AI that never sees your documents: summaries and translations on your own device
NomadVault can now summarise notes, translate previews and produce translated copies of Office documents. The model runs inside your browser, so nothing leaves your device.
End-to-end encryption and AI assistants usually pull in opposite directions. The assistant wants to read your text; encryption exists so that nobody but you can. Most products resolve that by decrypting on a server, sending the plaintext to a model provider and asking you to trust a privacy policy. We did not want to make that trade.
Chrome ships a language model inside the browser. Since version 138 the Summarizer, Translator and Language Detector APIs are available on desktop, backed by Gemini Nano running locally. That gave us a way to add the two things people asked for most, summaries and translations, without changing the one thing we promise: plaintext exists only on your device.
What you can do now
- Summarise a note. Open a note, press Summarize, and insert the result at the top or replace the content.
- Summarise or translate a file preview. PDFs, Office documents, text and Markdown get the same bar in the viewer. The text is extracted from the decrypted file in your browser and handed to the local model.
- Create a translated copy of a document. For Word, PowerPoint and OpenDocument files NomadVault rewrites the text runs inside the document and keeps headings, lists, tables, images and slide layouts. The copy is encrypted like any other file and saved next to the original.
How it stays private
The file is decrypted in your browser, as always. The extracted text goes to the Chrome API, which runs the model on your machine. The result never travels anywhere unless you choose to save it, and then it is encrypted before upload like everything else. Our servers see exactly what they saw before: an encrypted blob, its size, and a timestamp.
There is no cloud fallback. If the browser does not provide a local model, the feature is simply not shown. We would rather have a missing button than a silent trip to a third-party API.
What to expect
- First use downloads the model. Chrome fetches Gemini Nano once, around two gigabytes, after a user gesture. NomadVault shows the download progress. After that it is instant and works offline.
- Desktop Chrome only, for now. Edge and other Chromium browsers are adding the same APIs. Safari and Firefox have nothing comparable yet.
- Language pairs vary. Translation packs are installed per pair. If a pair is not available on your device, NomadVault says so instead of guessing.
- Scanned PDFs have no text. The model needs text, and a scan is a picture. OCR on-device is on our list.
- Character-level formatting is approximate. A bold word in the middle of a sentence keeps the formatting of the first run of the paragraph in the translated copy. Paragraph and page structure survive intact.
Why we think this matters
“Private AI” has become a marketing phrase, and it usually means a contract rather than a design. Running the model where the plaintext already is removes the question entirely. There is no model provider to audit, no data processing addendum to sign, no retention window to argue about. The document stays where end-to-end encryption put it.
We will extend this as the browser APIs grow. If you want to try it, open any note in your tenant with a current desktop Chrome and look for the Summarize button.