Node & browser

The parser core (src/parser.ts) is runtime-agnostic: it takes an ONNX InferenceSession, the runtime’s Tensor constructor, and a sub-word tokenizer, then builds the model’s word-grid input and decodes the raw tensors by argmax. Platform specifics live only in the two entry points.

Node

import { NlpGraph } from 'nlpgraph';

const nlp = await NlpGraph.load();               // onnxruntime-node (native), model auto-resolved
const doc = await nlp.parse('The controller shall notify the supervisory authority.');

load({ modelDir }) points at a directory containing config.json, the tokenizer files, and model.fp16.onnx. Omit it to use the resolve-or-download behavior described in Getting started.

Browser

import { NlpGraph } from 'nlpgraph/browser';      // onnxruntime-web (WASM / WebGPU)

const nlp = await NlpGraph.load({
  model: '/model/model.fp16.onnx',                // URL or bytes
  config: '/model/config.json',                   // URL or object
  tokenizer,                                       // a transformers.js PreTrainedTokenizer, or tokenizerId
});
const doc = await nlp.parse('Access rights must be reviewed annually.');

onnxruntime-web also runs under Node (WASM), which is how the browser path is tested without a browser. This very page’s live demo is the browser backend: the model and the ONNX WASM runtime stream from Cloudflare R2, parsing happens client-side, and nothing is sent to a server.

Determinism

Decoding is greedy argmax with no randomness, so the same input yields byte-identical output on every run and on both runtimes — important for reproducible compliance/reg-tech pipelines.