When you type or drop a document, Kaviar spots sensitive information on your computer and replaces it before anything is sent to the AI. Nothing you type leaves your device to be analysed. This page explains what happens under the hood, and why two mechanisms work together.
Two passes, one result
No single system catches everything. Kaviar therefore combines two that complement each other.
- The instant pass. Precise rules recognise known formats: email addresses, phone numbers, IBANs and bank cards (with checksum validation), official numbers, dates, amounts, postal addresses, plus tens of thousands of first names, cities and companies. It reacts to every keystroke, within milliseconds: it is what highlights your text as you write.
- The second read. A small entity-recognition model, also running on your device, re-reads the text a fraction of a second later. It has no list: it recognises the shape of a person’s or company’s name in context, even one it has never seen: a rare surname, an unknown SME, a foreign name. Whatever it finds is added to the highlights, with no visual distinction.
The rules always have the last word. The second read never removes or reclassifies what the rules decided; it only adds what they missed. An IBAN, a date or an amount never depends on the model.
What each level enables
Standard covers your structured personal data only (emails, phones, addresses, IBAN, cards, official numbers): rules only.
Reinforced, the default, adds people, the companies around you professionally, and dates. The second read looks for names of people and organisations.
Strict adds well-known brands, places, GPS positions and amounts. The second read also looks for places.
Nothing leaves before the read is done
If you press Enter while the second read is still running, Kaviar waits for it to finish before sending. The wait is usually imperceptible and never exceeds a second and a half once the model is loaded; it guarantees that no name spotted at the last moment leaves in the clear.
What if it gets it wrong?
A model recognises shapes, not meaning: it may mask one word too many (a brand inside a common phrase, for instance). Hover the highlighted word and click Leave visible: Kaviar remembers it for the conversation. Conversely, if something seems missed, add it to the Vault: it will always be protected, in every conversation.
On your device, truly
The model ships with Kaviar (about 190 MB, 21 languages) and runs inside the app, offline. No data is sent anywhere to be analysed, neither to Kaviar nor to a third party. In the browser, the model is downloaded once and kept by the browser; if it is unavailable, Kaviar relies on the instant pass alone.