DocCraft

English | 简体中文

A cross-platform PDF → Markdown, Image → Markdown, and Markdown → Excel desktop converter with built-in screenshot OCR, built with Tauri 2, React, TypeScript, shadcn/ui and pdf-inspector (Firecrawl's pure-Rust PDF classification / extraction engine). The UI is bilingual — English (default) and Simplified Chinese — switchable at runtime.

Detailed architecture & data-flow docs: docs/index.md

Highlights

• ⚡ Fast extraction — pure-Rust text-layer extraction via pdf-inspector (no browser engine, no network): typical text PDFs convert in milliseconds.
• ✏️ Draw-a-table extraction — draw vertical / horizontal separator lines over a PDF page or imported image to cut tables exactly where you want them.
• 🧠 Smart text merging — configurable paragraph policy (Guided / Smart / None) joins wrapped OCR / text lines into clean paragraphs while keeping tables, lists and headings intact.
• 🎯 Region exclusion — draw rectangles over headers, watermarks or page numbers to suppress unwanted content before conversion.
• 📐 Paddle layout analysis — the bundled PP-DocLayoutV3 DETR model (MNN) restores reading order for OCR pages, preserving skewed / multi-column layouts.
• 🔌 Offline PaddleOCR — the built-in local PaddleOCR engine (ocr-rs) runs fully on-device: scanned pages are recognized with zero network access and data never leaves the machine.

Screenshots

Features

PDF → Markdown

• Hybrid text + OCR — text pages extracted locally by pdf-inspector; scanned / image-only pages rendered to PNG and sent to a configured OCR provider (remote AI vision or local PaddleOCR), then reassembled in document order.
• Smart PDF routing — each PDF is classified (~10–50ms) as TextBased / Scanned / ImageBased / Mixed with the exact list of pages needing OCR. Pure-text PDFs never touch the network.
• Offline PaddleOCR — the built-in local PaddleOCR engine (ocr-rs) runs fully on-device with no network required; scanned pages are recognized on a resident engine (toggleable cache) with per-page confidence scores.
• Paddle layout analysis — for OCR pages, the bundled PP-DocLayoutV3 DETR model (MNN) detects reading order and region structure, so multi-column / skewed scans keep their original layout; degrades gracefully to plain Y→X sorting when the model is missing.
• Region exclusion — draw rectangles over headers, watermarks or page numbers to suppress unwanted content; apply per-page or to all pages at once, in both direct conversion and draw-table extraction.
• Smart text merging — a configurable paragraph policy (Guided / Smart / None) decides how extracted lines are joined: wrapped lines merge into clean paragraphs while tables, lists and headings stay intact. Screenshot OCR results follow the same policy.
• Configurable OCR providers — any OpenAI-chat-completions-compatible vision API (multiple vendors, multiple models per vendor) or the built-in local PaddleOCR engine. API keys are encrypted at rest (DPAPI on Windows) and never exposed to the frontend. A unified OCR mode selector offers five options: ForceLocal, ForceAi, NonTextLocal, NonTextAi, Disabled.
• Graceful OCR fallback — when no OCR provider is available, conversion still completes: pages needing OCR are skipped with <!-- OCR skipped --> comments. Per-page failures degrade to <!-- OCR failed --> comments. A bell icon in the status bar collects these as structured notices with clickable page chips and retry actions.
• Draw-a-table extraction — manually draw vertical and horizontal separator lines over a rendered PDF page to define table regions, then extract them into Markdown. Supports undo/redo, per-page lines, "apply to all pages" mode with a page limit, and OCR fallback for scanned pages (local PaddleOCR block cutting or remote AI vision with drawn-line hints).
• Batch queue — drag & drop many PDFs, worker-pool conversion with a user-configurable concurrency limit (1–16), retry / remove / export-all.
• Editor workspace — toolbar (convert), split view (PDF preview | Markdown preview), status bar (type / pages / confidence / OCR needs / notices bell).

Image → Markdown

• Dedicated workspace tab accepting PNG / JPEG images via drag & drop or file picker, with deduplicated list and thumbnails.
• Each image is recognized by the OCR engine selected by the current OCR mode (local PaddleOCR or remote AI vision).
• Results are previewed as a merged GFM document and can be exported per-image or merged into a single .md file.
• Draw-table on images — imported images can be opened in a draw-table overlay where you draw vertical lines, then the image + line positions are sent to the backend for column-based extraction (local PaddleOCR text blocks
  • column cutting, or AI vision with line hints, depending on the OCR mode).

Screenshot OCR

• Press the global hotkey (default F8) or use the tray menu to start a screen capture; a per-monitor overlay with magnifier lets you pick any screen region.
• The selected region is recognized by the engine chosen by the current OCR mode: local PaddleOCR runs on a dedicated engine instance that never queues behind batch jobs, or it goes to remote AI vision.
• Results follow the selected paragraph mode and OCR text cleanup (zero-width character stripping, whitespace collapsing, CJK ↔ Latin spacing normalization).
• Results appear in a glassmorphism popup window with pin-on-top, copy-to-clipboard and close actions, optional auto-copy, adjustable opacity, and the window position is remembered between sessions.

Markdown → Excel

• Drop or pick .md files; each is parsed for GitHub-Flavored Markdown tables.
• Inline table preview with table/row counts, single or bulk export to .xlsx.
• Configurable tables-only mode: exports only GFM tables; when off, the whole document content is written into the workbook.
• Optional strip Markdown syntax (**bold**, `code`, links) and numeric cells so numbers sort and sum; sample tables inside code fences are never exported.
• Each table is labeled with its source PDF page (Page N) when produced by this app's PDF conversion.

Performance & memory

• Per-page OCR streaming — the frontend renders and uploads one page at a time; peak memory stays at a single page image instead of the whole document.
• Virtualized PDF preview — only pages near the scroll viewport are rendered to canvas; off-screen bitmaps are released.
• Lazy Markdown / Excel preview — paginated rendering and windowed rows for large documents.
• State-preserving tabs — switching between tabs keeps every view mounted (hidden, not unmounted), so loaded files, results and queues survive tab switches.

System tray

• System tray icon with right-click menu (Open, Start Screenshot, Exit) and left-click to show the main window. Close button hides to tray instead of quitting. Configurable in Settings.

Getting Started

Prerequisites: Node ≥ 20, pnpm ≥ 10, Rust ≥ 1.85.

pnpm install       # install frontend deps
pnpm tauri dev     # run the desktop app
pnpm tauri build   # package the project

Useful checks:

pnpm exec tsc --noEmit               # frontend type check
pnpm build                           # frontend production build
cargo check --manifest-path src-tauri/Cargo.toml

Model Conversion

The bundled layout model (PP-DocLayoutV3, MNN) is converted from the original PaddlePaddle inference model via the pipeline documented in script/README.md: Paddle (inference.json / .pdmodel + .pdiparams) → ONNX → MNN.

Pre-converted MNN weights are published on ModelScope: tansen87/PP-DocLayoutV3_mnn — you can download them directly instead of converting on your own.

Installing a model

Each layout model lives in its own subdirectory named after the model, containing the MNN weights and its layout-meta.json:

<resources>/models/layout/
└─ PP-DocLayoutV3/
   ├─ PP-DocLayoutV3.mnn
   └─ layout-meta.json

In build mode the resources are mirrored to doccraft_resources/models/layout/ next to the executable (Tauri copies src-tauri/resources/ to <target>/<profile>/doccraft_resources/, see src-tauri/build.rs), so a manually installed model goes to:

doccraft_resources/models/layout/PP-DocLayoutV3/PP-DocLayoutV3.mnn
doccraft_resources/models/layout/PP-DocLayoutV3/layout-meta.json

The engine discovers any directory under models/layout/ that contains a valid layout-meta.json — dropping a new directory is enough, no code change needed. See src-tauri/resources/models/layout/README.md for the layout-meta.json format.

Configuration

• ocr-config.json — per-vendor name, base URL, protected API key, models.
• app-settings.json — maxConcurrent, cacheExtractedText, excelTablesOnly, stripMdSyntax, writeNumeric, ocrMode, screenshotHotkey, snipResultPopup, snipAutoCopy, snipResultOpacity, enableTray, textSeparator, paragraphMode, ocrTextCleanup, ocrLayoutMode, ocrLayoutModel, layoutScoreThreshold.

License

MIT