omna-pii-mask

The compiled detection engine behind Omna's PII features. This is the same unified Rust kernel that ships in the Omna Mac app and browser extension, exposed to Python as a single self-contained wheel.

It is the backend for pip install "omna[pii]" — you normally don't install it directly; the omna package depends on it.

What it does

A six-layer (L1–L6) detection pipeline over text:

• Regex + rule layers — emails, phone numbers, cards, URLs, and 220+ secret rules (AWS keys, GitHub tokens, JWTs, …) with entropy checks.
• Checksum-validated IDs — Luhn, IBAN, Verhoeff, NHS, and 30+ international formats, so a string that merely looks like an ID isn't flagged unless it validates.
• On-device AI (L3) — an optional ONNX model that catches contextual PII regex can't (e.g. a bare prose name). Runs in-process on the same ONNX Runtime the wheel already links; downloads once on first use.

API

import omna_pii_mask

omna_pii_mask.mask("Email john@acme.com", model=False)   # -> {"masked": "...", "tokens": {...}}
omna_pii_mask.detect("SSN 123-45-6789")                   # -> [{"entity": ..., "confidence": ...}, ...]
omna_pii_mask.restore(masked_text, tokens)                # reverse the reversible tokens
omna_pii_mask.download_model()                            # pre-fetch the L3 model
omna_pii_mask.version()                                   # engine fingerprint (version+gitsha)

Reversible PII is replaced with stable tokens ([PERSON_1], [EMAIL_1], …). Secrets/credentials are always redacted irreversibly ([REDACTED:<KIND>]) and never written to the reversible token map.

Span start/end (in mask(...)["spans"] and detect(...)) are UTF-8 byte offsets, not Python string indices. Slicing text[start:end] is wrong as soon as a dash, emoji or accented letter comes before the span — use text.encode()[start:end].decode(), or mask(...)["masked"], which is already correct.

Distribution

Binary-only, proprietary. Wheels are built abi3 (one wheel per platform covers Python 3.9+). No source distribution is published — the Rust source is closed.