mupdf4llmPDFs → LLM-ready Markdown

TypeScript port of pymupdf4llm on top of the official mupdf WASM package. Tracks the Python upstream closely — exact on synthetic fixtures, with documented divergences on real-world PDFs.

Get started

View on npm

GitHub

Drop-in for pymupdf4llm

Same algorithms as the Python reference (`pymupdf4llm.helpers.pymupdf_rag.to_markdown`). Validated by parity tests that diff every fixture against the live Python output.

Pure JS, no native build

Runs on Node 20+ and Bun out of the box. The only runtime dependency is the official `mupdf` WASM package from Artifex.

Headers, tables, images, forms

Font-size or TOC-driven headers, five table strategies (including OCR of scanned tables), image extraction or base64 embedding, AcroForm field dump, per-word coordinates — all via the same `toMarkdown()` entry point.

LlamaIndex-ready

Optional `@nalinor/mupdf4llm/llama` subpath export ships a `PDFMarkdownReader` that yields one `Document` per page, with `llamaindex` as an optional peer dependency.

Why a TS port?

The Python pymupdf4llm is excellent, but RAG pipelines increasingly live in Node / Bun / browser-adjacent stacks. Spinning up a Python sidecar just to call to_markdown is friction. This package gives you the same output, the same options, and the same fixtures — straight from the JavaScript runtime you already have.

Install

npmbunpnpm

npm i @nalinor/mupdf4llm

bun add @nalinor/mupdf4llm

pnpm add @nalinor/mupdf4llm

30-second tour

import { toMarkdown } from "@nalinor/mupdf4llm";
import { readFileSync } from "node:fs";

const md = await toMarkdown(readFileSync("paper.pdf"));
console.log(md);

Need per-page chunks for RAG? Switch to toMarkdownPages:

import { toMarkdownPages } from "@nalinor/mupdf4llm";

const chunks = await toMarkdownPages(readFileSync("paper.pdf"), {
  extractWords: true,
});
for (const c of chunks) console.log(c.metadata.page, c.text.length);

See the quick start for more.