Settings
PII redaction
Every ingest path redacts before it stores, so only redacted text is ever stored, judged, reproduced or trained on.
#Placeholders
Each detected value becomes a per-conversation placeholder such as [EMAIL_1] or [PHONE_2], numbered by first appearance. The same value keeps its placeholder throughout the conversation, so judge quotes still verify and the same email stays recognizable. A value detected anywhere (a date of birth in a tool argument) is also replaced where its exact text appears unlabeled, except a postal code or account number where something else names it (an order_id field, "order no. 73301"): those have no checksum, so the same digits can be another identifier. A JSON tool result stays valid JSON. The mapping is never stored, and the conversation shows a redacted badge with how many values were replaced.
#Detectors
- Email, phone, IP
- Address syntax; valid phone numbers grouped as their country writes them (national numbers read in the workspace's region); IPv4 and IPv6.
- Card, SSN, IBAN
- Luhn-valid card numbers from a payment network; US SSNs outside never-issued ranges; IBANs with a valid mod-97 check.
- Street address
- US street lines. Off by default.
- Labeled values
- Dates of birth, postal codes, routing numbers and account numbers count only where the text or a JSON key names them, such as "date of birth is 1991-01-22", "ZIP 73301" or
account_number. A bare date or 5-digit number stays, so order numbers, amounts and appointment dates are not redacted. - National IDs
- Mexican CURP and RFC, Brazilian CPF and Colombian NIT, each with its check digits.
- Names
- Opt-in English and Spanish name models. They guess, so they are off by default; choose the model of your conversations' language.
- Custom patterns
- Up to 20 labeled regular expressions (RE2, so no pattern can stall ingestion), run after the built-ins.
The settings page's test box redacts pasted text under the unsaved settings and stores nothing. Owners and admins change redaction.
#When settings change
A changed setting applies to conversations ingested from then on. Stored conversations are never rewritten: post a conversation again under the same external_id to store it redacted under the current setting, which also judges it again.
Not detected: person names without a name model, labeled kinds without a label, bank account numbers outside IBAN and labeled ones, other national IDs and non-US street addresses. Reproductions replace every placeholder with a synthetic value before planning, so your tests hold no customer value either.