Pass-through hatch for AI prompts

Your document stays here. Only tokens go through.

Gemini Nano reads the document on this device and swaps every name, number and address for a token. Cloud Gemini works with the tokens. The answer comes back and is filled in here, and the log shows exactly what left.

Try a client email How the hatch works
Cutaway of a pass-through hatch: the original document stays in the device room, a tokenized copy goes to the cloud room.

Device room

Nothing in this room leaves your browser.

01

Document

Hatch

Waiting for a document

Cloud room

What Gemini in the cloud receives. Exactly this, nothing more.

The tokenized copy appears here after the scan.

What left this device

Every run is written here, in this browser only. Cloud runs keep the exact text that was sent.

    How the hatch works

    1. 01

      Find, on the device

      Checksum-verified patterns catch SINs, cards, health numbers and IBANs. Gemini Nano, through Firebase AI Logic in ONLY_ON_DEVICE mode with a JSON schema, finds names, organizations and addresses. If Nano is missing, the patterns still run and the cloud never stands in for it.

    2. 02

      Swap for tokens

      Each value gets a stable token like [[PERSON_1]]. The table of real values lives in this tab’s memory and is gone on reload.

    3. 03

      Check before the door opens

      The guard re-scans the outgoing text for every real value and every pattern. One hit keeps the cloud door shut.

    4. 04

      Ask the cloud with tokens

      Firebase AI Logic in ONLY_IN_CLOUD mode sends the tokenized text to Gemini. Each answer shows its inferenceSource.

    5. 05

      Fill in, on the device

      Tokens in the answer are replaced with the real values here. Tokens the model made up are flagged, and the run goes to the log.

    Built on RedactGate, our earlier server-side redaction gateway. This version is new: the detection runs in the browser with Gemini Nano, and the cloud step goes through Firebase AI Logic.

    Run the demo