Garuda

An offline-first PDF + Image Text Extraction tool built with SvelteKit.

Features

• Extraction of selectable text layers from PDFs.
• OCR for scanned pages (image-only PDFs)using Tesseract.js.
• Support for PNG, JPEG, and WebP.
• On-device storage for document management (import, list, delete).
• Search across your entire "local library" (titles and extracted text).
• Best-effort detection of tables (CSV export) and AcroForm fields.

Architecture

Concurrency Model

• Main Thread: Manages UI state and user interactions.
• Worker Threads: Dedicated workers for PDF parsing (PDF.js) and OCR (Tesseract).
• Orchestration: A central orchestrator sequences tasks, manages cancellation, and reports progress to the main thread.

Offline Strategy

• PWA: Service worker caching for the application shell and OCR language data.
• Local Persistence: All document metadata, original blobs, extraction results, and derived artifacts are stored in IndexedDB.
• Search Index: A local-only index built from extracted content, rebuildable on startup.

Data Model

Document Persistence (IndexedDB)

• documents: Metadata including kind (pdf/image), name, size, and contentHash.
• blobs: Original file content associated with a document ID.
• runs: Execution history for extraction modes (pdf_text, ocr_image, ocr_pdf).
• pages: Extracted text and stats per page index.
• tables: Structured row/column data and confidence scores.
• form_fields: AcroForm metadata (name, value, type).

Tech Stack

• Framework: SvelteKit in SPA mode.
• Parsing: PDF.js for document rendering and text extraction.
• OCR: Tesseract.js (WASM) for client-side optical character recognition.
• Storage: Dexie.js for IndexedDB management, providing schema versioning and migrations.
• Styling: Tailwind v4

Principles

• Local-First: All processing happens on your device. No data is ever uploaded to a server.
• Deterministic: Extraction logic (table clustering, text normalization) is designed to produce consistent results for identical inputs.
• Deduplication: Files are identified by SHA-256 hashes to prevent redundant storage and extraction runs.