sigocr

sigocr

Fast PDF text extraction with paragraph layout and bounding regions.

Install

pnpm add sigocr

Usage

import { sigocr } from "sigocr";

// Extract structured text from a PDF (returns null if scanned/no embedded text)
const doc = await sigocr.pdf("/path/to/file.pdf");
// -> Document | null

// Extract from a buffer
const doc = await sigocr.buffer(pdfBuffer);

// Check if a PDF has embedded text (fast, doesn't extract)
const has = await sigocr.hasText("/path/to/file.pdf");
// -> boolean

Batch extraction

Extract many files in a single native call. Each file uses internal page-chunk parallelism via Rayon.

const docs = await sigocr.files([
  "/uploads/contract.pdf",
  "/uploads/invoice.pdf",
  "/uploads/report.pdf",
]);
// -> (Document | null)[]

Output shape

The output shape:

interface Document {
  content: string;                    // full text in reading order
  pages: Page[];
  paragraphs: Paragraph[];
  tables: Table[];
}

interface Paragraph {
  content: string;
  role?: string;                      // "title" | "sectionHeading" | "pageHeader" | "pageFooter" | "pageNumber"
  spans: Span[];                      // character offset + length into Document.content
  boundingRegions: BoundingRegion[];  // page number + polygon in points
}

Benchmarks

Measured on Apple M3 Pro, Node.js v24.

Structured extraction (with positions) - 200-page PDF

Library	Mean	ms/page	vs sigocr
sigocr (native)	16.5ms	0.08	1x
pdf.js-extract (JS)	71.0ms	0.36	4.3x slower
pdf2json (JS)	280ms	1.40	17x slower

Plain text extraction (no positions) - 200-page PDF

Library	Mean	ms/page
pdf-parse (JS)	51.6ms	0.26
unpdf (JS)	63.9ms	0.32

Batch extraction - 100 x 3-page PDFs

Library	Mean	vs sigocr
sigocr.files (native batch)	17.2ms	1x
sigocr.pdf (sequential loop)	97.3ms	5.7x slower
pdf.js-extract (sequential loop)	134.0ms	7.8x slower

hasText detection

	Mean
sigocr hasTextBuffer	1.7ms

sigocr is 4.3x faster than pdf.js-extract and 17x faster than pdf2json while producing richer output: grouped paragraphs with roles, bounding regions, and document-level spans. pdf.js-extract gives raw text items that still need assembly. sigocr is also 3.1x faster than plain-text-only extractors despite doing strictly more work. Batch extraction distributes files across cores via Rayon - 100 PDFs in 17ms.

Run benchmarks yourself:

pnpm bench

How it works

• null for scanned PDFs: No OCR engine - this library only handles embedded text
• Pure Rust, no C deps: Uses pdf_oxide for character-level extraction. No PDFium/MuPDF binaries to ship. ~1 MB package
• Parallel page chunks: Large PDFs are split into chunks processed in parallel via Rayon, each opening its own PDF instance
• Paragraph roles via heuristics: Font size for heading detection, page position for header/footer/page number
• hasText() fast path: Check if a PDF has embedded text without full extraction. Checks first 3 pages only
• UTF-16 span offsets: Span offsets and lengths are in JavaScript string units for direct use with String.slice()