Use case · PII and PHI in prompts
A clinician pastes a discharge note. A developer pastes a customer record. Before the model sees the prompt, Raidu finds the personal data, replaces it with coded stand-ins such as PERSON_1, and sends the clean version to the model.
Runs inside your AWS, GCP or Kubernetes install.
Last reviewed
Illustrative. No customer data.
The difference
79 personal-data and credential types detected and replaced with coded stand-ins in your cloud, before the prompt reaches a model.
Source: Personal data detection service and entity registry
How Raidu handles it
Tower receives the prompt, including system prompt and every prior turn. The personal-data scan runs on the full assembled text before the model sees any of it. 79 entity types are checked in parallel: standard personal-data types and credential types.
Each detected item is replaced with a coded stand-in in the version sent to the model. The replacement is in your cloud. The model receives clean text and answers as if it had seen the original, but it hasn’t.
If the model echoes a coded stand-in in its answer, Raidu checks the response on the way back and masks any item that should not leave the session. Even streaming responses are checked as they arrive.
Products that do it
61 registered guardrail policies run on every prompt, including personal-data detection across 79 entity types. Responses scanned on the way back.
Read more ›60+ providers and 467 models, with personal-data detection on every turn before the model sees the prompt.
Read more ›Every detection event lands on a tamper-evident, write-once record. Content can be erased for GDPR while the record keeps the chain verifiable.
Read more ›The product
Product windows from the films, with sample data. No customer data.
The record
Every prompt with personal data lands on a tamper-evident record: what was found, what the model saw, and what the response contained.
Common questions