Skip to main content

Privacy first, by design

Tuteliq does not store your users’ content. Messages, images, audio and documents are analysed and discarded. Where Tuteliq has to keep some text of its own, such as the summary of an incident written after a detection, that text is redacted before it is written anywhere. Redactyl is the engine that does this, and you can call it on your own data: redact chat logs before they reach your moderators, your analytics, your support tools or a training set.
  • No third party sees the text. Redaction makes no call to any external service or large language model. It runs entirely inside Tuteliq’s infrastructure, using Tuteliq’s own model and rules.
  • Deterministic. The same input always gives the same output, so a result can be reproduced and audited.
  • Nothing kept. The endpoint returns the redacted text and a count per type. The text you send and the values removed are never stored or logged.

What Redactyl removes

Redactyl finds 23 types of personal data by default. Identifiers, recognised by their own rules. Some formats carry a check digit or checksum, and Redactyl verifies it, so an order number is not mistaken for a card number. The others are recognised by their structure, and the ones that look like any other number (an account number, a passport number) only when a label such as “account number” or “passport no.” comes just before them. A value recognised by structure or label alone has no check to confirm it, so an unusual format can be missed and a look-alike can be removed. Names and places, read in context. A proprietary multilingual model, trained by Tuteliq, finds names and places in all 32 supported languages. It reads the sentence, not a list: “Rose” can be a person or a flower, and “Paris” a city or a name.

Readable after redaction

Each value becomes a typed placeholder, such as [PERSON_1], [SCHOOL_1] or [PHONE_1]. Within one call the same value always gets the same placeholder, so a moderator can still follow who said what to whom without seeing the personal data. Numbering is per call and follows the order values first appear: the first person found is [PERSON_1], the next [PERSON_2]. A conversation redacted in two calls, or with its messages in a different order, can number the same person differently, so send a conversation in one call, in order, when you need its placeholders to line up.

Safety signals are kept

By default Redactyl removes what identifies a person and keeps what a safety review needs. A child’s age, school year, “my parents are away”, or a mention of the park after school stay in the text, because they are often the very details that show a child is at risk. Removing them would hide the risk from the people reviewing it. If you share text outside safety review, for example in analytics or a dataset, you can remove those details too with strict anonymisation.

Use Redactyl on your own data

counts always lists every type, including the strict-only ones, with zero for the types not found.

A whole conversation

Send messages instead of text to redact the messages of one conversation together. The same person or value gets the same placeholder in every message.
counts here counts every replaced occurrence, so PERSON is 3 for one person named three times.

Request fields

Send exactly one of text or messages.

Response fields

Modes

Both modes remove every type above in the same way. They differ only on two kinds of near miss, which are usually a mistyped real number:
  • a number shaped like a card, IBAN or national ID number whose check digit fails;
  • a phone number one digit longer or shorter than its format allows.
Every other type, including TAX_ID, PASSPORT and HEALTH_ID, is the same in both modes.

Strict anonymisation

Set strict_anonymisation: true to also remove the details a safety review would keep:
Strict anonymisation removes the details that show a child is at risk. Use it only for text that has already been through safety analysis and is being shared for another purpose. Never redact text this way before analysing it.
English only. The strict-only types above are recognised in English phrasing today. In text in any other language they are not removed, even with strict_anonymisation: true in the request and the response. The default types are removed in every supported language as usual. Before sharing non-English text as anonymised, check that it does not carry these details.

When model_applied is false

Name and place recognition has two layers. When model_applied is false, only the first ran, because the second could not be loaded:
  • The call still succeeds with 200 and costs 1 credit.
  • Every identifier type is removed as usual.
  • Names, places, addresses, schools and organisations are still found, but fewer of them: those written without a clear cue, such as a bare name, can be missed.
Treat that result as possibly under-redacted. Before it reaches people or a dataset, retry the call, or hold the text until a call returns model_applied: true.

Credits

Each call costs 1 credit, whatever the length of the text or conversation and whichever options you choose. Calls with a sandbox key are not charged. See Pricing & Credits.

Errors

Every invalid request is refused with 400 and error code VAL_3001, before anything is redacted or charged:
  • both text and messages, or neither;
  • a text with nothing but spaces;
  • a text, or any message, over 10,000 characters;
  • an empty messages list, or more than 100 messages;
  • a mode other than protective or precise.
Error messages never quote your input. Rate limits and running out of credits behave as on every endpoint: see Rate limits and Error handling.

Where Tuteliq uses it internally

Every summary, stored detection result and error report Tuteliq writes passes through Redactyl first. Contact details and identifiers are removed from all of them today. Name, place, school and organisation coverage is rolling out across the same storage paths.