Lipi OCR | The Software Co.

The Privacy Problem

Most web OCR utilities compromise on privacy by uploading sensitive legal, regulatory, and financial documents to remote cloud servers, creating massive compliance and security liabilities.

Zero-Server Execution

All document rasterization, WebAssembly OCR workers, and document assembly execute 100% inside local browser memory sandbox. Once loaded, base OCR functions completely air-gapped and offline.

Strict 1:1 Page Fidelity

Input Page N maps directly to Output Word (.docx) Page N, preserving page breaks, orientation (portrait vs landscape), margins, and headings without paragraph bleed.

Engineered for multi-page legal and regional records.

Linguistics

Full Pan-Bharatiya Script Coverage

Supports all 22 official Eighth Schedule languages of the Republic of India with automated Unicode script detection across Devanagari, Bengali, Dravidian, Gujarati, Gurmukhi, Odia, Perso-Arabic, Ol Chiki, and Latin.

Dual-Engine

Hybrid Cocktail Architecture

Layer 1 runs parallel local Tesseract.js WASM workers on your CPU cores. Layer 2 offers optional vision model refinement via Google Gemini Multimodal Vision API with an automatic fallback cascade (gemini-3.6-flash → gemini-2.5-flash → gemini-2.0-flash → gemini-1.5-flash).

Word Precision

Strict 1:1 Word (.docx) Page Fidelity

Every page is isolated in an OpenXML SectionType.NEXT_PAGE section. Landscape scans automatically receive PageOrientation.LANDSCAPE, while dense pages auto-scale typography (down to 8.5pt) and tighten margins to prevent spillover.

Formats

Broad Input & Streaming Export

Accepts scanned PDFs (1 to 300+ pages with streaming memory management), Multi-page TIFF (.tif/.tiff), PNG, JPG, JPEG, and WebP. Exports to Microsoft Word (.docx), standalone offline HTML viewer, Markdown (.md), and plain text (.txt).

All 22 Eighth Schedule languages of Bharat.

Script Family	Supported Languages	Classification Mode	Target Documents
Devanagari	Hindi, Marathi, Sanskrit, Nepali, Bodo, Dogri, Konkani, Maithili	Automated Unicode block detection	Legal deeds, revenue records, gazettes
Bengali & Assamese	Bengali, Assamese, Manipuri (Meitei)	Phonetic ligature mapping	Land registry, municipal records
Dravidian Scripts	Tamil, Telugu, Kannada, Malayalam	Sub-word character clustering	Judicial orders, civil certificates
Gujarati	Gujarati	Compound conjunct parsing	Commercial contracts, notarized deeds
Gurmukhi	Punjabi	Diacritic & tone mark alignment	Agricultural titles, identity papers
Odia	Odia	Curved glyph ligature segmentation	Historical records, government gazettes
Perso-Arabic	Urdu, Kashmiri, Sindhi	RTL directional stream parser	Waqf documents, historic certificates
Ol Chiki	Santali	Tribal Unicode script engine	Cultural & indigenous documentation
Latin	English and mixed multilingual records	Multi-lingual interleaved OCR	Corporate filings, hybrid court orders

Memory-managed client streaming pipeline.

Processing 300-page scanned PDFs in the browser can easily crash tab memory. Lipi implements streaming page rasterization, recycling high-res page blobs immediately after OCR extraction.

[Input Document: Scanned PDF / Multi-Page TIFF / Image] | v +---------------------------+ | Client-Side Document | | Streaming Extractor | +---------------------------+ | +-------------------+ | | v v [High-Res Page Blob] [Low-Res Thumbnail] | | v v +---------------------------+ | | Unicode Script & Language | | | Classifier | | +---------------------------+ | | | v | +---------------------------+ | | Parallel WASM Worker Pool | | | (Hardware Concurrency) | | +---------------------------+ | | | v | [Raw Base OCR Draft] | | | +---------+---------+ | v +-------------------------------+ | Layer 2: Gemini Flash AI | (Optional user-enabled) | Multimodal Vision Refinement | (3.6 -> 2.5 -> 2.0 -> 1.5) +-------------------------------+ | v +-------------------------------+ | Strict 1:1 Page Exporter | | (DOCX / HTML / MD / TXT) | +-------------------------------+

Your documents stay strictly in your hands.

Zero Storage

Zero Remote Storage

Documents are never uploaded to any remote server or third-party database. All document rasterization and analysis happens exclusively in client memory and is purged on tab close.

Air-Gapped

Air-Gapped Base Engine

The base WASM engine and PDF workers are bundled locally in dist/assets/. Once loaded, base OCR functions completely offline without any active internet connection.

Key Privacy

Local API Key Storage

When using Layer 2 AI refinement, your Google Gemini API key is stored strictly in your browser's private localStorage and sent directly to Google's API endpoint over HTTPS.

How Lipi processes high-volume documents.

01 · Load

Stream Ingest

Drag in scanned PDFs, multi-page TIFFs, or photos without full-file RAM buffering.

02 · Classify

Detect Script

Unicode classifiers identify regional Indian scripts and configure OCR dictionaries.

03 · Extract

Parallel WASM

Multi-core WebAssembly workers extract raw text with optional Gemini vision refinement.

04 · Export

1:1 Word DOCX

Download a clean Word document with exact section breaks, headings, and margins.