Desktop PDF viewer optimised for high magnification viewing with AI-guided "rail reading".
Built with .NET/Avalonia, PDFtoImage (PDFium) for PDF rasterisation, SkiaSharp for GPU-accelerated rendering, and Docling Heron-INT8 (ONNX) for layout detection.

Download · Microsoft Store · Website · User Guide

RailReader2 is licensed under the MIT License.

Versions prior to 3.0.0 were released under the GNU General Public License v3 (GPLv3). Starting with version 3.0.0, the project is licensed under the MIT License. As the sole author and copyright holder, this relicensing applies to all new releases going forward. Previous releases remain available under the terms of the GPLv3.

As a visually impaired user, I need a PDF viewer that works comfortably at high magnification for sustained reading. The tech industry rarely builds for this "missing middle": the market is too niche for standard software companies, and tools designed for full blindness aren't appropriate when you have poor-but-usable vision.

High magnification introduces real navigation challenges: context loss, inefficient scrolling, and UI elements that break at zoom. RailReader2 addresses these by using AI layout analysis to guide line-by-line reading through detected text blocks, like a typewriter carriage return across the page.

RailReader2 was built with Claude Code. It is an example of how AI-assisted development can enable niche accessibility tools that would otherwise never exist.

PDF pages are rasterised by PDFium (via PDFtoImage) at a DPI proportional to the current zoom level — from a 150 DPI floor up to a cap set by the chosen render-quality preset (default High caps at 525 DPI; presets range from 350 to 800 DPI, with a Custom option up to 1200). The resulting bitmap is uploaded to the GPU as a mipmapped SKImage and drawn on Avalonia's composition thread via a CompositionCustomVisual. Camera pan and zoom are applied atomically inside the Skia draw call (not via an Avalonia MatrixTransform), which keeps panning jitter-free across platforms; the bitmap only re-renders when the DPI tier changes, not on every pan/zoom frame.

At high zoom levels, navigation switches to "rail mode" — the viewer locks onto detected text blocks and advances line-by-line, like a typewriter carriage return. This is powered by ONNX layout detection models: Docling Heron-INT8 (default, detects document regions with broad class coverage) or PP-DocLayoutV3 (alternative, optimised for academic papers and includes native reading order). When using Heron or PP-S, reading order is determined via the XY-Cut++ algorithm. Non-active regions are dimmed so you can focus on the current block and line. See the layout model guide for details and model selection.

• Rail toolbar — docked vertical toolbar with toggle buttons (P/J/F/H) for auto-scroll, jump mode, line focus dim, and line highlight, plus sliders for scroll speed (or jump distance) and motion blur intensity; auto-scroll and jump mode are mutually exclusive
• Semi-automatic auto-scroll — flows through prose line by line in rail mode (P key) and parks (waits for a keypress) on arrival at a non-prose unit — equation, table, figure, heading — and at column/page breaks; press D/S to continue. Pan/zoom stay live while parked, and a "Parked — press D" affordance shows in the status bar + on-page. A brief reading beat is held on every line; which block types park is configurable (Settings > Auto-Scroll > Park On)
• Auto-scroll trigger — optionally auto-start auto-scroll after holding D/Right for a configurable delay (default 2s). Off by default, configurable in Settings > Auto-Scroll
• Jump mode — saccade-style reading (J key) that advances by a configurable percentage of the visible width; Shift+Right/Left for half-distance short jumps
• Line focus dim — smooth feathered dimming of non-active lines to reduce peripheral distraction, with configurable intensity and padding
• Line highlight toggle — independently toggle line highlight tint (H key); works with or without line focus blur
• Line highlight tint — configurable colour tint on the active line in rail mode (Auto, Yellow, Cyan, Green, or None) with adjustable opacity. Auto adapts to the active colour effect
• Click-to-select block — click on any detected block in rail mode to jump to it
• Free pan in rail mode — hold Ctrl while dragging to pan and zoom freely (even below rail threshold) to inspect images or equations. The page draws clean while you pan — the rail dim and overlay are suppressed — and release Ctrl to snap back to your original reading position and zoom level
• Zoom position preservation — zooming in rail mode no longer snaps to line start; horizontal scroll position and line screen position are preserved
• Vertical position preservation — maintains your panned vertical offset when navigating lines in rail mode
• Line snap shortcuts — Home/End keys snap to the start/end of the current line in rail mode
• Pixel snapping — quantises camera positions to the pixel grid to eliminate sub-pixel text shimmer at high zoom
• Edge-hold page navigation — in non-rail mode, hold Down/S at the page bottom for 400ms to advance to the next page. Same for Up/W at the top edge
• Margin cropping — fit/centre operations target the detected content area instead of the full page, so whitespace margins don't waste screen space at high zoom. Toggle with Ctrl+Shift+M or via Settings. Capped just below the rail zoom threshold so cropping never accidentally switches you into rail mode
• Analysis lookahead — pre-analyzes upcoming pages in the background for instant navigation
• Analysis indicator — status bar shows "Analyzing..." during layout inference
• Configurable navigation — choose which block types are navigable in rail mode via Settings → Advanced
• GPU-accelerated layout analysis — optional WebGPU execution for the ONNX layout model (Docling Heron or PP-DocLayoutV3) on supported GPUs, roughly an order of magnitude faster than CPU inference. Toggle via Settings → Advanced → GPU Acceleration (falls back to CPU automatically if unavailable; needs a restart to apply). Heron's GPU model is a separate one-time download
• View rotation — quarter-turn the displayed page for sideways scans or wide tables (Ctrl+R / Ctrl+Shift+R, or View → Rotate). When rail mode reaches a sideways block, a one-time toast points at Rotate to Read (U) to make just that block upright and re-run analysis on it; press U again to reset. A status bar badge shows the current rotation with a one-click reset. Annotation authoring is disabled while rotated (stored annotation geometry is always in the unrotated frame)

• Freeze panes — pin part of a page in place, like Excel's Freeze Panes, so a header row, a label column, or both stay visible while the rest of the page scrolls. Open the Freeze button (snowflake) on the toolbar and pick a mode — Rows (freeze everything above a horizontal guide line), Columns (everything left of a vertical line), or Both (a crossing pair; Z arms "both" directly). Click to drop the split exactly where you point — no snapping to detected boundaries, and no dependence on table detection: it is page-wide and works on any page. Zoom is locked while frozen so the panes and body can't drift. Each split pane and tear-off window freezes independently; a freeze clears when its view leaves the page. Release via Unfreeze in the flyout, the ❄ Frozen — Unfreeze chip in the pane's corner, or Z

• OCR for scanned pages — a scan is a picture of text, with no text layer for rail reading, search, selection, table cells, Markdown export, or VLM grounding to work from. Opt-in OCR recovers one so those features behave as they do on a born-digital page. Three modes in Settings → OCR: Off (default, no cost), Lines (line geometry only — enough to restore line-by-line rail reading), and Full (also recognises the text). Changing the mode re-analyses affected pages immediately, no restart
• Automatic skew correction — scans are rarely square on the glass, and line grouping is exactly the step a tilt defeats: well under a degree is enough to fragment a paragraph into a couple of huge rail lines, or fuse neighbouring lines into one. The page's tilt is measured from the OCR results and compensated for when grouping text into lines (no pixels are rotated, and square pages are left untouched). On by default
• Multilingual language packs — the bundled recogniser reads Latin-script text; optional PP-OCRv6 packs add broad coverage (Latin + CJK and more) in three size/accuracy tiers, downloaded in-app to the config folder so they work from a read-only AppImage. Recognition cost varies sharply across the tiers, and the settings panel states each one's cost before you commit to it

• Colour effects — GPU-accelerated accessibility filters (High Contrast, High Visibility, Amber, Invert) with adjustable intensity. Per-document: each tab keeps its own effect, persisted across sessions
• Colour effect cycling — press C to cycle through colour effects on the active tab, with a brief status bar toast showing the current effect
• Dark mode — toggle via Settings → Appearance; switches the Avalonia Fluent theme to dark variant
• UI font scaling — adjustable font size via Settings for high-DPI or accessibility use
• Render quality — choose a render-DPI preset in Settings → Rendering (Ultra / Quality / High / Balanced / Medium / Performance, or Custom with your own max-DPI and tier-step). Higher presets re-rasterise at a greater DPI cap for sharper text and deeper zoom; lower ones favour fluidity and lower memory. Changes apply to the open page immediately — no restart. High is the default
• Smooth zoom — scroll wheel and +/- key zooms animate over 180ms with cubic ease-out; rapid scrolling accumulates smoothly
• Motion blur — subtle directional blur during horizontal scroll and zoom for perceptual smoothness, with configurable intensity
• Fullscreen mode — F11 hides all chrome for distraction-free reading; Escape exits
• Colorblind-safe colors — status bar, link indicators, debug overlay, and annotation highlights use a colorblind-safe palette

• Multi-tab support — open multiple PDFs with independent per-tab state. Right-click a tab to duplicate or close. Opening the same file twice (or duplicating a tab) shares one underlying document — the PDF handle, layout/text caches, and annotations are shared (no duplicate analysis work), while each tab keeps its own page, zoom, and rail position
• Tab bar overflow — tabs shrink with ellipsis when many are open. Horizontal mouse wheel scrolls the tab bar. Overflow dropdown button lists all tabs
• Split panes & tear-off windows — view one document at several positions at once: split the editor into N side-by-side resizable panes (View ▸ Split Editor ▸ Split Right, Ctrl+) or move a pane into its own floating always-on-top window. Each pane/window is an independent viewport with its own page, zoom, and rail; click a pane to focus it so keyboard, scroll, and menu commands act on it
• Continuous scrolling — optional page-anchored camera that lets you scroll smoothly across page boundaries in browse mode instead of jumping page-to-page. Off by default; enable via Settings → Rendering → Continuous scrolling (no restart). The minimap gains a thin document-position strip while enabled. Rail mode is unaffected — it stays page-local and wheel still zooms at rail zoom
• Side panel (accordion) — a single-open accordion: opening one section collapses the others, and the open section fills the panel. Sections are Outline (table of contents, Ctrl+Shift+O), Bookmarks (named bookmarks, Ctrl+Shift+B), Index (figures/tables/equations browser with thumbnails and extracted equation text, Ctrl+Shift+I), Search (full-document text search, Ctrl+F), Comments (a list of annotation notes and in-PDF reviewer comments), and Portals (linked context viewports — keep a referenced figure/table/equation in view while you read; see below). Toggle the whole panel with the sidebar button at the left of the tab strip
• Whole-document figure scan — the Index section's Scan All button sweeps every page for figures, tables, and equations (beyond the background lookahead), building a complete browsable index with thumbnails
• Focus follows navigation — clicking an entry in any side-panel section (an outline heading, search result, bookmark, or figure) moves keyboard focus back to the page, so scrolling immediately drives the document rather than the list
• Named bookmarks — bookmark any page with a custom name (B key or + button in the Bookmarks section). Navigate to bookmarks with a single click. Rename and delete inline. "Back to previous location" button for quick return after jumping. Bookmarks persist in the document's annotation store (keyed by the PDF's path)
• Interactive minimap — click or drag inside to navigate. Drag the top-edge grip to move; drag the inner corner to resize. Switches to the primary's high-DPI bitmap when enlarged so it stays crisp. Position and size persist
• On-screen nav buttons — ◀/▶ buttons in the status bar for mouse-only page navigation
• Search — full-document text search in the side panel's Search section, with results grouped by page, text snippets with highlighted match terms, regex and case sensitivity toggles, and match highlighting on the page (Ctrl+F)
• Copy as LaTeX — send any detected equation, table, or figure to a Vision Language Model and copy the result to clipboard. Equations → LaTeX, tables → Markdown, figures → description. Access via Ctrl+L (current block), Ctrl+right-click (context menu), or the Edit menu (Copy Block as LaTeX / Markdown / Description / Image). Works with cloud APIs (OpenAI gpt-5.4-nano-2026-03-17 recommended) or local models (Ollama, vLLM). Configure in Settings > VLM. See the VLM setup guide for all options
• Portals (linked context viewports) — link a reference in the text ("see Figure 3") to the figure, table, or equation it points to, and the target stays in view as you rail-read past the reference. The linked target shows in a docked Portals side-panel section and, optionally, a detachable always-on-top pop-out window for multi-monitor setups. The pop-out is a live viewport — not a static thumbnail — so rail reading, freeze panes, and annotation all work inside it, and it re-aims itself as the reading position moves. Author a link by right-clicking a detected block; line-precise sources mean several references in one paragraph each surface their own target in turn. Always-on on-page markers (a gutter dot at the source line, a corner badge on the target block) show where portals are anchored — click one to show its target or jump to its source. Open in Portal (Temporary) peeks any block in the pop-out without saving a link (right-click a detected block or an Index entry), auto-dismissing as you read on unless you Lock it. Portals persist per-document in a sidecar (keyed by the PDF's path). The concept — and the name — is borrowed from Sioyek, whose Portals feature inspired this

• PDF links — click internal cross-references (citations, figure refs, TOC entries) to navigate to the exact target position, or external URLs to open in the browser with a confirmation prompt. Back/forward history with Alt+Left/Alt+Right. Hand cursor on hover
• Annotations — toggle Annotation Mode from the toolbar (or Ctrl+E) to reveal the tool row: text-markup tools Highlight, Underline, Strikethrough, and Squiggly (drag over text, sticky), plus Pen, Rectangle, Text Note, Text Box (typewriter-style FreeText, drag a box), and Eraser. A shared five-colour palette (Yellow/Green/Red/Blue/Black) applies to every colour-capable tool via the toolbar's Colour flyout; a Thickness flyout (thin/normal/thick) applies to Pen and Rectangle. Annotations render in z-order: highlights below strokes/rectangles, text notes and text boxes on top. Collapsible popup notes with folded-corner icon. Select, move, and resize annotations in browse mode. Delete selected annotations with the Delete key.
• Comments pane — the side panel's Comments section lists every annotation note and imported / in-PDF reviewer comment across the document; click an entry to jump to it, filter by source (all / reviewer / yours), and change a reviewer comment's review state inline
• Text selection — select and copy text from PDF pages via the toolbar
• Toolbar — floating Browse/Text Select/Copy toolbar for quick mode switching
• Annotation export — export PDFs with embedded annotations (File → Export with Annotations)
• Annotation JSON export — export annotation data as JSON (File → Export Annotations as JSON)
• Annotation import — import annotations from a JSON file and merge with existing (File → Import Annotations). Share annotations with other RailReader2 users
• Undo/redo — annotation history with Ctrl+Z / Ctrl+Y
• Annotation mode indicator — status bar shows active tool name in amber with a clickable exit button
• Annotation tool cursors — each annotation tool shows a distinct mouse cursor (crosshair for drawing tools, I-beam for text select, no-entry for eraser) so you always know the active mode
• Tab-switch tool reset — switching tabs automatically exits any active annotation mode to prevent accidental edits

• Render pages as PNG — export PDF pages as images with optional colour effects (high contrast, high visibility, amber, invert) and annotation overlay baked in
• Extract document structure — output outline, ONNX layout blocks, and per-block text as JSON
• Export annotations — export annotations as rich JSON (with extracted text, layout block correlations, and nearest section headings) or as an annotated PDF
• Batch VLM transcription — send detected equations, tables, and figures to an OpenAI-compatible vision API and write LaTeX/Markdown/descriptions as JSON. Supports strict JSON schema output, per-class endpoint routing (e.g. local for equations + cloud for figures), and $OPENAI_API_KEY env-var fallback
• Export to Markdown — convert a PDF to structured Markdown with heading hierarchy (matched against the PDF outline), LaTeX equations, pipe tables, figure images/descriptions, and annotation blockquotes. Degrades gracefully: full fidelity with ONNX + VLM, text-with-placeholders with ONNX only, plain text without either
• Ships as separate standalone binaries for Linux and Windows on GitHub Releases

• Menu bar — File, Edit, View, Rail, Navigation, Help menus. Every command is reachable by name (the Rail menu surfaces the rail toggles, the Edit menu the block-copy actions), items grey out when unavailable (e.g. Export with Annotations on an encrypted PDF), and each carries an Alt+letter access key (mnemonic) for keyboard and assistive navigation
• Vector icons — the toolbar and panel controls use crisp Lucide SVG icons that inherit the theme text colour and scale with the UI font-size setting
• Settings panel — live-editable rail reading parameters with persistence
• Tabbed settings — organised settings panel with Appearance, Rendering, Rail Reading, Auto-Scroll, Advanced, OCR, and VLM tabs
• Keyboard shortcuts dialog — press F1 or Help → Keyboard Shortcuts for a complete reference
• Tooltips — all interactive controls have descriptive tooltips
• Splash screen — startup splash while ONNX model loads
• About dialog — version info and credits (Help → About)
• Diagnostic logging — session log file written to the config directory; export via Help → Export Diagnostic Log, or copy the path from Help → About for bug reports
• Disk cleanup — removes cache, old logs, temp files (Help → Clean Up Temp Files)
• Debug overlay — visualise detected layout blocks with class labels and confidence

Download the latest release from GitHub Releases. The AI layout model is bundled in all packages.

Advanced users: A standalone CLI for automated PDF extraction is also available from the same release page — see Command-line interface below. Download railreader2-cli-linux-x64.tar.gz (Linux) or railreader2-cli-win-x64.zip (Windows), extract, and run.

Download railreader2-linux-x86_64.AppImage, make it executable, and run it:

chmod +x railreader2-linux-x86_64.AppImage
./railreader2-linux-x86_64.AppImage

There are two ways to install on Windows:

Microsoft Store (recommended): Install directly from the Microsoft Store. This provides automatic updates, no SmartScreen warnings, and clean install/uninstall. Note: the Store release may lag behind the GitHub release by a few days due to certification review.

Standalone installer: Download railreader2-setup-x64.exe from GitHub Releases and run it. This always has the latest version immediately.

Windows SmartScreen warning (standalone installer only)

Windows may show a "Windows protected your PC" SmartScreen warning because the standalone installer is not code-signed. This does not apply to the Microsoft Store version.

To proceed:

1. Click More info in the SmartScreen dialog.
2. Click Run anyway.

If your browser warns that the file "may be harmful", choose Keep (Chrome) or Keep anyway (Edge) before running it. The source code is fully public on GitHub — you can verify what is being installed.

After installing, launch RailReader2 from your application menu or desktop shortcut. You can also open a PDF directly from your file manager by double-clicking (once associated) or by passing the file as a command-line argument:

railreader2 <path-to-pdf>

Use File → Open (Ctrl+O) to open a PDF from within the app.

The headless CLI (railreader2-cli) provides five commands for automated PDF extraction:

# Render pages as PNG with amber colour effect
railreader2-cli render paper.pdf --pages 1-5 --effect amber --output-dir ./out

# Extract document structure with layout analysis
railreader2-cli structure paper.pdf --analyze --include-text --output structure.json

# Export annotations with extracted text and block context
railreader2-cli annotations paper.pdf --include-text --output annotations.json

# Transcribe every equation and table to LaTeX/Markdown via a vision LLM
railreader2-cli vlm paper.pdf --classes equation,table \
    --endpoint https://api.openai.com/v1 --model gpt-5.4-nano-2026-03-17 \
    --output transcriptions.json

# Export PDF to structured Markdown (headings, LaTeX, tables, figures, annotations)
railreader2-cli export paper.pdf --output paper.md

Run railreader2-cli --help or railreader2-cli <command> --help for all options. See docs/user-guide.md for full reference.