Open PDF Translate

Translate PDF documents in Obsidian using OpenRouter, OpenAI, Anthropic, Gemini, Ollama, and other LLM providers. Features overlay rendering, OCR for scanned documents, batch translation with background queue, and a BBox edit mode for manual overlay adjustments.

How It Works

Translation Pipeline

The plugin follows a multi-stage pipeline to translate PDF content:

1. Text extraction. When you open a PDF and trigger translation, the plugin extracts text from the PDF page. Two extraction engines are available:

• Internal engine (default, works everywhere): Uses PDF.js (bundled with Obsidian) to read text items, their positions, font sizes, and font families directly from the PDF's content streams. No external dependencies.
• Python engine (desktop-only, more accurate on complex layouts): Spawns a PyMuPDF (fitz) child process that parses the PDF with a more sophisticated layout model. Requires Python 3.8+ and pip install pymupdf.

Both engines produce the same output format: an array of text spans with bounding rectangles (relative to the page, as 0–1 fractions), font metadata, and the raw text content.

2. Layout detection. The extracted spans are grouped into paragraphs by the layout detector

Settings like contourIndentThreshold, fontSizeTolerance, maxMergePasses, columnGapThreshold, and decorationThreshold control how aggressive the splitting/merging is. The preserveStyle option (enabled by default) treats bold and italic as different font families, producing finer-grained paragraphs.

3. Translation. Each paragraph's text is sent to your configured LLM provider.

The maxBatchChars setting controls the maximum characters per batch. If the total exceeds this, paragraphs are split into multiple batches. A sequentialDelayMs setting adds a pause between batches to avoid rate-limiting.

4. Overlay rendering. Translated text is rendered as a position: absolute div overlay on top of the original PDF page.

5. Persistence. Translations are saved to a .translations.md file alongside the PDF (in the same folder, or a configurable storage folder). The file uses the V4 format (see below) with YAML frontmatter and Obsidian %% {...} %% comments containing per-overlay JSON metadata. Writes are atomic (temp file + rename) and serialized via a per-file write lock (writingPromises).

Overlay Visibility

You can show or hide overlays without losing translations:

• Command: Toggle PDF overlay visibility (or hotkey) flips all overlays globally
• Per-page: Overlays are loaded lazily as you scroll. If visibility is off, overlays are still created in the DOM but with visibility: hidden — they become visible instantly when you toggle back on
• Settings: showOverlayByDefault controls whether overlays appear when a PDF is first opened

BBox Edit Mode

BBox Edit Mode is a non-invasive overlay editing system. When enabled:

• Normal cursor, scrolling, and all Obsidian UI work — nothing is blocked
• Click on overlay → select it (clears previous selection)
• Ctrl/Shift+click on overlay → toggle selection (additive — select multiple)
• Click on empty space → clear selection
• Shift+LMB+drag → marquee selection (a temporary overlay appears only during the drag, then is removed)
• Escape → clear selection
• Right-click on overlay → context menu:
  • Retranslate selected (paragraphs mode) — re-detects paragraphs in the selected region and retranslates
  • Merge selected (block mode) — merges selected overlays into one block and retranslates
  • Delete selected — removes selected overlays (uses REPLACE semantics, no resurrection)
  • Edit translation — opens a modal to manually edit the translated text
  • Copy translation — copies plain text (multi-line preserved via innerText)
  • Copy as callout / citation / footnote — copies with formatting templates
  • Increase/decrease font size — adjusts overlay font size (persisted to disk)
  • Increase/decrease line height — adjusts overlay line height (persisted to disk)
  • Force refresh overlays — reloads all visible overlays from disk
  • Go to translation file — opens the .translations.md file at the current page heading

BBox Edit Mode does not pause the background translation queue. You can edit overlays while background translation is running — writes are serialized via the per-file lock.

Background Translation Queue

The background queue translates entire PDFs (or page ranges) without blocking the UI:

• Entry points: Watcher modal (per-file Translate button), "Translate multiple pages..." modal, "Run All" in watcher
• Processing: The PdfLayoutQueue processes pages sequentially within a file (extraction is serial, translation is parallel up to backgroundTranslationConcurrency). Multiple files are processed in FIFO order.
• Backpressure: The extraction-to-translation queue has a high-watermark (concurrency × 10). If the queue is full, extraction pauses until translation drains it — prevents OOM on 1000+ page PDFs.
• Document reuse: PDF documents are loaded once per batch (not per page), reducing extraction time by ~30–60s on 100-page PDFs.
• Cancellation: Cancel stops in-flight work at the next chunk boundary (5–10 seconds). The queue auto-resumes when new tasks are enqueued.
• Progress: The watcher modal shows live progress per file (pages done/total + progress bar). Active translations appear in a scrollable accent-bordered section; available files appear in a compact list below.

V4 Translation Format

Translation files (.translations.md) use the V4 format:

---
pdf-source: "[[document.pdf]]"
timestamp: 2026-08-13T12:00:00.000Z
format-version: 4
engine: openrouter/openai/gpt-4o-mini
layoutSettingsHash: a1b2c3d4e5f67890
---

# Translations for document

> Last updated: 2026-08-13 12:00

## Page 1

[[document.pdf#page=1|→ View page]]

%% {"r":{"l":0.12,"t":0.34,"w":0.56,"h":0.08},"page":1,"ot":"Original text","fs":12,"ff":"sans-serif","ofs":[12,12,11],"id":"a1b2c3d4e5f67890","engine":"openrouter/openai/gpt-4o-mini"} %%

Translated text here

Per-overlay fields

Field	Type	Description
r	{l,t,w,h}	Relative bounding rect (0–1 fractions of page dimensions, 4-decimal precision)
page	number	Page number
ot	string	Original text content (used for edit-modal lookup)
fs	number	Font size (persisted but inert — rendering uses ofs instead)
ff	string	Font family
ofs	number[]	Original font sizes array (drives overlay rendering — dominant/mode size)
id	string	Stable hash hash(page + rect@2dec + textContent) — enables exact overlay lookup and merge-by-id
engine	string	<provider>/<model> that produced this translation

Frontmatter fields

Field	Description
pdf-source	Wikilink to the source PDF (ties .translations.md to its PDF)
timestamp	ISO-8601 creation/update time (display only)
format-version	4 (V3 files are auto-migrated on first edit)
engine	Primary engine used for this file
layoutSettingsHash	Hash of layout settings — isCached invalidates on mismatch, forcing re-translate when layout preset changes

V3 → V4 Migration

Automatic on first edit/translate after plugin update:

• V3 files (format-version: 3) continue to work without changes
• On next edit/translate, id and engine are stamped on all overlays, format-version bumped to 4
• layoutSettingsHash added to frontmatter
• No manual migration needed

Features

• Multi-provider support: 15 LLM providers including OpenAI, Anthropic, Gemini, OpenRouter, RouterAI, DeepSeek, xAI, Groq, Mistral, Together, Qwen, Ollama, LM Studio, vLLM, and custom endpoints.
• PDF overlay rendering: Translated text appears as an overlay on top of the original PDF, with adjustable positioning, font size, opacity, and colors. Overlays are virtualized (only 15 closest pages loaded in DOM).
• BBox Edit Mode: Non-invasive editing — click to select, shift+click to toggle, shift+drag for marquee. Right-click for context menu (retranslate, merge, delete, edit, copy, font adjust).
• OCR for scanned documents: Recognize text in scanned PDFs using vision models (GPT-4o, Claude, Gemini, LLaVA, etc.) and generate translation notes.
• Background translation queue: Translate entire documents with a background worker. Watch a folder for new PDFs and auto-translate them. Live progress with per-file status and progress bars. Backpressure prevents OOM on large PDFs.
• Layout detection: Internal (JS-based) and Python (PyMuPDF) layout engines. Configurable paragraph splitting (indent, font tolerance, merge passes, column gaps, decoration threshold).
• Watched folders: Automatically detect and translate PDFs dropped into a watched folder.
• Export: Export PDFs with baked-in translations (desktop + Python required).
• V4 translation format: Stable overlay IDs, engine stamps, layout settings hash. Automatic V3→V4 migration.
• Bilingual support: Show original and translated text side-by-side.

Installation

From Obsidian Community Plugins

1. Open Obsidian Settings → Community Plugins
2. Search for "OpenRouter PDF Translator"
3. Click Install, then Enable

Manual installation

1. Download the latest release from GitHub Releases
2. Extract main.js, manifest.json, and styles.css to your vault's .obsidian/plugins/openrouter-pdf-translator/ folder
3. Enable the plugin in Obsidian Settings → Community Plugins

Quick Start

1. Configure provider: Open plugin settings → Provider section → Select your LLM provider → Enter API key
2. Open a PDF: Open any PDF in Obsidian
3. Translate: Right-click on the PDF → "Translate multiple pages..." → Select page range → Start
4. View overlay: Translated text appears as an overlay on the PDF
5. Toggle overlay (optional): Command palette → "Toggle PDF overlay visibility" to show/hide
6. Edit overlays (optional): Toggle BBox Edit Mode → click overlays to select, right-click for context menu (edit, retranslate, delete, copy)

Desktop-only Features

The following features require desktop Obsidian (Windows, macOS, Linux) and are not available on mobile (iOS/iPadOS):

• Python layout engine (layoutEngine: 'python'): Uses PyMuPDF for advanced layout detection. Requires Python 3.8+ and PyMuPDF installed on your system:

  pip install pymupdf
  

• PDF Export with translations (Export PDF with translations command): Renders a new PDF with translated text baked in. Same Python+PyMuPDF requirement.

All other features (translation, overlay, OCR via cloud vision models, basic layout detection, BBox edit mode) work on both desktop and mobile.

Supported Providers

Provider	API Key Required	Default Model	Notes
OpenAI	Yes	gpt-4o-mini	Full support including vision
Anthropic	Yes	claude-3-5-sonnet-latest	Full support including vision
Google Gemini	Yes	gemini-2.0-flash-exp	Full support including vision
OpenRouter	Yes	openai/gpt-4o-mini	Aggregator — access many models
RouterAI	Yes	openai/gpt-4o	Russian aggregator (routerai.ru) — OpenAI-compatible API
DeepSeek	Yes	deepseek-chat	Text only
xAI (Grok)	Yes	grok-2	Text only
Groq	Yes	llama-3.3-70b-versatile	Fast inference
Mistral	Yes	mistral-small-latest	Text only
Together AI	Yes	meta-llama/Llama-3-8b-chat-hf	Open-source models
Alibaba Qwen	Yes	qwen-plus	Text only
Ollama	No	llama3	Local, requires Ollama running
LM Studio	No	(user-set)	Local, requires LM Studio running
vLLM	No	(auto-fetched)	Local, requires vLLM server
Custom	Varies	(user-set)	Bring your own endpoint

Commands

Translation

• Translate multiple pages... — Open the batch translation modal (page range, start/cancel)
• Translate and add overlay to current PDF page — Quick single-page translate
• Reprocess/retranslate a text region — Re-translate a selected region (shift+drag)
• Retranslate using saved overlay layout... — Bulk re-translation from saved overlays

Overlay Management

• Save current PDF overlay — Save current page's overlays to disk
• Refresh current PDF overlay — Reload overlays from disk
• Clear current PDF overlay — Remove all overlays from current page
• Toggle PDF overlay visibility — Show/hide all overlays globally
• Toggle BBox Edit Mode — Enable/disable overlay editing (non-invasive)

OCR

• OCR: recognize PDF to translated note — OCR with translation (multi-page)
• OCR: recognize current page — Single-page OCR

Background Queue

• Background translation: open watched-folder queue — Open watcher modal (Card Stack layout: Active + Available sections)

Layout

• Layout: extract entire PDF (background) — Extract layout data in background
• Layout: create file with originals — Create originals-only translation file

Export & Maintenance

• Export PDF with translations — Desktop-only, requires Python
• Repair translation links — Fix pdf-source frontmatter
• Repair translation file — Re-parse and re-write translation files (triggers V3→V4 migration)
• Rebuild PDF-to-translation file map — Refresh internal map
• Clean unused translation files — Find and delete orphaned files

Settings

The plugin has 14 settings sections:

1. Provider — Select and configure your LLM provider (15 providers supported)
2. Language — Source and target languages (auto-detect supported)
3. Translation — Batch size (maxBatchChars), concurrency, retry settings, reasoning mode
4. Storage — Where translation files are saved (.translations.md V4 format)
5. Prompts — Custom system prompts for translation (single + batch templates, supports {TEXT}, {sourceLang}, {targetLang} placeholders)
6. Visual — Overlay appearance (font size scale, opacity, line height, colors)
7. Layout Engine — Internal vs Python, detection parameters (indent threshold, font tolerance, merge passes, column gap, decoration threshold)
8. Watcher — Auto-translation of dropped PDFs, watched folder path
9. OCR — Vision model configuration + Advanced settings (DPI, image quality, pacing)
10. PDF Export — Python script paths
11. UI Behavior — Context menu, hover behavior, debug mode, layout debug mode
12. Copy Formats — How copied translations are formatted (callout, citation, footnote templates)
13. Paragraph Filter — Skip page numbers, single letters, URLs (configurable rules with regex)
14. Advanced — Debug mode, layout debug mode, cache settings

License

GLM-3.0 License — see LICENSE file for details.

Bundled Python Scripts

This plugin includes two Python scripts (layout_engine.py and pdf_export.py) as embedded base64 resources. These are required for the Python layout engine and PDF export features (desktop-only). Both scripts are licensed under the GLM-3.0 License.

Privacy

• API keys are stored locally in Obsidian's plugin data folder (data.json)
• PDF content is sent to your configured LLM provider for translation
• No telemetry or analytics are collected by this plugin
• Translation files (.translations.md) are stored in your vault — you own all data