Entity types - FrostWolf

Each detection has a type and a category. You can choose which types to look for with entities, and choose a strategy per type or per category.

{ cloak: { context, entities: ["email", "payment_card", "phone_number"] } }

An unknown type is rejected up front rather than quietly detecting less than you asked for.

Categories

Category	Meaning
`pii`	Personal data and identifiers
`pci`	Payment card data
`credential`	Secrets that are not personal data but do not belong in a prompt
`phi`	Health data. Reserved. Nothing is reported under it yet.

Types

Type	Category	On by default
`person_name`	`pii`	Yes
`date_of_birth`	`pii`	Yes
`email`	`pii`	Yes
`phone_number`	`pii`	Yes
`address`	`pii`	Yes
`location`	`pii`	No
`government_id`	`pii`	Yes
`passport_number`	`pii`	Yes
`drivers_license`	`pii`	Yes
`tax_id`	`pii`	Yes
`bank_account`	`pii`	Yes
`iban`	`pii`	Yes
`routing_number`	`pii`	Yes
`payment_card`	`pci`	Yes
`card_expiry`	`pci`	Yes
`card_cvv`	`pci`	Yes
`username`	`pii`	Yes
`ip_address`	`pii`	Yes
`account_id`	`pii`	Yes
`sensitive_date`	`pii`	Yes
`date`	`pii`	No
`password`	`credential`	Yes
`secret`	`credential`	Yes
`recovery_code`	`credential`	Yes

location and date are off by default because they fire on ordinary prose. A city or a document date is in most sentences you write, and redacting them by default would leave a model with nothing to reason about. Name them in entities to include them. secret covers API keys and access tokens as well as other secrets.

Confidence

Detection returns a score for each entity. minConfidence (0 to 1) sets the lowest you accept. The default is set by the control plane.

Notes on types

A US social security number is reported as tax_id. It is still detected and protected, only the type name is arguable.
A street and a postcode may be reported as two address entities. The city between them is left alone unless location is enabled. Enable location if you want the whole address covered.
A medical record number is reported as account_id.