AI Book Shelf / AI 书架

A persistent DSH Web UI plugin that adds a full-page multi-format reader to the sidebar, an AI explanation rail for selected text, a paper library with translation and bilingual comparison, and a dossier library with searchable files and AI summaries.

Highlights

• Read many formats — EPUB, PDF, TXT, Markdown, HTML, FB2 and DOCX.
• Ask without leaving the page — select text for background, context, explanations and follow-ups in the right rail.
• Search the whole shelf — full-book Ctrl+F, online book search (Project Gutenberg, Standard Ebooks, Chinese Wikimedia sources) and custom OPDS sources.
• Paper mode — paragraph/whole-paper translation, bilingual comparison, AI explanations and follow-ups.
• Dossier mode — folders, Ctrl+A multi-import, keyword search, AI summaries of results, and per-folder skills.

Install

1. Download dsh-ai-book-shelf-2.6.11.tgz from the latest release, or install from npm with dsh-ai-book-shelf@2.6.11.
2. In DSH open Plugins → Add plugin and choose the downloaded tgz, or paste the npm package spec.
3. Reload the plugin and open AI 书架 / AI Book Shelf in the sidebar.

Chinese step-by-step guide: INSTALL-zh.md. Publishing and npm instructions: PUBLISHING.md.

What it does

• Shelf — scans the archive shelf (the first registered workspace by default, which is D:\BookLib on this deployment), plus any extra libraryDirs, for supported book files.
• Copies external books — a book opened or imported from outside the archive shelf is copied into archiveDir first. Identical files are reused; a name collision gets a (2) suffix rather than overwriting anything.
• Dossier mode — a separate dossier library with folders, multi-select import (Ctrl/Cmd or Ctrl+A in the file picker), folder-scoped full-text search, streamed AI summaries, and folder skills. Creating a folder adds an empty skill you can fill in; an imported SKILL.md can replace it, and AI summaries follow it.
• Public-domain search — the Search books tab queries Project Gutenberg (Gutendex, with the Gutenberg website search as a fallback), a Chinese-only Project Gutenberg catalog (lang=zh), Chinese Wikisource, Chinese Wikibooks, and Standard Ebooks, then saves the chosen book (EPUB, or an HTML book for the Wikimedia sources) into the archive shelf. Simplified Chinese queries are also tried as traditional Chinese for the Chinese catalogs. Results stream into the page one source at a time; slower sources keep searching while the first results are already usable. The search store lives outside the Library component, so opening a book or switching Harness pages does not abort the stream; returning restores the query and results, and the Host keeps result ids downloadable for 30 minutes. Only public-domain / openly licensed catalogs are used; DRM and unauthorized book sources are not supported. Choose which catalogs to query with Choose sources; the selected source homepages are shown as small links. A custom OPDS search URL can be added, but only entries that declare an open-access EPUB acquisition link are accepted. Several custom OPDS sources can be added with the + button. The built-in Chinese catalogs cover traditional Chinese Gutenberg titles and Wikisource; other catalogs are mostly English.
• Reader — parses the book into chapters, sanitizes HTML, and renders a comfortable serif reading page with per-book progress. EPUB keeps its OPF spine, headings, lists, tables, images and TOC; PDF/Text/Markdown/HTML/FB2/ DOCX are normalized into readable chapters.
• Ask while selecting — the floating selection bar contains a question field to the left of 解释. Type a specific question and press Enter or 解释, or leave it blank for a direct explanation. The question is sent together with the book, chapter, selection, and nearby context; the right rail opens and streams the answer.
• Follow-ups — keep the whole Q&A thread in the panel and send the previous turns to the model, so pronouns and earlier details stay in context. You can stop the stream, copy the answer, or use quick asks (translate / example / background).
• Per-book annotations — every completed explanation and its follow-ups are saved on the Host under the archive directory. Opening a book opens the right annotation rail with that book's saved records; click a record to reopen its Q&A thread and continue asking. Records survive browser reloads and restarts.
• Paper mode — a separate paper library for English PDFs, with virtual folders (create, drag papers in, rename), paragraph-level translation into a JSON translation file, manual correction, retranslation, whole-paper translation saved as a new Chinese paper, and a bilingual compare view. Paper reading reuses the reader's AI explanation and full-book search; the compare view has the same right-side AI rail, selection question bar, follow-ups, saved annotation history and per-answer collapse as books, and works for both the English source and the Chinese translation. Annotation history and threads can delete a whole thread or an individual follow-up question/answer pair, with an X in the record's top-right corner and a second-click confirmation. The paper library supports right-click menus for new folders, importing into a folder, and renaming; the new-folder control is an icon on the right. Books can be hidden/unhidden or deleted (moved to .epub-reader/trash) from the book-card context menu. Papers can be deleted from their card context menu (two-step confirmation); a source paper also removes its Chinese translation paper and moves the PDF/translation file to the paper trash. Paper copies live in .epub-reader/papers, separate from the book shelf. The compare toggle sits in the paper reader header and can be hidden again; each translation box is resizable from its bottom edge, auto-fits its text on first open, remembers its height/collapsed state, and can be collapsed individually or all at once. Translation text supports AI explanation.
• Zoom — Ctrl/Cmd + mouse wheel changes the reading font size in both books and papers; the paper compare view also zooms its source/translation text.
• AI panels — explanations are expanded by default and can be collapsed individually or with one “Collapse all” action.
• Find in book — Ctrl/Cmd+F opens a full-book search across every chapter, including PDF/HTML/DOCX books. Results show chapter, occurrence and a text snippet; clicking one jumps to the chapter and highlights every match. The reader toolbar shows the shortcut next to the search button. Simplified Chinese queries also match traditional text (陈文龙 finds 陳文龍).
• Keyboard — Ctrl/Cmd+Shift+E explains the current selection, Alt+← / Alt+→ move between chapters, Esc closes search/rail.

Supported formats

Format	Support
.epub	Full spine, chapters, TOC, images, cover
.pdf	Text-layer extraction, 8 pages per chapter; scanned PDFs need OCR first
.txt / .text	Plain text, automatically chunked
.md / .markdown	Headings, paragraphs, lists, quotes, code blocks
.html / .htm / .xhtml	Sanitized HTML body
.fb2	FictionBook XML
.docx	Word paragraphs and Heading 1–3 chapters
.mobi / .azw3 / .azw	Convert to EPUB or PDF first (for example with Calibre)

PDF extraction uses a vendored PDF.js runtime under vendor/pdfjs/; the package has no npm runtime dependency for it.

Configuration

Edit the installed row's config in the profile patch (or the plugin's cordis.patch.yml before installing):

- id: epub-reader
  name: '@local/dsh-epub-reader'
  config:
    # Where outside books are copied and what the shelf scans by default.
    # Empty = the first registered workspace (D:\BookLib here), then cwd.
    archiveDir: D:\BookLib
    # Extra folders to scan, if any. Absolute paths are allowed.
    libraryDirs:
      - D:\Books
    # Optional; otherwise the Host's default model is used.
    # Example: explainProvider: deepseek-account
    #          explainModel: deepseek-flash
    explainProvider: deepseek-account
    explainModel: deepseek-flash
    # Off by default: the API answers only loopback browser requests.
    allowRemote: false

The Host routes are mounted under /api/epub-reader. They accept only same-origin browser requests, expose only .epub files found under a configured shelf root, copy external files into the archive before parsing, and stream explanations over SSE.

Notes

• Books are opened read-only; progress is stored in the browser's localStorage.
• DRM-protected books are not supported.
• PDF support is text-first: a scanned/image-only PDF returns PDF_NO_TEXT instead of showing blank pages. Run OCR or convert it to a text PDF first.
• MOBI/AZW3 are not parsed directly; convert them to EPUB/PDF before importing.
• The plugin uses the official llm Host service; no API key is read by the plugin itself.
• Only the first 5 opened books are retained in memory; a re-open re-parses the file.
• Annotation history is stored in <archiveDir>/.epub-reader/annotations.json and keyed by the EPUB content fingerprint, so the same book keeps its records even at a different path. Copy that folder with the EPUB when moving a shelf to another machine if the annotations should travel with it.
• Search results are cached for 30 minutes so a result can be downloaded; a stale id returns SEARCH_EXPIRED and the UI asks you to search again.
• Clearing the search field (the browser's X) or pressing Stop searching aborts the in-flight request instead of leaving the spinner running.
• Custom OPDS sources must be lawful/open-access catalogs. The Host only accepts EPUB acquisition links marked open-access; it is not a generic downloader.
• Downloads are bounded to 120 MB and fetched only from allow-listed gutenberg.org / standardebooks.org hosts. Large books can take a while.
• The search feature needs outbound network access from the DSH Host process.

License

Plugin code is released under the MIT License. The bundled PDF.js runtime keeps its own Apache-2.0 license under vendor/pdfjs/LICENSE; bundled fonts keep the licenses in vendor/pdfjs/standard_fonts/.