsing, paragraph/style analysis, typesetting and
writing use the Rust crates from the exact BabelDOC 0.7.1 revision through a
bridge-owned PyO3 extension. HTTP, annotation and directory-row policies remain
in Python; Wenyi still communicates with this separate AGPL process over HTTP. - Precompiled installation: Platform wheels target Linux x64, Windows x64,
macOS Intel and Apple Silicon with the CPython 3.10 stable ABI. Supported
Python versions remain 3.10–3.13. Installing a matching wheel does not require
Rust; source builds require Rust/Cargo and a C/C++ toolchain. - Restart-safe snapshots: Compressed native graph checkpoints preserve
shared objects, lazy graphics-state history and the parser's page-object map.
Each fillback loads a fresh graph, and recovery does not reload the layout model. - Preserved integration behavior: Stable paragraph IDs, selected-page exports,
aligned formulas/styles/links, directory rows and protected vertical glyphs
continue through the production pipeline. The HTTP schema and Desktop protocol
remain unchanged.
Upgrade notes
- Update the optional BabelDOC plugin from Desktop Settings; publication does
not automatically upgrade an installed plugin. Python users can upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip check,
and restart the service. - BabelDOC 0.5/0.6 sessions are incompatible with the new snapshot format.
Keep existing translation state and session backups. Finish old tasks in their
original environment, or extract the original PDF again with the new bridge.
Old pickle/JSON snapshots are rejected, not loaded or automatically migrated. - The bridge wheel does not contain a Python interpreter, fonts or models.
Desktop runtimes are self-contained and use a separate shared model/font ZIP.
Python installations resolve dependencies normally and use BabelDOC's asset cache. - Validation uses synthetic PDFs and local roundtrip checks, including CJK text,
formulas/links, directory rows, vertical glyphs and restart recovery. It does
not establish real-model translation quality or pixel-identical preservation. - Windows x64, Linux x64, macOS Apple Silicon and Intel have separate runtimes
and share the model/font archive. The service remains AGPL-3.0-only with
protocol 1 and PyMuPDF 1.x; packages are not Windows code-signed or macOS notarized.
中文
本版将独立 bridge 升级到 BabelDOC 0.7.1,并提供 Python 预编译 wheel。
发布范围包含可选插件和 Python 包,不是 Wenyi Desktop 主程序更新。
原生引擎与 Python wheel
- 固定版本的 Rust 流程: 通过 bridge 自有的 PyO3 扩展调用 BabelDOC 0.7.1
准确提交中的 Rust crate,负责 PDF 解析、段落/样式分析、排版和写出。
HTTP、注释和目录行策略仍由 Python 负责;Wenyi 仍仅通过 HTTP 访问独立 AGPL 进程。 - 预编译安装: Linux x64、Windows x64、macOS Intel 和 Apple Silicon wheel
使用 CPython 3.10 稳定 ABI,支持 Python 3.10–3.13。安装匹配的 wheel 不需要 Rust;
从源码构建需要 Rust/Cargo 和 C/C++ 工具链。 - 可恢复的原生快照: 压缩图快照保留共享对象、惰性图形状态历史和解析器页面对象映射。
每次回填加载全新副本,恢复会话时不重新加载版面模型。 - 保留集成行为: 生产路径继续支持稳定段落 ID、选页导出、公式/样式/链接对齐、
目录行和竖排字形保护;HTTP schema 与 Desktop protocol 保持不变。
升级说明
- 在 Desktop 设置中更新可选 BabelDOC 插件。发布本版不会自动升级已安装的插件。
Python 用户可在专用环境中升级wenyi-babeldoc-bridge,运行uv pip check后重启服务。 - BabelDOC 0.5/0.6 会话与新快照格式不兼容。 保留已有译文状态及会话备份;
在原环境完成旧任务,或使用新 bridge 重新抽取原始 PDF。
旧 pickle/JSON 快照会被拒绝,不会加载或自动迁移。 - bridge wheel 不包含 Python 解释器、字体或模型。Desktop 运行包自带运行环境,
另配共用模型/字体 ZIP;Python 安装按常规解析依赖并使用 BabelDOC 资源缓存。 - 验证使用合成 PDF 和本地回填检查,覆盖中文、公式/链接、目录行、竖排和重启恢复。
这些验证不代表真实模型翻译质量评估,也不保证逐像素保留版面。 - Windows x64、Linux x64、macOS Apple Silicon 和 Intel 分别提供运行包,共用模型/字体包。
服务仍采用 AGPL-3.0-only、protocol 1 和 PyMuPDF 1.x;
运行包未做 Windows 代码签名或 macOS 公证。
BabelDOC Desktop plugin v0.1.8
English
This release fixes rotated sidebar preservation in the independent BabelDOC
bridge. It includes the optional plugin and Python package, not an update to
the Wenyi Desktop application itself.
Rotated text preservation
- Skip before translation: Paragraphs marked
vertical=Trueby BabelDOC
are no longer sent for translation. Their original glyphs, fonts, sizes,
coordinates, rotation and links are preserved without reflow, regardless of
layout class. - Protect existing sessions: Previously extracted rotated paragraphs also
retain their frozen glyphs and links, even when saved translations reorder
their text or annotation placements are stale. Paragraph IDs and frozen
snapshots remain unchanged across repeated exports and session recovery. - Replace the text-equality shortcut: Preservation now follows BabelDOC's
vertical flag rather than requiring translation text and annotation positions
to match the extracted source. Nonvertical paragraphs use normal text and
annotation fillback, including when their text is unchanged.
Upgrade notes
- Update the optional BabelDOC plugin from Desktop Settings; publication does
not automatically upgrade an installed plugin. Python users can upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip check,
and restart the service. - Export the PDF again after updating to apply the protection to an existing
compatible session. Previously generated PDFs are not rewritten automatically.
Keep translation state and session backups; frozen-session recovery still
requires matching Python/BabelDOC versions. - Protection depends on BabelDOC's
verticalflag; it does not add a separate
rotation detector or translate vertical text. - Windows x64, Linux x64, macOS Apple Silicon and Intel have separate runtimes
and share the model/font archive. The service remains AGPL-3.0-only with
protocol 1 and PyMuPDF 1.x; packages are not Windows code-signed or macOS notarized.
中文
本版修复独立 BabelDOC bridge 中旋转侧栏文字的保留问题。
发布范围包含可选插件和 Python 包,不是 Wenyi Desktop 主程序更新。
旋转文字保护
- 翻译前跳过: BabelDOC 标记为
vertical=True的段落不再送往翻译器。
不论版面类别,均直接保留原始字形、字体、字号、坐标、旋转及链接,不重新排版。 - 已有会话同样受保护: 即使已保存的译文调整了侧栏文字顺序,或标注位置已经过期,
仍保留冻结快照中的字形和链接。重复导出及会话恢复不会改变段落 ID 和冻结快照。 - 取代文字相等判断: 保护依据改为 BabelDOC 的竖排标记,不再要求译文及标注位置
与抽取原文完全相同。非竖排段落统一走正常文字/标注回填,即使文字未发生变化。
升级说明
- 在 Desktop 设置中更新可选 BabelDOC 插件。发布本版不会自动升级已安装的插件。
Python 用户可在专用环境中升级wenyi-babeldoc-bridge,运行uv pip check后重启服务。 - 更新后重新导出 PDF,即可将保护应用于已有的兼容会话。已经生成的 PDF 不会自动改写。
保留译文状态和会话备份;冻结会话恢复仍要求匹配的 Python/BabelDOC 版本。 - 保护依赖 BabelDOC 的
vertical标记,不增加独立的旋转检测,也不提供竖排文字翻译。 - Windows x64、Linux x64、macOS Apple Silicon 和 Intel 分别提供运行包,共用模型/字体包。
服务仍采用 AGPL-3.0-only、protocol 1 和 PyMuPDF 1.x;
运行包未做 Windows 代码签名或 macOS 公证。
BabelDOC Desktop plugin v0.1.7
English
This patch release improves PDF contents-row fillback and preserves original glyphs when a paragraph does not need to change. It updates the independent BabelDOC plugin and Python bridge, not the Wenyi Desktop application itself.
Fixes
- Separate contents entries: Capture confidently detected numbered contents rows before style grouping, then fill each translated entry into its saved region. Formula and annotation identities remain attached to their original entries, and each row stays contiguous in the PDF text stream for search and copy.
- Unchanged text keeps its original appearance: When paragraph text and all annotation placements are unchanged, preserve the frozen glyphs, fonts, sizes, coordinates and rotation instead of recreating and reflowing the text. This covers standalone page numbers and rotated labels such as arXiv identifiers; links are rebuilt from the unchanged glyphs. Explicitly moved or failed placements still follow normal annotation handling.
- Reliable fillback and recovery: Validate explicit target row ranges rather than guessing boundaries from newlines, and retain row metadata through frozen-session recovery. Repeated fillback continues to load a fresh frozen snapshot.
Upgrade notes
- Update the optional plugin separately in Wenyi Desktop Settings. Python users can upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip check, and restart the service. Use Wenyiv1.5.4or newer for the matching directory-row client support. - Re-extract affected PDFs into fresh state to enable directory rows. Existing parsed projects and sessions are not automatically migrated; retain translations and session backups. Recovery still requires matching Python and BabelDOC versions.
- Row detection is conservative, not general table or contents-column reconstruction. Original-glyph passthrough applies only when the complete paragraph and its placements are unchanged; it does not freeze every unchanged substring inside a translated paragraph.
- The bridge remains on BabelDOC
>=0.6.2,<0.7.0; BabelDOC 0.7.1 migration is not included. Windows x64, Linux x64, macOS Apple Silicon and Intel use separate runtimes and a shared model/font archive. The service remains AGPL-3.0-only with protocol 1; runtimes are not Windows code-signed or macOS notarized.
中文
本补丁版本改进 PDF 目录分行回填,并在段落无需变化时保留原始字形。发布范围是独立 BabelDOC 插件和 Python bridge,不是 Wenyi Desktop 主程序更新。
问题修复
- 目录条目独立回填: 在样式重分组前记录可靠识别的编号目录行,再将各条译文回填到保存的原始区域。公式和标注身份保持归属于原条目,各行在 PDF 文本流中连续,便于搜索和复制。
- 未改动文字保持原貌: 段落文字及全部标注位置不变时,保留冻结字形的字体、字号、坐标和旋转,不再重新生成和排版文字。适用于独立页码、arXiv 标识等旋转标签;链接按未改动字形重建。显式移动或失败的标注仍走正常标注处理路径。
- 回填与恢复: 严格校验显式译文行范围,不依靠换行符猜测边界;冻结会话恢复后仍保留行元数据。重复回填继续从冻结快照加载新副本。
升级说明
- 在 Wenyi Desktop 设置中单独更新可选插件。Python 用户可在专用环境中升级
wenyi-babeldoc-bridge,运行uv pip check后重启服务。目录分行的配套客户端支持需要 Wenyiv1.5.4或更新版本。 - 将受影响的 PDF 重新提取到全新状态,才能启用目录分行。已有解析项目和会话不会自动迁移;请保留译文及会话备份。恢复仍要求匹配的 Python 和 BabelDOC 版本。
- 目录行检测采用保守策略,不是通用表格或目录分栏重建。原字形直通仅适用于完整段落及标注位置均不变的情况,不会冻结已翻译段落中所有未改动的子串。
- 仍使用 BabelDOC
>=0.6.2,<0.7.0,不包含 0.7.1 迁移。Windows x64、Linux x64、macOS Apple Silicon 和 Intel 分别提供运行包,共用模型/字体包。服务仍采用 AGPL-3.0-only、protocol 1;运行包未做 Windows 代码签名或 macOS 公证。
BabelDOC Desktop plugin v0.1.6
English
This release adds aligned PDF formula, marker, style and hyperlink preservation
to the independent BabelDOC bridge. It includes the optional plugin and Python
package, not an update to the Wenyi Desktop application itself.
PDF annotations
- Original inline objects: Restore frozen formula and corner/citation-marker
objects and mixed styles at the target offsets supplied by Wenyi's shared
annotation aligner, without regenerating mathematical content. - Clickable translated text: Capture URI, internal page and resolved named
links before BabelDOC clears annotations, then rebuild their clickable areas
from the actual typeset glyphs. Internal destinations follow selected output
pages; links to omitted destination pages are omitted. - Independent best-effort fillback: Keep successful items even when another
item in the same paragraph fails. Missing, stale, invalid or fallback positions
omit only the affected annotations; formula/style conflicts omit their
participants. Annotation failures do not block export or replace translated
paragraphs with source text. Fillback itself makes no translation-model calls. - Diagnostics: Advertise
annotation-best-effort-v1, return per-request
validation/application statistics, and save per-item skip reasons in
annotation_stats.json. The compatible Wenyi client records the summary in
its diagnostic log.
Extraction and recovery
- Preserve original glyph offsets before style regrouping and split discontinuous
style/link runs, keeping stable paragraph and composition identities. - Select linked glyphs by their centers so tall citation-link rectangles do not
accidentally include words on neighboring lines. - Load fresh frozen IL for every fillback. Regression tests cover partial failure
within one paragraph, formulas/markers, page selection, retries, recovery,
deletion and cropped/rotated link geometry.
Upgrade notes
- Update the optional BabelDOC plugin from Desktop Settings; publication does
not automatically upgrade an installed plugin. Python users can upgrade
wenyi-babeldoc-bridgein their dedicated environment, runuv pip check,
and restart the service. - New annotation preservation requires an annotation-capable Wenyi build.
Older text-only projects need fresh extraction to acquire annotation metadata;
existing sessions are not silently migrated or supplemented. Keep translation
state and session backups, and retain matching Python/BabelDOC versions for
frozen-session recovery. - Placement-validation and application rates are not semantic accuracy scores.
A model can mark an overly broad formula range as aligned, including in date
fields. BabelDOC's formula recognition, fixed paragraph boxes and font scaling
still limit fidelity; this release does not claim those quality issues are solved. - Windows x64, Linux x64, macOS Apple Silicon and Intel have separate runtimes
and share the model/font archive. The service remains AGPL-3.0-only with
protocol 1 and PyMuPDF 1.x; packages are not Windows code-signed or macOS notarized.
中文
本版为独立 BabelDOC bridge 增加 PDF 公式、角标、样式及超链接的对齐回填。
发布范围包含可选插件和 Python 包,不是 Wenyi Desktop 主程序更新。
PDF 标注
- 恢复原始行内对象: 按文译共用对齐器提供的译文位置,恢复冻结的公式、
引用角标对象及混合样式,不重新生成数学内容。 - 译文链接可点击: 在 BabelDOC 清理注释前保存网址、内部页码和已解析命名目标,
再按排版后的实际字形重建点击区域。内部跳转映射到选定输出页;目标页未导出时省略链接。 - 同段逐项回填: 同段中一项失败不影响其他成功项。位置缺失、过期、无效或处于
回退状态时,仅忽略相关标注;公式/样式冲突时忽略参与冲突的项。
标注失败不会阻止导出,也不会把已译段落替换回原文。回填本身不调用翻译模型。 - 诊断统计: 声明
annotation-best-effort-v1能力,返回本次回填的位置校验
通过率和应用统计,并在annotation_stats.json中记录逐项跳过原因。
兼容的文译客户端会将摘要写入诊断日志。
提取与恢复
- 在样式重分组前保存原始字形偏移,拆开不连续的样式/链接范围,保留稳定的段落和组成身份。
- 按字形中心判断链接范围,避免过高的文献链接矩形误选相邻行文字。
- 每次回填重新加载冻结 IL。回归测试覆盖同段部分失败、公式/角标、页面选择、
重试、恢复、删除,以及裁切/旋转页面的链接几何。
升级说明
- 在 Desktop 设置中更新可选 BabelDOC 插件。发布本版不会自动升级已安装的插件。
Python 用户可在专用环境中升级wenyi-babeldoc-bridge,运行uv pip check后重启服务。 - 新标注保留能力需要支持 PDF 标注对齐的文译版本。旧的纯文本项目需重新提取以补齐
标注元数据;不会静默迁移或补写现有会话。保留译文状态和会话备份;
冻结会话恢复仍要求匹配的 Python/BabelDOC 版本。 - 位置校验通过率和应用率不代表语义正确率。模型仍可能把过宽的公式范围标成成功,
日期栏就是一种情况。BabelDOC 的公式识别、固定段落框和字体缩放仍会限制保真度,
本版不宣称这些质量问题已经解决。 - Windows x64、Linux x64、macOS Apple Silicon 和 Intel 分别提供运行包,共用模型/字体包。
服务仍采用 AGPL-3.0-only、protocol 1 和 PyMuPDF 1.x;
运行包未做 Windows 代码签名或 macOS 公证。
BabelDOC Desktop plugin v0.1.5
English
This maintenance release restores complete Desktop plugin packaging and keeps
the independent BabelDOC bridge on PyMuPDF 1.x. It includes the optional plugin
and Python package, not an update to the Wenyi Desktop application itself.
Fixes
- Intel macOS builds: Standard installations and Desktop builds now require
pymupdf<2, avoiding the PyMuPDF 2.x source-build failure on Intel Macs.
Compatible 1.x minor versions remain resolver-selected; Desktop still pins
BabelDOC 0.6.2, while ordinary installations support BabelDOC 0.6.
Development and documentation
- Separate language guides: English installation, API, Desktop and publishing
guides live indocs/, with matching Chinese guides indocs/zh/.
READMEs provide concise introductions and navigation. - Bundled documentation: Runtime ZIPs include both language sets so their
README links continue to work. Regression tests cover documentation packaging
and the PyMuPDF major-version boundary.
Upgrade notes
- Update the optional BabelDOC plugin from Desktop Settings; publication does
not automatically upgrade an installed plugin. Python users can upgrade
wenyi-babeldoc-bridgein their dedicated environment and runuv pip check. - v0.1.4 was Python/source-only. Desktop users upgrading from v0.1.3 or earlier
move to BabelDOC 0.6: finish 0.5 sessions in the old environment or re-extract
the original PDF. Frozen sessions are not migrated across BabelDOC versions.
The updated model/font archive includes native-parser CMaps and needs a new
download when the matching verified archive is not already cached. - PyMuPDF 1.x is a compatibility choice, not a long-term security-support
guarantee; upstream does not maintain a separate 1.x branch. - Windows x64, Linux x64, macOS Apple Silicon and Intel have separate runtimes
and share the model/font archive. The service remains AGPL-3.0-only with
protocol 1; packages are not Windows code-signed or macOS notarized.
中文
本次维护版本恢复完整 Desktop 插件打包,并将独立 BabelDOC bridge 保持在 PyMuPDF 1.x。
发布范围包含可选插件和 Python 包,不是 Wenyi Desktop 主程序更新。
问题修复
- Intel macOS 构建: 普通安装与 Desktop 构建统一要求
pymupdf<2,
避免 Intel Mac 上 PyMuPDF 2.x 的源码构建失败。兼容的 1.x 小版本仍由依赖解析选择;
Desktop 继续固定 BabelDOC 0.6.2,普通安装支持 BabelDOC 0.6。
开发与文档
- 中英文分离: 英文安装、API、Desktop 和发布指南放在
docs/,
对应中文指南放在docs/zh/;README 只保留简介与导航。 - 随包提供文档: 运行包包含两种语言的文档,避免 README 链接失效。
回归测试覆盖文档打包和 PyMuPDF 大版本约束。
升级说明
- 在 Desktop 设置中更新可选 BabelDOC 插件。发布本版不会自动升级已安装的插件。
Python 用户可在专用环境中升级wenyi-babeldoc-bridge,再运行uv pip check。 - v0.1.4 仅发布了 Python/源码。从 v0.1.3 或更早版本升级的 Desktop 用户将切换到
BabelDOC 0.6:请在旧环境完成 0.5 会话,或重新提取原 PDF;冻结快照不跨版本迁移。
新模型/字体包包含原生解析器的 CMap;没有缓存匹配且已验证的资源包时,需要重新下载。 - 保留 PyMuPDF 1.x 是兼容性选择,不代表长期安全支持;上游不维护独立的 1.x 分支。
- Windows x64、Linux x64、macOS Apple Silicon 和 Intel 分别提供运行包,共用模型/字体包。
服务仍采用 AGPL-3.0-only 和 protocol 1;运行包未做 Windows 代码签名或 macOS 公证。
Wenyi BabelDOC Bridge v0.1.4 — Python / source
English
Python/source-only release. This release does not include the self-contained
Wenyi Desktop plugin runtimes, model archive, or plugin manifest. It is deliberately
not marked as GitHub's latest release; the Desktop release remains v0.1.3.
Changes
- Upgrade the bridge to BabelDOC 0.6 (
>=0.6.2,<0.7.0) and its native PDF parser.
Keepbitstring>=4.3.0,<5for the requiredBitStreamAPI. - Derive package and HTTP API versions from Git tags with
hatch-vcs.
The release workflow builds, checks, and uploads Python distributions to PyPI
using Trusted Publishing. - Default to one dedicated Python 3.12 environment for
wenyi-cliand
wenyi-babeldoc-bridge. They remain separate processes communicating over HTTP.
Remove the PDF2zh-specificzoteroextra; neither PDF2zh frontend is required. - Add parser, packaging, and synthetic PDF regressions for stable paragraph IDs,
translated fillback, page selection, session recovery, and cleanup. - PRs validate models and Linux/Windows runtimes with HTTP smoke tests.
Main/tag/manual builds retain the four-platform Desktop pipeline.
Installation and upgrade
uv venv .venv-wenyi --python 3.12
uv pip install --python .venv-wenyi/bin/python "wenyi-babeldoc-bridge==0.1.4"
uv pip check --python .venv-wenyi/bin/python
.venv-wenyi/bin/wenyi-babeldoc-bridgeThe Python package becomes available after the release's Publish bridge to
PyPI workflow succeeds. Install a compatible published wenyi-cli separately
into this same environment when its CLI/Core releases are available. Windows
uses Scripts/python.exe and Scripts/wenyi-babeldoc-bridge.exe.
- Finish BabelDOC 0.5 sessions in the old environment or re-extract the original
PDF with 0.6. Frozen snapshots are not migrated across BabelDOC versions. - Fonts and models still use BabelDOC's resource download/cache mechanism; they
are not bundled into the Python wheel. - The bridge remains AGPL-3.0-only. Sharing an environment does not remove any
component's or dependency's license obligations.
Release scope and validation
The owner explicitly approved a Python/source-only release without waiting for
the queued macOS Desktop jobs. Linux/Windows runtime and HTTP smoke checks have
passed. No macOS runtime or complete Desktop plugin is published or claimed as
validated here. A subsequent full Desktop publication should use a new version;
this release and its tag will not be replaced.
Synthetic PDF tests use generated documents and no translation API. No private
PDFs or real-model translation-quality evaluation were used.
中文
本版仅发布 Python 包与源码。 不包含 Wenyi Desktop 的独立运行包、模型包或插件
清单,且不会标为 GitHub latest;Desktop 发布版仍为 v0.1.3。
变更
- bridge 升级到 BabelDOC 0.6(
>=0.6.2,<0.7.0),使用原生 PDF 解析入口;
保留bitstring>=4.3.0,<5,以提供所需的BitStream接口。 - 包版本与 HTTP API 版本通过
hatch-vcs从 Git 标签生成。Release 工作流自动
构建、检查 Python 分发包,再通过可信发布上传 PyPI。 - 默认将
wenyi-cli与wenyi-babeldoc-bridge安装到一个专用 Python 3.12
环境;两者仍是通过 HTTP 通信的独立进程。移除 PDF2zh 专属的zoteroextra,
不需要安装 pdf2zh 或 pdf2zh-next。 - 增加解析、打包与合成 PDF 回归,覆盖稳定段落 ID、译文回填、页码选择、会话恢复
和清理。 - PR 验证模型包及 Linux/Windows 运行包与 HTTP 冒烟;main、标签和手动构建仍保留
四平台 Desktop 流程。
安装与升级
uv venv .venv-wenyi --python 3.12
uv pip install --python .venv-wenyi/bin/python "wenyi-babeldoc-bridge==0.1.4"
uv pip check --python .venv-wenyi/bin/python
.venv-wenyi/bin/wenyi-babeldoc-bridge本 Release 的 Publish bridge to PyPI 工作流成功后,才可从 PyPI 安装。
CLI/Core 的兼容版本发布后,再将 wenyi-cli 安装到同一个环境。Windows 使用
Scripts/python.exe 与 Scripts/wenyi-babeldoc-bridge.exe。
- 在旧环境中完成 BabelDOC 0.5 会话,或在 0.6 环境重新提取原 PDF;冻结快照不跨
BabelDOC 版本迁移。 - 字体与模型仍通过 BabelDOC 的资源下载和缓存机制获取,不包含在 Python wheel 中。
- bridge 仍采用 AGPL-3.0-only;共用环境不会豁免任何组件及其依赖的许可证义务。
发布范围与验证
本次经维护者明确同意,不等待仍排队的 macOS Desktop 作业,先发布 Python 包与源码。
Linux/Windows 运行包及 HTTP 冒烟已通过。本版不发布 macOS 运行包或完整 Desktop
插件,也不声称已经完成这些验证。后续完整 Desktop 发布应使用新版本,不替换本版
Release 或标签。
合成 PDF 回归使用临时生成的文档,不调用翻译 API;未使用私有 PDF,也未进行真实
模型的翻译质量评估。