Knowbel Reader

Knowbel Reader is a PDF and EPUB reader for studying dense documents: textbooks, papers, lecture notes, and articles saved from the web. It does everything you'd expect for highlighting and note-taking, and adds an AI study layer. You can select a passage and "deep dive" into it, ask questions about the whole document, and build a glossary of its terms and notation. Answers are grounded in the document itself and cite the pages they draw on.

It runs in the browser or as a native desktop app via Tauri. So far it has only been tested on macOS. Tauri can also build it for Windows and Linux, but those builds haven't been tried yet.

Features

Library

• Import PDFs and EPUBs (⌘O). Each one's title, author, page count and cover are read from the file, and importing the same file twice opens the existing entry.
• An EPUB becomes a PDF when it's added: the browser lays it out (using the book's own stylesheets) as a 6×9in paperback set in Literata. The result has real text, the book's pictures, working links, and its table of contents as bookmarks. Every feature below then works on it the same way it does on a PDF. Known gaps: CJK text and emoji aren't drawn (the bundled fonts don't cover them), small caps come out as ordinary letters, and right-to-left scripts aren't laid out right to left.
• Add from URL: save a web page (a blog post, an article, documentation) as a document, or add a PDF or EPUB straight from a link. Paste a link anywhere on the library, drop one in from a browser, or use the button.
  • A page becomes a PDF the same way an EPUB does, set like the rest of the library. The article is found with Mozilla's Readability (Firefox's Reader View). Anything Readability drops from inside the article (a code sample, a table, a note, words cut from the middle of a paragraph) is checked and put back. The page's own navigation, ads and related links stay out.
  • Figures the page draws with CSS (boxes and arrows in styled HTML) are captured as pictures from a render of the page with its own stylesheets. Their words stay searchable, so Ask and the glossary can use them.
  • Pictures are saved into the document. Video, audio and embeds (YouTube, tweets, CodePens, Gists) become a link with a thumbnail, since a PDF can't play them. Formulas are kept as their TeX source.
  • When something couldn't be kept exactly (a picture that wouldn't download, a video saved as a link), you're told before the document opens. Otherwise it opens straight away.
  • An arXiv abstract page fetches the paper's PDF, titled from the abstract page. A paper's landing page that offers a PDF asks which one you want.
  • The page is saved as it was the day you added it, with its address, its author and date, and Open original page in the View menu. Adding the same address again offers the saved copy, or a fresh one if it changed.
  • Known gaps: pages that build their content with JavaScript, and pages behind a sign-in or paywall, can't be saved yet.
• "Continue reading" brings you back to where you left off.
• Export and import all your notes as a single backup file.

Reading and annotating

• Text highlights with comments. Highlight mode highlights each selection as you make it, in the color and style you choose.
• Area highlights (hold Alt and drag) and sticky notes (right-click the page).
• ⌘Z undoes adding, editing or deleting a highlight or note.
• Table of contents, full-text search (⌘F) and an annotations sidebar.
• Page navigator with typed page numbers (printed labels like "xii" work) and Back / Forward (⌘← / ⌘→) over every jump.
• Zoom, two-page view, continuous scrolling, and light/dark themes.
• Cross-references: "Eq. (3.2)", "Theorem 4.1", "Fig. 7" and "[12]" link to their targets, with a preview of the target exactly as the document prints it.
• Export an annotated copy of the PDF.

AI study tools

• Deep Dive: select text or a region of a page and get an explanation that draws on the surrounding text and related passages elsewhere in the document. You can ask follow-ups or dive deeper into a phrase in the answer, building a tree of nested dives.
• Ask (⌘J): ask questions about the whole document, a chapter, or a page range ("what does section 3.2 prove?"). Only the relevant passages are sent to the model, and you can attach text or page regions to a question.
• Citations: each answer cites the passages it relies on. Each citation is a chip that jumps to the exact spot in the PDF, and claims from outside the document are flagged as such.
• Glossary: look up a term and get a definition based on how this document uses it, along with every place it appears and a matching dictionary/Wikipedia entry.

Requirements

• Node.js 22.18 or newer (the test suite runs TypeScript directly with Node).
• An API key for at least one supported model provider (see Configure API keys). The reader and annotation features work without one, but the AI study tools don't.
• For the desktop app: the Rust toolchain and Tauri's platform prerequisites.

Getting started

npm install
cp .env.example .env    # then add your API key (next section)
npm run dev

npm run dev starts both the Vite frontend and the API server (server/, on port 4300). Open the URL Vite prints, usually http://localhost:5173.

Configure API keys

All model settings live in .env in this directory. The file is git-ignored, so keys stay local. Start from .env.example.

With nothing else set, every feature uses Meta's Muse Spark API (muse-spark-1.3-contributor), so the minimum setup is one line:

MUSE_SPARK_API_KEY=your-key-here

Providers

Provider	Key in .env	Notes
muse-spark	MUSE_SPARK_API_KEY	Default. Meta's Muse Spark, via the OpenAI SDK
anthropic	ANTHROPIC_API_KEY	Claude API
openai	OPENAI_API_KEY	OpenAI API
google	GOOGLE_API_KEY	Gemini API

You only need keys for the providers you actually use.

Choosing a provider per feature

Each feature runs under a role, and each role can use a different provider and model through <ROLE>_PROVIDER and <ROLE>_MODEL:

Role	Used by	Default
DIVE	Deep Dive	muse-spark · muse-spark-1.3-contributor
ASK	Ask	Whatever DIVE uses
GLOSSARY	Glossary definitions and notation lookups	muse-spark · muse-spark-1.3-contributor, low reasoning effort
QUICK	Short one-shot answers (/api/dive/quick) and the grounding eval's --judge	muse-spark · muse-spark-1.3-contributor

For example, to run Deep Dive and Ask on Claude while keeping the glossary on Muse Spark:

DIVE_PROVIDER=anthropic
DIVE_MODEL=claude-sonnet-5
ANTHROPIC_API_KEY=sk-ant-...
MUSE_SPARK_API_KEY=...

Restart npm run dev after editing .env.

Other settings

Variable	Purpose
<ROLE>_MAX_TOKENS	Cap on a role's reply length (defaults: DIVE 1024, QUICK 512, GLOSSARY 1024)
REFERENCE_CONTACT	Your email or URL, sent to Wikimedia with glossary lookups. Raises their rate limit from 10 to 200 requests a minute
PORT	API server port (default 4300)
LOG_USAGE=1	Log token usage per model call in production (always on in development)

Desktop app

npm run tauri dev     # native window against the Vite dev server
npm run tauri build   # installable bundle in src-tauri/target/release/bundle/

How the packaged app differs from the browser version:

• It needs Node.js at runtime. The app starts its bundled API server with the system's node, so Node must be installed on any machine that runs it.
• It reads .env at build time. npm run tauri build copies your .env into the bundle. Rebuild after changing keys or providers. Because your keys ship inside the bundle, don't distribute a build made with your personal .env.
• It stores data as plain files. Your library and notes go in the app data folder (on macOS, ~/Library/Application Support/com.knowbelreader.app/library/), one folder per document. The browser version uses IndexedDB instead.

Scripts

Command	What it does
npm run dev	Start the frontend and API server for development
npm run build	Type-check and build the frontend into dist/
npm start	Serve the built frontend and API from one Node process
npm test	Run the unit tests
npm run lint	Lint with oxlint
npm run build:server	Bundle the API server for the desktop app (run by tauri build)
npm run tauri dev / build	Run or package the desktop app
npm run eval:grounding	Evaluate retrieval and citation quality against a corpus (see scripts/eval-grounding.mjs)

Project layout

src/            React + TypeScript frontend
  epub/         EPUB import: unpacking, layout, and conversion to PDF
  deepdive/     Deep Dive panels, prompts and rendering
  ask/          Ask panel and question threads
  glossary/     Glossary lookups and definitions
  notation/     Symbol lookups
  grounding/    Retrieval, prompt sources and citation checking
  xref/         Cross-reference detection and previews
  library/      Library home, import, backup
  web/          Add from URL: fetching, finding the article, capturing figures,
                and building the document (see web/index.ts for the stages)
  storage/      Per-document storage (files on desktop, IndexedDB on web)
server/         Express API server that calls the model providers, and fetches
                web pages for Add from URL (web.js)
  providers/    One module per provider
src-tauri/      Tauri desktop shell
scripts/        Build and eval scripts

Adding a provider

Each provider in server/providers/ exposes the same small interface (streamReply and complete) and is registered in server/providers/index.js. Any service that speaks the OpenAI Chat Completions API (Groq, Together, Fireworks, and so on) can be added the way museSpark.js wraps openAiCompatible.js with its own base URL.

Built with

React, TypeScript, Vite, Express, Tauri, react-pdf-highlighter-plus (patched via patch-package), pdf.js, pdf-lib, fflate, KaTeX, Mermaid, Readability and html2canvas-pro. EPUBs are set in Literata and DejaVu (both under open font licenses).