turbo-parsepdf

Fast native PDF text / table / image extraction for Node — a pure-Rust core (N-API). Output as a structured object, HTML, Markdown, or JSON.

npm install turbo-parsepdf

Benchmark vs pdf.js

Wall-clock to extract every page's text, best-of-N (Apple M-series, release). Reproduce: node benches/competitive/bench.mjs (after python3 benches/gen-corpus.py).

| document | turbo-parsepdf | pdf.js (pdf-parse) | |---|---|---| | 100 pages | 5.6 ms | 54 ms · 9.7× faster | | 20 pages | 1.0 ms | 8.3 ms · 8.3× faster |

Extracted text is byte-identical to PyMuPDF (100% word recall) — fast and correct.

import { parse, parseToMarkdown, parseToHtml, parseToJson } from "turbo-parsepdf";
import { readFileSync } from "node:fs";

const pdf = readFileSync("doc.pdf");

const doc = parse(pdf);
// { version: "1.7", pages: [{ width, height, needs_ocr,
//   lines: [{ text, x, y }], tables: [{ rows, cols, cells }],
//   images: [{ name, format, width, height, bits_per_component, color_space }] }] }

parseToMarkdown(pdf); // string — reading text + pipe tables, page-broken
parseToHtml(pdf);     // string — one <section> per page
parseToJson(pdf);     // string — the doc above, pretty-printed

// Encrypted PDFs: pass the user or owner password.
parse(readFileSync("locked.pdf"), "secret");

A fatal parse fault throws an Error whose message carries a stable code (InvalidHeader, BadStream, …). Image-only/scanned pages come back with needs_ocr: true and no text (OCR is out of scope).

| feature | support | |---|---| | xref tables + streams, object streams, hybrid, incremental | ✓ | | Flate / LZW / ASCII85 / ASCIIHex / RunLength + predictors | ✓ | | /ToUnicode, WinAnsi/MacRoman/Standard + Differences (AGL), Type0/CID | ✓ | | ruled tables, image XObjects (JPEG passthrough + raw) | ✓ | | encryption: RC4 + AES-128/256, R2–R6, empty/user/owner password | ✓ |

Part of the turbo-parsepdf workspace (also on PyPI as turbo-parsepdf and as turbo-parsepdf-wasm). MIT.