docutext

Zero-dependency PDF text extraction built for RAG and AI pipelines. Parses PDFs from scratch -- no PDF.js, no WASM, no native addons. Works in Node.js and the browser.

Performance

Library	Small PDF (31 KB)	Large PDF (1.3 MB)	Bundle (gzip)	Dependencies
docutext	3 ms	40 ms	~24 KB	0
pdfjs-dist	5 ms	244 ms	~1.3 MB	0 (but large)
pdf-parse	5 ms	279 ms	~780 KB	1 (pdfjs)
unpdf	6 ms	232 ms	~320 KB	2 (pdfjs)

Median of 3 runs, Node.js, Apple Silicon.

docutext is purpose-built for text extraction in RAG/AI workflows -- not a general PDF toolkit. It does one thing and does it fast.

Install

# Node.js (zero dependencies)
pnpm add docutext

# Browser / bundler (add fflate for decompression, ~3 KB gzip)
pnpm add docutext fflate

Requires Node.js 18+. Browser builds target ES2020+.

Quick Start

import { DocuText } from 'docutext';

const doc = await DocuText.load('document.pdf');
console.log(doc.text);

Structured Markdown

import { DocuText } from 'docutext';
import { docToMarkdown, pageToMarkdown } from 'docutext/markdown';

const doc = await DocuText.load('document.pdf');
console.log(docToMarkdown(doc));       // headings, bold, links
console.log(pageToMarkdown(doc.pages[0]));

Browser

import { DocuText } from 'docutext';

const response = await fetch('/document.pdf');
const bytes = new Uint8Array(await response.arrayBuffer());
const doc = DocuText.fromBuffer(bytes);
console.log(doc.text);

Page-by-page

for (const page of doc) {
  console.log(`Page ${page.number}: ${page.text}`);
}

Layout-Fidelity Opt-In

import { DocuText } from 'docutext';

const doc = await DocuText.load('document.pdf', { textMode: 'layout' });
console.log(doc.text);

textMode: 'layout' keeps more literal spacing and text-object behavior. The default textMode: 'clean' favors semantic text reconstruction and fixes fragmented form PDFs.

Key Features

• Zero dependencies in Node.js. Single optional peer dep (fflate) for browser.
• ~24 KB gzipped browser bundle -- 50x smaller than pdfjs-dist.
• 6x faster than alternatives on real-world documents.
• Clean semantic text by default -- fragmented form PDFs are reconstructed without spurious intra-word splits.
• Plain text + structured markdown output (headings inferred from font size, bold/italic, links).
• Opt-in layout fidelity mode via textMode: 'layout' when you want more literal spacing/object ordering.
• Column-aware text flow -- side-by-side columns (e.g. signature blocks) are read column-first to keep related data together.
• Lazy extraction -- accessing page.text only processes that page.
• Full PDF parsing -- xref tables, stream filters, font encodings, ToUnicode CMaps, form XObjects.
• Encrypted PDF support -- handles permission-encrypted PDFs (empty password, RC4/AES-128).
• Runs everywhere -- Node.js and browser via conditional exports.
• ESM only -- modern module system, tree-shakeable.
• TypeScript first -- full type definitions included.

Documentation

For the full API reference, architecture details, and a live playground, see the documentation site.

Security

Please see SECURITY.md for vulnerability reporting guidance.

License

MIT