redact

redact removes sensitive data before you log text or send it to a model. detect_pii finds it without changing anything. Detectors are validated (Luhn checksums, SSN ranges, digit counts, a real IPv6 parser) to keep false positives down, and overlapping matches are resolved by a fixed rule.

redact(text, *, policy="default", exclude=(), patterns=None, terms=None,
       replacement="[REDACTED:{label}]") -> str
detect_pii(text, *, policy="default", exclude=(), patterns=None, terms=None) -> list[dict]

>>> redact("Email me at alice@example.com")
'Email me at [REDACTED:email]'
>>> redact("Card: 4111-1111-1111-1111", policy="strict")
'Card: [REDACTED:credit_card]'

Try it

Validated detectors (Luhn, SSN ranges, digit counts) and deterministic replacements.

Textpolicyreplacement

Output

Loading beacon's Rust core (WebAssembly)…

IdleSame Rust core as the Python package, compiled to WebAssembly

Detectors

Label	Detects	Validation
`private_key`	PEM `-----BEGIN … PRIVATE KEY-----` blocks
`jwt`	`eyJ….eyJ….signature`
`bearer_token`	the token after `Bearer` (the scheme is kept)
`api_key`	OpenAI, Anthropic, GitHub, Slack, Google, Stripe, Hugging Face keys
`aws_access_key`	`AKIA…` / `ASIA…`
`email`	addresses, including internationalized local parts
`credit_card`	13–19 digits with optional spaces/dashes	Luhn checksum
`ssn`	`NNN-NN-NNNN`	no 000/666/9xx area, 00 group, 0000 serial
`phone`	`+CC …` international, `(NNN) NNN-NNNN`, `NNN-NNN-NNNN`	digit counts, no adjacent digits/letters
`ipv6`	IPv6 addresses	parsed by a real IPv6 parser
`ipv4`	dotted quads 0–255

Policies

Policy	Labels
`default`	everything except `ipv4`/`ipv6`
`strict`	everything
`secrets`	`private_key`, `jwt`, `bearer_token`, `api_key`, `aws_access_key`
`pii`	`email`, `credit_card`, `ssn`, `phone`
`none`	nothing (use with `patterns`/`terms`)

exclude turns off built-in labels: redact(text, exclude=["phone"]).

Your own detectors

redact(
    "EMP-123456 works on Project Falcon",
    patterns={"employee_id": r"EMP-\d{6}"},          # Rust regex syntax, compiled per call
    terms={"codename": ["Project Falcon", "falcon"]}, # whole words, ASCII case-insensitive
)
# '[REDACTED:employee_id] works on [REDACTED:codename]'

Labels must match [a-z][a-z0-9_]*.

Replacements

replacement may use {label} and {hash}, the first 8 hex digits of BLAKE3 of the matched text. With {hash} the same value always gets the same token, so you can still see that two mentions refer to the same person without seeing who it is:

>>> redact("alice@x.io wrote to bob@x.io; alice@x.io replied", replacement="<{label}:{hash}>")
'<email:4610c0e7> wrote to <email:72b1cf05>; <email:4610c0e7> replied'

Write {{ and }} for literal braces; any other placeholder is a ValueError.

Overlaps

All matches are sorted by start offset, then longer first, then priority (the table order above, then patterns in the order given, then terms). A match is kept only if it doesn’t overlap one already kept. This is how a credit-card number that contains a phone-shaped run is redacted once, as a card.