@pdfmuse/core

WebAssembly build of pdfmuse — a deterministic PDF/DOCX parser for RAG / LLMs. No native binary: the same package runs in the browser (the file never leaves the tab) and on Node, including CommonJS-only hosts where a .node addon is not an option.

npm i @pdfmuse/core

Usage

Browser / ESM — the wasm is fetched, so init() is async:

import init, { to_text, parse } from "@pdfmuse/core";
await init();

const bytes = new Uint8Array(await file.arrayBuffer());
const text  = to_text(bytes);                  // plain reading-order text
const doc   = JSON.parse(parse(bytes));        // full IR: chars/blocks with bboxes

Node / CommonJS — the wasm is read from disk and instantiated at require time, so there is no init() step:

const { to_text, to_markdown, parse } = require("@pdfmuse/core");

const bytes = new Uint8Array(require("fs").readFileSync("resume.pdf"));
const md    = to_markdown(bytes);              // headings + tables preserved

Node ≥ 18. The module uses the reference-types and multivalue wasm proposals, which Rust enables by default for wasm32-unknown-unknown and current wasm-bindgen requires. Node 16 rejects it with invalid value type 'externref', Node 14 with return count of 4 exceeds internal limit of 1.

Use this build when a native addon won't load — bundled serverless runtimes, sandboxed platforms, anywhere the deploy artifact is JS-only. If you can ship a binary, @pdfmuse/node is faster.

Extracts text with exact coordinates, tables and structure; deterministic, byte-identical to the Python/Node/Rust bindings. Scanned/image-only pages return a NeedsOcr warning to hand off server-side.

• Live playground: https://casperkwok.github.io/pdfmuse/
• Docs & source: https://github.com/casperkwok/pdfmuse

MIT OR Apache-2.0