Use cases
Redaction is not a product you sit in front of. It is a call you put in front of something else — and which mode you want depends entirely on whether the document has to come back.
An agent handling a support inbox has to read the whole message, including the address and the card. Send it through here first, let the model reason over tokens, then rehydrate the reply before a person sees it.
Two calls per turn. The model never holds a real value, and the customer still gets an answer addressed to them.
The fastest way to have a breach is to log a request body. Mask the payload on the way into your observability stack and the incident becomes a shape rather than a disclosure.
mask returns no mapping, so a log line that leaks later leaks nothing.
A translation vendor, an annotation contractor, an offshore QA team. Tokenize on the way out, rehydrate on the way back, and the document is intact without the contract ever having covered personal data.
The mapping never leaves your side. We do not hold it and cannot be compelled to produce it.
Fine-tuning sets and eval suites outlive the incident that created them. Mask once at ingestion and the archive is safe to hold for years.
Read the not-detected list first — names survive masking, and a transcript is mostly names.
People paste .env files into chat. Vendor-prefixed keys, JWTs and whole PEM blocks are detected by shape, so a config dump can be scrubbed before it reaches a model or a ticket.
High-entropy strings with no known prefix are reported as unsure rather than removed — a hash and a secret look identical.
An access request means handing someone a bundle that mentions other people. Tokenize the third parties, keep the requester's own values in place by narrowing the entity list, and the redaction is reproducible.
Deterministic means two runs a month apart produce identical output — which is what makes it defensible.
This tool removes specific, well-defined patterns and tells you exactly which ones it can and cannot find. Whether that is enough under GDPR, HIPAA, PCI-DSS or your own contracts is a question for someone qualified to answer it — and the honest place for them to start is the list of what is never detected.
What it will not do