grobid-js

A pure JS / TypeScript port of GROBID — scientific PDF → TEI-XML extraction in Node and the browser, no Docker, no Java.

Status

• Wapiti CRF path: ties upstream Java GROBID byte-for-byte on the test corpus.
• BiLSTM_CRF_FEATURES path (opt-in): ONNX runtime, modest quality wins on author + affiliation extraction.
• Auto-download: Wapiti model weights and pdfalto binary are fetched on first run.

Install

npm install grobid-js

Quick start

import { Grobid } from "grobid-js/node";

const g = new Grobid();
const tei = await g.processPdf("path/to/paper.pdf");
console.log(tei);

Models and the pdfalto binary auto-download to ~/.cache/grobid-js/ on first call. To pin specific paths:

const g = new Grobid({
  modelsDir: "/path/to/grobid-home/models",
  lexiconDir: "/path/to/grobid-home/lexicon",
  pdfaltoOptions: { binaryPath: "/path/to/pdfalto" },
});

Opting into the BiLSTM (ONNX) header model

import { Grobid } from "grobid-js/node";
import {
  GloveFileProvider,
} from "grobid-js/grobid/engines/tagging/glove-file-provider";

const g = new Grobid({
  bidlstmModels: {
    header: {
      modelDir: "/path/to/header-BidLSTM_CRF_FEATURES",
      glove: new GloveFileProvider(vocabPath, vectorsPath),
    },
  },
});

Requires onnxruntime-node (optional peer dep) and a vocab-restricted Glove dump produced by scripts/extract-glove-vocab.py.

Architecture

Faithful port of the upstream Java layers:

• pdfalto: native binary for PDF token + bounding-box extraction (same binary upstream uses)
• Wapiti: pure-JS implementation of the Wapiti CRF decoder (src/grobid/jni/wapiti-*.ts), byte-equivalent to upstream's C decoder
• Engines: header, citation, segmentation, fulltext, affiliation, date, author name, reference segmenter, figure, table, funding-acknowledgement — all ported (src/grobid/engines/)
• Features: every per-token feature vector upstream produces (src/grobid/features/)
• Lexicon: same gazetteers and lookup tables
• TEI emission: same <teiHeader> / <text> / <biblStruct> structure

Building from source

npm install
npm run build
npm test

For the BiLSTM ONNX conversion path, see scripts/convert-bidlstm-crf-features.py.

License

Apache 2.0