Privacy filter

Privacy Filter is a token-level model that scans text for personally identifiable information (PII), such as names, emails, phone numbers, addresses, SSNs, secrets and more, and returns the detected spans together with the entity type.

createPrivacyFilter runs privacy-filter by OpenAI.

Quick start

import { createPrivacyFilter, models } from 'runntime/zoo';

// 1. Load the model once. The weights download from the Hugging Face Hub.
const filter = await createPrivacyFilter(models.privacyFilter.PRIVACY_FILTER.DEFAULT);

// 2. Scan a text.
const spans = await filter.detect('Contact Jane Doe at jane.doe@example.com.');
console.log(spans);
// [
//   { label: 'private_person', start: 8, end: 16, text: 'Jane Doe', placeholder: '<PRIVATE_PERSON>' },
//   { label: 'private_email', start: 20, end: 40, text: 'jane.doe@example.com', placeholder: '<PRIVATE_EMAIL>' },
// ]

// 3. Free the GPU memory when done.
filter.dispose();

The engine must be set up first, see Getting started. The weights are 1.55 GB, so pass a cache in the loading options to download them once.

Try it yourself

Your text through detect() on your GPU, every span it finds marked with its label. Redact swaps each one for its placeholder. Nothing leaves the page.

Methods

function createPrivacyFilter(
  model?: PrivacyFilterModel,
  options?: LoadOptions,
): Promise<PrivacyFilter>;

interface PrivacyFilter {
  detect(text: string): Promise<PrivacySpan[]>;
  dispose(): void;
}

• detect(text) returns the personal data it found, in text order. Empty text gives an empty list.
• dispose() frees the GPU memory.

Spans

Each span is one piece of personal data:

interface PrivacySpan {
  readonly label: string; // 'private_email', 'private_phone', ...
  readonly start: number;
  readonly end: number; // exclusive: text.slice(start, end) is the match
  readonly text: string;
  readonly placeholder: string; // '<PRIVATE_EMAIL>', ...
}

• label - the kind of data, one of private_person, private_email, private_phone, private_address, private_date, private_url, account_number, secret.
• start, end - where it sits in the text, as string indices. end is exclusive, so text.slice(start, end) is the match.
• text - the matched text.
• placeholder - a tag to put in its place, like <PRIVATE_EMAIL>.

Options

The first argument is the model config, a plain object with the field below. It is optional: createPrivacyFilter() with no arguments loads models.privacyFilter.PRIVACY_FILTER.DEFAULT, the official checkpoint with int8 weights, 1.55 GB, the one hosted format.

Pass it as is, or point at your own copy of the file:

interface PrivacyFilterModel {
  readonly modelPath?: ModelPath;
}

type ModelPath = string | RangeSource; // a URL, or bytes you already have

Examples:

// Your own copy of the file.
await createPrivacyFilter({ modelPath: '/models/privacy-filter/model.safetensors' });

• modelPath - the weights, a model.safetensors file: the hosted int8 export, or your own int8, f16 or bf16 export of OpenAI’s checkpoint. The weight format is read from the file.

Loading options

The second argument controls the download. Every field is optional.

interface LoadOptions {
  cache?: WeightCache;
  cacheId?: string;
  onProgress?: (name: string, doneBytes: number, totalBytes: number) => void;
  onBytes?: (chunkBytes: number) => void;
  signal?: AbortSignal;
}

Example:

import { createOpfsCache, createPrivacyFilter, models } from 'runntime/zoo';

const filter = await createPrivacyFilter(models.privacyFilter.PRIVACY_FILTER.DEFAULT, {
  cache: await createOpfsCache('my-app'),
  onProgress: (name, done, total) => console.log(`${Math.round((100 * done) / total)}%`),
});

• cache - keeps the downloaded weights in the browser, so the next visit loads them without the network. createOpfsCache(name) is the built-in one.
• onProgress - called while the weights upload to the GPU, with the tensor name, bytes done and bytes total. Use it for a progress bar.
• signal - an AbortSignal. Abort it to stop the load between steps. A step already running, like the download or the GPU upload, finishes first.

Supported models

Model	Config	Language	Parameters	Weights
privacy-filter	models.privacyFilter.PRIVACY_FILTER	English	1.5B, 50M active	int8, 1.55 GB