writinglint-parser-node

The default backend is the owned compact ONNX dependency parser. It runs fully inside Node through onnxruntime-node; no Python process is required. The Stanza bridge remains available as a development oracle and explicit fallback.

The compact INT8 model is included in the npm package. A source checkout can stage the exact same verified files with npm run setup-model. Set WRITINGLINT_ONNX_MODEL or pass onnxModelDir only to override the bundled release:

import { loadParser } from 'writinglint-parser-node';

const parser = await loadParser({
  backend: 'onnx',
  onnxModelDir: '/path/to/another-model',
  // Optional: lower this on especially memory-constrained machines.
  onnxMaxBatchSentences: 16,
});

The runtime performs owned sentence/word segmentation, BERT WordPiece encoding, the main ONNX graph, deterministic valid-tree decoding, and selected-head relation scoring. Token offsets are document-global UTF-16 indices. Individual model calls remain bounded to 256 encoder subwords. Longer unpunctuated spans are split at token boundaries without dropping text; graph edges do not cross those defensive chunk boundaries. Documents are also parsed in sequential batches of at most 16 sentence chunks. This keeps ONNX tensor allocation bounded as a file or corpus grows instead of batching a whole large document into memory at once.

Stanza oracle

To explicitly use the development oracle, install its isolated Python environment and English models once:

npm run setup-stanza

Then use the parser normally:

import { loadParser } from 'writinglint-parser-node';

const parser = await loadParser({ backend: 'stanza' });
const sentences = await parser.parse('Trust the flags, not the number.');

The Stanza backend starts one long-lived Python process and communicates over JSON-lines. The model is loaded once rather than once per document.

For non-workspace installs, configure:

• STANZA_MODEL_DIR: downloaded Stanza model directory.
• WRITINGLINT_PYTHON: Python executable with stanza installed.

The parser output uses document-global UTF-16 offsets, matching JavaScript string indices exactly.

License

Runtime source code is MIT. The bundled model has an explicit file-level license boundary: its trained ONNX graphs are distributed under CC BY-SA 4.0 for conservative compliance with the UD English EWT training lineage, while the BERT-derived tokenizer retains its Apache 2.0 lineage. See MODEL_LICENSE.md for attribution and provenance.