Playground

Paste any text and click Detect to see the spans the openai/privacy-filter model finds. Click Redact to replace each span with its category marker. Everything runs in your browser — the text never leaves the page.

1. Configuration — change anything, then click Reload

Model source URL (an HTTP origin that hosts onnx/model_q4f16.onnx + config.json — usually HuggingFace Hub or a mirror)

Enabled categories (fixed by the model — these are the 8 BIOES classes openai/privacy-filter was trained on)

private_person private_email private_phone private_address private_url private_date account_number secret categories are fixed by the upstream model. Unchecking just drops those spans from results — adding new categories would require retraining.

Markers (JSON: label → replacement. Example pre-filled — leave blank to use defaults like [private_email]) { "private_email": "[EMAIL]", "private_phone": "[PHONE]", "secret": null } String → replace span with that string. null → keep original text untouched (useful for "detect-only" categories). Omit a label → fall back to [label].

Reload model with this config

2. Input text Hi support team, My name is John Smith and I'm writing on behalf of my mother, Margaret Smith, who lives at 456 Oak Avenue, Apt 3B, Boston, MA 02134. She's been a customer since March 12, 2018, and her account number is 9876543210. Yesterday around 2:45pm she received a phone call from someone claiming to be from your fraud team — they asked her to confirm her bank routing number (021000089) and the last four digits of her SSN (6789). She gave them the info before realising it was probably a scam. Please reach out to me directly at (617) 555-0199 or john.smith@example.com so we can sort this out. You can also reference our earlier ticket at https://support.example.com/tickets/TKT-48211. For verification, the password reset link she clicked was https://login.example.com/r?t=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.fake-jwt-payload. We've already disabled the card ending 4242 at her request. Thanks, John (cell: 617.555.0123, work: jsmith@acme.io)

Detect Redact idle

3. Result

Equivalent code — same call you'd make from your own JS/TS project (works in browsers + Node from the single textsift ESM entry; no separate textsift/browser import is needed)

Raw result — the full object your code receives. detect() returns { spans, summary, ... }; redact() returns the same plus redactedText, containsPii. Each span has label, start/end character offsets, the matched text, the marker it'd be replaced with, and a confidence score 0..1.

4. Storage — model weights cached on this device

Refresh Wipe cached weights

…

What’s happening

The first click loads the textsift JS bundle (~2.6 MB) and the underlying model_q4f16.onnx from the Hugging Face Hub (~770 MB). Subsequent clicks are served from the browser’s HTTP cache, so the slow path only runs once per browser. WebGPU is auto-selected when available; the Backends page covers fallbacks.

The default backend selection mirrors what PrivacyFilter.create({}) does in your own code — see Quickstart and the API reference.