Before AI: replace personal data, redact it or block it?

Reduce the data sent to a model

A summary or classification request does not always need a person's name, phone number or identity document. On protected RN routes, the barrier examines content before the model call. It combines person-name detection with rules for personal data categories.

The purpose is to reduce exposure in the inference workflow. Detection has limits and is not a guarantee of irreversible anonymization. Validate results with representative fictional examples and retain your application's access controls.

Choose the organization behavior

• placeholder replaces detected data with a marker, such as [PESSOA_1] or [CPF_1], to preserve some context.
• redact removes the detected value. Use it when identification does not help the task.
• block refuses the request with 422 when the barrier detects personal data.
• monitor records detection but preserves content. It does not clean a request before transmission.

Choose a path that supports the content

Protected chat uses the rota/* catalog. The /v1/privacy/clean endpoint processes text without calling a response model. For documents or images, check /v1/privacy/extract and capacity availability before sending content. A text barrier cannot directly inspect image pixels or audio.

The /v1/providers/* gateway preserves the original format and does not apply this processing. An organization can require the barrier and refuse that path. A request policy can only strengthen the organization's policy. Conversation storage, external transmission and personal data processing remain separate decisions.

Frequently asked questions