A prompt sent to a hosted model leaves your infrastructure — along with the retrieved context, the tool output and the trace. Databunker Pro runs PII cleanup first, so plaintext never leaves the vault.
The Challenge
Anything personal in a prompt is held by a third party, on their retention schedule, in logs you do not control. Plain redaction breaks the task.The Solution — Four Capabilities
1. Typed placeholders
[EMAIL:uuid] tells the model what kind of value it stands for, so the task survives the PII cleanup.
2. Optional unique, safe to share
unique on, one person stays one person across turns; off, every mention is unlinkable. Either way the tokens are safe to hand to a third party.
3. Expiry without a cleanup job
finaltime retires a token when the conversation ends; slidingtime extends it on every access.
4. Tenant isolation and scoped access
PostgreSQL row-level security separates tenants; bulk resolve takes its own unlock step and is audited.
How It Works
PII cleanup runs on the way out and is reversed on the way back — one bulk call each direction.# the prompt your application built
Reply to sarah.chen@example.com re the $420 charge on card 4532015112830366.
# what reaches the model, after TokenCreateBulk
Reply to [EMAIL:cfd20fe2-8872-…] re the $420 charge on card [CARD:600c5f6d-…].The reply carries the same placeholders; BulkListTokens restores the values before it goes anywhere.
The Payoff
- No personal value in the provider’s logs or your traces.
- One person stays one person across every turn.
- Ship AI features without owning a plaintext mapping table.
- Cards come back the same length and still Luhn-valid.
Honest limits: detection sets the ceiling, and tokenized prompts remain personal data under GDPR.