neural-formmap

A tiny local model that maps messy form labels and attributes to standard HTML autocomplete field names.

import { mapField } from "neural-formmap";

mapField({
  label: "Where should we send updates?",
  name: "contactEmail",
  placeholder: "[email protected]",
});
// { type: "email", autocomplete: "email", source: "model", ... }

CLI

Pass a plain label:

npx neural-formmap "Work email"

Or pass a JSON field descriptor:

npx neural-formmap '{"label":"Apt / Suite","name":"unit"}'

Piped input is read as JSON Lines. A JSON array maps a complete form.

How it works

The mapper follows a strict evidence order:

1. A valid autocomplete attribute wins.
2. Unambiguous HTML types such as email, tel, url, and search win.
3. A dependency-free 128 → 24 → 22 neural network classifies hashed features from the label, name, id, placeholder, input type, and nearby context.

The model has 3,646 parameters and runs locally. There is no server, telemetry, inference framework, or runtime dependency.

API

• mapField(field | label) maps one field and returns the best type, confidence, evidence source, and three alternatives.
• mapForm(fields) maps a list while preserving its order.
• FIELD_KINDS exposes the supported vocabulary.

The vocabulary covers names, credentials, contact information, addresses, payment fields, search, and unknown. It follows HTML autocomplete names where an equivalent exists.

Scope

This package classifies field purpose. It does not read the DOM, fill values, store personal data, or bypass form validation. Integrations should gather accessible labels and attributes, then decide whether a confidence threshold is sufficient for their use case.

The checked-in model is trained and evaluated on deterministic synthetic variations. The proof is reproducible, but it does not establish accuracy on arbitrary production forms, languages, or adversarial markup. See MODEL_CARD.md.

Development

Requires Bun 1.3 or newer.

bun run train
bun run check
bun run benchmark

CI retrains the model and verifies that src/weights.ts is byte-for-byte reproducible. See docs/prior-art.md for the project boundary.

License

MIT