bz Reader

by 叫我包仔
5
4
3
2
1
Score: 35/100

Description

把 Obsidian 变成一个能读书的地方。桌面端 + 移动端、多格式(包括 PDF,漫画),多语言、AI 锚定追问 / 全书问答 / 语义搜索 / 知识库、离线听书、双链,统计,数据全部本地。

Reviews

No reviews yet.

Stats

15
stars
4,856
downloads
1
forks
16
days
0
days
2
days
0
total PRs
0
open PRs
0
closed PRs
0
merged PRs
3
total issues
1
open issues
2
closed issues
666
commits

Latest Version

2 days ago

Changelog

更新日志

✨ 新增

  • 多分类 :一本书可以同时归进好几个分类,菜单里点一下加进去、再点一下移出来。 QQ 频道 ↗
  • 点图放大 :书里的图片点一下就放大看,能缩放、拖动、复制和保存,设置里默认关。 QQ 频道 ↗
  • 导出书摘 :导出的书摘末尾多了一节「阅读记录」,把阅读时长和每次会话一起带走。 QQ 频道 ↗
  • 导出书摘 :每条划线下面写上了它是什么时候划的。 QQ 频道 ↗
  • 彻底删除 :移出书架时可以选择「彻底删除」,连记录和阅读数据一起清掉,源文件不动。 QQ 频道 ↗
  • 阅读统计 :阅读会话开始记录中断次数。 QQ 频道 ↗
  • 字体 :字体多了一档「跟随 Obsidian」,跟宿主主题走。 QQ 频道 ↗
  • 更新日志 :更新日志弹窗重做,正文可在「更新 / 路线图 / 反馈」之间切换。

🐛 修复

  • 加入复习 :推给 Review 的划线卡带上跳回书里的链接。

安装

  1. 打开 设置 → 第三方插件,关闭安全模式。
  2. 点 浏览,搜索 bz Reader,安装。
  3. 启用后左侧出现 bz 图标,从书架开始。

反馈

问题与建议请提到 GitHub Issues。

README file from

Github

安装

在社区市场安装(首次上架审核通过后可见):

  1. 打开 设置 → 第三方插件,关闭安全模式。
  2. 点 浏览,搜索 bz Reader,安装。
  3. 启用后左侧出现 bz 图标,从书架开始。

审核期间或搜不到时,用 BRAT 添加本仓库安装,或手动下载 Release 里的 main.js、manifest.json、styles.css,放进 <你的库>/.obsidian/plugins/bz-reader/ 后启用。

功能

类别 功能 基础版 高级版
阅读与书架 EPUB / TXT 打开即读 ✓ ✓
单栏 / 双栏排版 ✓ ✓
目录、书签、四色划线、想法 ✓ ✓
划线做成书摘图片 ✓ ✓
书内搜索 ✓ ✓
跳转历史:一键回到上一处 ✓ ✓
脚注浮窗:小注不用离页 ✓ ✓
分类管理:自建分类归书 ✓ ✓
打开与导入 Kindle 系(MOBI / AZW3)、FB2、CBZ 漫画、PDF — ✓
剪藏成书:剪藏本合成一本 EPUB · 看教程 — ✓
PDF 阅读 PDF 阅读:缩放、旋转、单双页、滚动 — ✓
页缩略图:几百页也能一屏找 — ✓
复制页链接:页码、视图、嵌入、批注 — ✓
本页反链:笔记引了这页就画出来 — ✓
文件自带批注:别人标的也能读 — ✓
导出本页为图片 — ✓
扫描页识别:像文字版一样选、搜、划 — ✓
笔记与导出 划线存成双链:笔记里点回原文 — ✓
引用页签:看哪些笔记引了这本书 — ✓
导出整本摘录:每条带双链 — ✓
数据与搜索 统计回顾:字数速度、节律热力、分类与成就 — ✓
阅读海报:年度总结与数据长图 — ✓
全库搜索:按字面搜所有书,含 PDF — ✓
语义搜索:按意思找,本书或全库 — ✓
AI · 看教程 AI 解释、翻译、摘要与追问:即划即用 — ✓
AI 全书问答:划线、正文、截图与笔记四种依据都可添加进去 — ✓
AI 对话留档并导出至笔记 — ✓
AI 伴读:边读边问,按章导读 — ✓
AI 知识库:增量阅读、知识输入、整理与输出的闭环 · 看教程 — ✓
朗读与听书 · 看教程 朗读:EPUB、TXT、PDF,念完自动翻页 ✓ ✓
三种引擎:系统语音、在线、本地自建 ✓ ✓
音色与试听 ✓ ✓
语速与高亮跟随 ✓ ✓
整章预生成 — ✓
整本预生成 — ✓
离线听书:mp3 拷进手机照样听 — ✓
阅读方式 竖排排版:EPUB / TXT — ✓
并排对照:与另一本书同步段落 — ✓
全书鸟瞰:各章划线数与一句话摘要 — ✓
PDF 重排:固定版式变流式 — —
漫画模式:分格、右到左、跨页 — —
整章翻译:原文译文对照 — —
复习:全文回顾,AI 抽认卡与出题 — —
墨水屏:黑白更清、翻页不闪 — —
界面动效:翻页与切章过渡,可关 — —

付费说明

基础版免费。高级版需要付费购买激活码:购买页面,具体价格以该页面为准。

  • 激活码在 书架 → 设置 → 授权 里填入,本地验签,不联网。
  • 不绑定设备数:换电脑或手机时,在新设备的同一位置再填一次即可。
  • 桌面端与移动端是同一个包,买一次两边都能用。

AI 是怎么工作的

  • 本地维护每本书的画像(chunk + 向量 / 摘要)。
  • 提问先做问题路由,再走混合检索,受预算约束的依据进 prompt。
  • SSE 流式回传后做引文回核。

为什么嵌入借 bz 算

  • 三件套发布(main.js / manifest.json / styles.css)带不动模型权重与推理运行时;内置固定小模型等于锁死检索精度上限。
  • 嵌入层交给 bz,模型选择权在你:本地大模型或云端服务,算力越强可上越重的模型,定位越准。
  • 书与笔记须共享向量空间(笔记命中也是依据之一),模型名与 bz 不一致会静默错配。
  • 模型名由 bz 注入画像,换模型自动判废;向量本身仍只存本地 portraits/。

画像与分块

  • portraits/<书 id>.json:plan / model / dim / chunkChars / overlap / blocks[] / overview / guides / done,块内 chapter / vec / digest / summary / text。
  • 向量用 base64 Float32Array 存盘,本地余弦相似度,原子写;查询向量 50 条 LRU 缓存。
  • 切块按段落边界、章首过半收口。
  • A 向量化:1000 字/块、重叠 120 字、8 块一批送嵌入;B 信息压缩:8000 字/块,对话模型压成要点,不参与向量检索。
  • 全局单例,移动端禁用。

混合检索与重排

  • 字面通道:中文二字组 BM25(k1 1.2 / b 0.75),章名字段权重 ×2.5。
  • 向量通道:查询改写后近邻召回,绝对门槛 0.2 / 相对门槛 0.6。
  • RRF 名次融合(k = 60):余弦与 BM25 不同量纲,按名次合并而非归一化。
  • 自适应收口:floor 3 / ceiling 30 / rel 0.3 / abs 1e-4,按章去重。
  • 邻居扩展:top-3 候选的同章 ±1 块以 0.85 权重入池;章首补偿保留命中章的第一块。
  • 预算:top 6 / 8000 字;跨书语义搜每书 3 条、全局 20 条封顶。

结构路由、生成与回核

  • 章级地图:一章一行摘要、节号补真名、字面选章权重 3/1、序号硬匹配、零模型调用。
  • 整章读:目录:正文 3:7、单章 ≥600 字、超预算截断留痕;全书问题走 bookSpread 每章均匀取样。
  • 剧透门:按当前章末块截断、清空 overview,三条检索路与追问补充段落共用一道门,默认关。
  • prompt:8000 字依据 + 最近 6 条历史;引文用「」括起并注明章节。
  • SSE 流式:TextDecoder(stream:true) 处理汉字跨帧、按字节间隔计时、80ms 节流增量 DOM 重绘、半截答案保留。
  • 引文回核:4–80 字按二字组覆盖 ≥0.8 比对本次实际依据,短句逐字;来源 ≤8 条、可跳原文。

无索引的纯 AI 路径

  • 划词解释 / 翻译 / 摘要:仅发选区,摘要 ≤100 字、解释 ≤120 字。
  • 追问:原文 ≤2000 字 + 可选同书 3 块 / 1200 字;查询用「划线前 120 字 + 本问」解决指代缺失。
  • 侧栏上下文:单条 1200 字、合计 6000 字封顶。

扫描版 PDF 是怎么划和搜的

  • 识别:头栏「识别扫描页」提供识别本页 / 接下来 20 页 / 整本 / 重试失败页;逐页串行、每批 20 页。走视觉模型或自建 / 局域网 OCR 服务(手机端只能用前者),整本按页计费会先弹确认。
  • 结果:一本书一个侧车 JSON,默认放在库内 bz-reader/storage/ocr/<书 id>.json(数据目录可改)。原 PDF 一个字节都不改。重开仍在、随同步目录走。失败页记下来供「重试失败的 N 页」重跑,没有自动重试。
  • 文字层:识别结果按行铺成透明、可选中的文字层压在扫描图上。带行框的页才能划词;只有文字没行框的页可读可搜,但不能划。
  • 划记:OCR 页与普通 PDF 共用同一套划线、四色笔、想法、双链;锚点是「页码 + 文字层字符偏移」,所以文字版那套笔记、导出、复习照常可用。没有文字层的页拖动会变成矩形框选,仍可写想法、挂双链、导出书摘图,但框内引文要认过字才有内容。
  • 搜:全书字面搜索与 PDF 内查找直接读侧车 JSON,不用建索引。命中点开跳到对应页,在透明文字层上按矩形画高亮,每页最多 40 处。全书命中 300 条、同页 50 条封顶。
  • 按意思搜:要先建画像。画像正文来自逐页取字,所以只有先 OCR 再建画像才搜得到新识别的文字;之后再认的页必须重建画像。检索还要求画像已完成、嵌入模型与当前一致。语义命中没有精确锚点,点开后用片段开头再做一次全文搜索定位。

朗读是怎么工作的

  • 取文本:EPUB / TXT 按当前可见段落或整章取样;PDF 只能念当前页的文本层,扫描页要先 OCR,且没有按章预生成。
  • 三种引擎:系统语音(完全离线、不缓存不预取)、在线语音(官方接口或 OpenAI 兼容接口)、本地自建语音服务。朗读配置与 AI 配置相互独立;手机端强制走在线。
  • 切句:先区分旁白与对白,再按句末标点切分;单句目标约 240 字,超长在逗号、顿号等软边界回退,过短碎片并入前句,纯符号不念。系统语音另按约 220 字切,最长 4000 字。
  • 调度:逐句合成、逐句播放,播完自动接下一句。默认预取 3 句(本地自建服务会尽量备整章),同时最多两个请求在途。换进度时只作废旧句请求,不打断预取。审核拒绝的句子跳过,预取失败静默重试一次,其他失败停下来给「继续听」。
  • 缓存:音频按「书 + 引擎 + 模型 + 音色 + 文本」做键存在本机,命中只刷时间、不重复合成;默认上限 300 MB,可调到 2 GB。
  • 高亮:在线与本地朗读把每一句映射回正文做页面高亮,不依赖语音事件也不靠计时;不支持时退回覆盖层。离线整章音频用时间轴反查当前句,系统语音靠开始回调。
  • 预生成:整章音频拼成一个文件,另附逐句时间与字节区间的清单。写盘原子替换、写完读回校验。进度分排队 / 合成 / 打包 / 完成 / 取消 / 失败六态,不做时间预估,停止在句边界生效。已有缓存即进度,可续做或重做。整章是高级版功能,整本仅本地引擎可用、手机不可用。
  • 离线听书:把整个包目录(音频 + 清单)放进库的 bz-reader/offline/,插件只扫一层,按书名、书 ID 依次识别。只拷 mp3 不会被发现,设置里可「重新扫描」。逐句按清单切片播放,没有区间就整章播并按时间轴推高亮。正文改过会自动对齐,播完自动进下一章。
  • 续听:只保留最近一条,停止那一刻写到本机。粒度是句。恢复按「同章原位 → 全章查找 → 退回页首」三级。换设备不跟随。

界面语言

界面跟着 Obsidian 的语言走,也可以在 书架 → 设置 → 界面 里手动指定:简体中文、繁體中文、English、日本語、Deutsch、Français、Español、Português。

切换后界面立即生效;命令名需重启 Obsidian 才更新。繁体不是简体的字形转换,是另一套台湾用语(設定 / 預設 / 匯入 / 匯出 / 檔案)。更新日志是唯一的例外:它每次发版手写,只做中文与英文两版,其余六种语言显示英文条目。

网络访问

AI 与支持者名单之外完全离线。AI 请求走你配置的 OpenAI 兼容端点(插件无自建服务器),Key 存本地、仅作为请求头发往该端点。建画像时正文分块发往 bz 配的嵌入服务或你配的对话接口;提问只发命中依据、书名章名与最近 6 条历史;划词只发选区,语义搜只发查询文本,字面搜不出网。支持者名单取 GitHub / 码云公开静态 JSON,谁先回用谁,不发送个人信息、不统计,本机缓存。

常见问题

  • 支持哪些格式? EPUB 与 TXT 开箱即读;mobi、azw3、fb2、漫画与 PDF 在高级版。
  • 会和其他插件打架吗? 阅读页是 Obsidian 原生视图,目录是左侧栏里的一个页签;划线可以复制成带定位的 Markdown 链接,点一下跳回原文那一句,不另建索引。
  • 换电脑怎么办? 库同步过去,书架、进度和标注跟着走;高级版在新设备上重新填一次激活码。
  • 手机上要单独买吗? 不用,同一个包,桌面与移动都包含在内。

插件推荐

  • bz —— 日记、备忘、读书观影、游戏、剪藏、番茄钟,复习、安全加密,ai知识库,工具坞,还有一只桌宠猫

支持作者

插件的开发、维护和教程视频都是作者一个人在做。

打赏后会出现在 bz 所有插件的感谢名单里(插件里走 设置 → 支持作者 就能看到)。

成为赞助人:点击这里

许可

本项目以 MIT 许可发布。渲染内核来自 foliate-js 1.0.1(MIT,以 npm 依赖分发,本地补丁见 patches/[email protected])。


English

Installation

Install from the community directory (visible once the initial review passes):

  1. Open Settings → Community plugins and turn off Restricted mode.
  2. Click Browse, search for bz Reader, and install.
  3. Enable it — a bz icon appears in the left ribbon; start from the shelf.

While the review is pending, or if search doesn't find it, install with BRAT, or download main.js, manifest.json, and styles.css from the Release and place them in <your vault>/.obsidian/plugins/bz-reader/, then enable the plugin.

Features

Group Feature Free Pro
Reading & Shelf EPUB / TXT, open and read ✓ ✓
One or two columns ✓ ✓
Contents, bookmarks, four-color highlights, notes ✓ ✓
In-book search ✓ ✓
Back to your previous spot: one tap after any jump ✓ ✓
Footnotes in a popover: no need to leave the page ✓ ✓
Highlights as quote card images ✓ ✓
Shelf categories ✓ ✓
Formats & Export Kindle family (MOBI / AZW3), FB2, CBZ comics, PDF — ✓
Highlights as wikilinks: click back from your notes — ✓
Citations tab: see which notes cite this book — ✓
Export every excerpt, each with a wikilink — ✓
PDF Reading Open a PDF: zoom, rotate, one or two pages, scroll; view remembered per book — ✓
Page thumbnails: find a page in a long book — ✓
Copy page links: page / view / embed / annotation — ✓
Backlinks on this page and annotations in the file — ✓
Export a page as an image — ✓
Scanned page OCR: select, search and highlight like real text — ✓
Stats & Search Stats recap: pace, rhythm heatmap, per-book ETA, badges — ✓
Reading poster: yearly recap and stats long image — ✓
Whole-library search, PDF text included — ✓
Semantic search: by meaning, in this book or the shelf — ✓
AI AI explain, translate, summarize and follow up: select and go — ✓
AI Q&A over the whole book: highlights, page text, screenshots and notes as evidence — ✓
AI chats are kept and can be exported to a note — ✓
AI companion: ask as you read, with a per-chapter guide — ✓
AI knowledge base: incremental reading, intake, and a loop from notes to output — ✓
Advanced Vertical typesetting: EPUB / TXT — ✓
Read aloud: EPUB, TXT and PDF, auto page-turn — ✓
Read two books side by side — ✓
Book overview — ✓
PDF reflow: fixed layouts read as flowing text — —
Manga mode: panels, right-to-left, spreads — —
Chapter translation: source and translation side by side — —
Review: reread in full, with AI flashcards and quizzes — —
E Ink optimization: crisper black and white, no page-turn flashing — —
Interface motion: page turns, chapter changes and shelf transitions, can be turned off — —
Clipbook to EPUB: bind clipped articles into one EPUB and read it on the shelf — —

Interface languages

The interface follows Obsidian's language, and you can also pick one of eight by hand under Shelf → Settings → Interface: 简体中文, 繁體中文, English, 日本語, Deutsch, Français, Español, Português.

The switch applies to the interface immediately; command names update after restarting Obsidian (the host registers them at load time). Traditional Chinese is not a glyph conversion of Simplified — it is a separate pack using Taiwan vocabulary (設定 / 預設 / 匯入 / 匯出 / 檔案). The changelog is the one exception: it is written by hand for each release and exists only in Chinese and English; the other six languages show the English entries.

Paid features

The free edition covers everyday reading. Pro features require a paid license key; see the purchase page for the current price.

  • Enter the key in Shelf → Settings → License. It is verified locally, with no network call.
  • Keys are not tied to a device count — just enter yours again on a new device.
  • One build covers desktop and mobile — no separate purchase.

How AI works

  • Each book keeps a local portrait (chunks + vectors / summaries).
  • A question goes through query routing, then hybrid retrieval; only evidence within budget enters the prompt.
  • The answer comes back over SSE and is followed by quote verification.
Why embeddings come from bz
  • The plugin ships as three files (main.js, manifest.json, styles.css) and cannot carry model weights or a runtime. Bundling a small model would cap retrieval accuracy for everyone.
  • The embedding layer comes from bz (Second Brain). You choose the model: a heavier local model or a cloud service. More compute allows a stronger model and more precise hits.
  • Book chunks and notes must share one vector space (note hits are evidence too). A model name differing from bz's used to cause silent mismatches.
  • bz supplies and records the model name; changing it invalidates old portraits. Vectors stay in local portraits/.
Portrait and chunking
  • portraits/<book id>.json: plan / model / dim / chunkChars / overlap / blocks[] / overview / guides / done; each block holds chapter / vec / digest / summary / text.
  • Vectors are base64 Float32Array, cosine similarity runs locally, writes are atomic; query vectors use a 50-entry LRU cache.
  • Chunks follow paragraph boundaries, with chapter-opening cutoff.
  • A Vector search: 1000 chars per chunk, 120 overlap, 8 chunks per batch. B Information compression: 8000 chars per chunk, compressed by the chat model, not part of vector retrieval.
  • One book at a time; unavailable on mobile.
Hybrid retrieval and reranking
  • Literal channel: Chinese bigram BM25 (k1 1.2 / b 0.75), chapter-name field ×2.5.
  • Vector channel: rewritten query, nearest neighbours, absolute 0.2 / relative 0.6.
  • RRF rank fusion (k = 60): cosine and BM25 are not comparable, so ranks are merged instead of normalizing scores.
  • Adaptive cutoff: floor 3 / ceiling 30 / rel 0.3 / abs 1e-4, deduplicated by chapter.
  • Neighbour expansion: ±1 chunks around the top 3 candidates enter at weight 0.85; chapter-head compensation keeps the first chunk of a hit chapter.
  • Budget: top 6 / 8000 chars; cross-book semantic search: 3 per book, 20 in total.
Structural routing, generation and verification
  • Chapter map: one line per chapter, section numbers resolved to real names, literal chapter scoring 3/1, hard match on "chapter N", no model call.
  • Whole-chapter read: contents:body 3:7, at least 600 chars per chapter, truncation marked; whole-book questions use bookSpread to sample every chapter.
  • Spoiler gate: cuts at the current chapter's last chunk and clears the overview; shared by all three paths and by follow-ups; off by default.
  • Prompt: 8000 chars of evidence plus the last 6 turns; quotes in 「」 with chapter names.
  • SSE streaming: TextDecoder(stream:true) for split UTF-8, byte-gap timeout, 80 ms throttled DOM updates, partial answers preserved.
  • Quote verification: 4–80 char quotes are compared with the actual evidence at bigram overlap ≥ 0.8, short quotes must match exactly; at most 8 sources, each jumping back to the text.
Pure AI path without an index
  • Explain / translate / summarize: only the selection is sent (≤100 / ≤120 chars).
  • Follow-up: source text ≤2000 chars plus optional 3 chunks / 1200 chars; query = 120 chars before the highlight + the current question.
  • Sidebar context: 1200 chars per item, 6000 in total.
  • OCR: the "Recognize scanned pages" menu offers this page / next 20 / whole book / retry failed pages. Serial processing, 20 pages per batch. A vision model or a self-hosted / LAN OCR service (mobile: vision only); whole-book runs ask for confirmation because pages are billed.
  • Results: one sidecar JSON per book, by default bz-reader/storage/ocr/<book id>.json (data folder configurable); the PDF itself is untouched. Results survive restarts; failed pages are kept for retry, with no automatic retry.
  • Text layer: recognized lines are laid over the scan as a transparent, selectable layer. Only pages with line boxes allow text selection; text-only pages can be read and searched, but not selected.
  • Markup: OCR pages use the same highlights, four colours, comments and wikilinks as normal PDFs; anchors are "page + text-layer character offset". Pages without a text layer fall back to rectangle selection, with comments, wikilinks and quote-card export; quoted text requires OCR.
  • Search: whole-library literal search and in-PDF find read the sidecar directly, with no index to rebuild. Hits jump to the page and draw rectangles on the transparent layer; 40 highlights per page, 300 hits overall and 50 per page.
  • Semantic search: needs a portrait. Portrait text comes from per-page OCR, so OCR must run before the portrait is built; pages recognized later require a rebuild. The portrait must be complete and the embedding model must match. A semantic hit has no exact anchor, so opening it runs a text search using the snippet.

Read aloud: how it works

  • Text: EPUB / TXT use the visible passage or a whole chapter; PDF reads only the current page's text layer, scan pages need OCR first, and PDF has no chapter pre-generation.
  • Three engines: system voice (fully offline, no cache or prefetch), online (official or OpenAI-compatible endpoint), and a self-hosted voice service. TTS settings are separate from AI settings; mobile is forced online.
  • Sentence splitting: narration is separated from dialogue, then split on sentence punctuation; target 240 chars, longer sentences fall back to comma-like boundaries, short fragments merge, symbol-only fragments are skipped. System voice uses 220 / 4000 chars.
  • Playback: sentence-by-sentence audio with auto-advance; prefetch 3 sentences (self-hosted tries the whole chapter), at most two requests in flight; changing position cancels only stale requests, not prefetch. Rejected sentences are skipped, prefetch failures retry once silently, other failures stop with "Continue listening".
  • Cache: keyed by book + engine + model + voice + text in local storage; a hit only refreshes the timestamp. 300 MB by default, adjustable up to 2 GB.
  • Highlight: sentences map back to the page as anchors rather than speech events or timers; offline whole-chapter audio uses a time axis, system voice uses the start callback.
  • Pre-generation: a chapter's audio is joined into one file plus a manifest of per-sentence times and byte ranges; atomic writes and read-back verification; six states (queued / synthesizing / packing / done / cancelled / failed), no time estimate, stop takes effect at sentence boundaries; the cache is the progress and can be resumed or rebuilt. Whole-chapter is Pro; whole-book works only with a local engine and not on mobile.
  • Offline listening: copy the whole package folder (audio + manifest) into bz-reader/offline/ inside the vault. Only one level is scanned, matched by book title and then book ID; copying mp3 files alone will not be detected. Playback slices by byte ranges, or falls back to whole-chapter audio with a time axis; edited text is realigned and the next chapter starts automatically.
  • Resume: only the latest position is kept, written when playback stops, at sentence granularity. Recovery goes same-chapter offset → whole-chapter search → page start. Positions do not follow you to another device.

Network access

Only the AI features and the supporter list in Support the author go online; everything else works fully offline.

  • Requests go to the OpenAI-compatible endpoint you configure (your own server or a third-party API). The plugin has no server of its own.
  • Building the whole-book index sends the book's text to your embedding service in chunks; a question only sends the few retrieved chunks, plus the book title, chapter name, and the last few turns when needed. Vectors and the index stay in the plugin folder and are never uploaded.
  • Embeddings come from the bz (second brain) plugin: without it there is no vector search, only the compressed-key-points path.
  • Your API key is stored with the rest of your settings in data.json inside the plugin folder. It is never sent to the plugin author; it only travels to the endpoint you configured, as a request header.
  • Opening Support the author (or the changelog) reads a public supporter list: the same static JSON from GitHub and Gitee is requested at the same time, and whichever answers first is used. Only that list is fetched — no personal data, no analytics. It is cached locally; if both sources fail, "couldn't load" and a retry button appear. The list is public data, identical in both plugins.

FAQ

  • Which formats? EPUB and TXT out of the box; mobi, azw3, fb2, comics, and PDF with Pro.
  • Does it conflict with other plugins? The reader is a native Obsidian view and the contents are a tab in the left sidebar. Highlights can be copied as locator Markdown links that jump back to the exact passage — no separate index is built.
  • Switching computers? Your vault carries the shelf, progress, and annotations. Re-enter the license key on the new device.
  • Do I pay again on mobile? No. One build covers desktop and mobile.
  • bz — Diary, memos, books and films, games, web clippings, pomodoro, review, secure encryption, AI knowledge base, tool dock, and a desktop pet cat.

License

Released under the MIT license. The rendering engine is foliate-js 1.0.1 (MIT, distributed as an npm dependency; local patches in patches/[email protected]).