Native OCR or privacy extraction: which path fits?

Reading a document is not the same as protecting its text

OCR converts visual content into text and structure. The privacy barrier processes personal data in content it can inspect. An image sent directly to a model is not protected merely because another part of the conversation passed through text rules.

Choose the path before integrating. If the organization requires processing before transmission, check protected extraction and available capabilities. If you need a provider-native contract, assess authorization to send the original document.

When to use native OCR

The /v1/providers/mistral/v1/ocr path preserves the body accepted by Mistral, including the model and supported provider options. Input can reference a document or image in the native format, or an owned file registered in the gateway.

This path requires an authorized connection and native consent. It sends the original document without the privacy barrier. An organization that requires the barrier receives 403 before transmission. The gateway preserves provider errors rather than modifying the request to make it pass.

When to use protected extraction

The /v1/privacy/extract endpoint is the entry for obtaining processed text before using protected chat. Check supported formats, maximum size and capacity availability in the API. The existence of a route does not mean a worker is available at that moment.

Start with a representative fictional document. Check reading order, tables and detected categories. OCR can misread characters and structure; personal data detection has limits too. A successful response does not establish perfect extraction or automatic legal compliance.

Frequently asked questions