Redaction
Remove sensitive details, keep everything else exactly as it was.
Notes, transcripts and tickets need personal data removed before they are shared, stored or analysed. Sending them to a third-party model to be redacted defeats the purpose. A small model in your own account does the job where the data already is.
The problem
Redaction must happen before data leaves
If the text has to be sent out to be cleaned, the sensitive data has already left.
A single leak matters
Averages hide the one example where a name or account number slipped through.
Over-redaction ruins the text
Removing too much makes notes and transcripts useless for the people who read them.
Where it applies
- Call and chat transcripts
- Clinical and case notes
- Support tickets
- Logs before analytics
- Documents before sharing
How the switch is checked
Before you switch, the fine-tuned model and the LLM you use today answer the same held-out examples from your data.
- Leak rate: the share of examples where any word your reference removed still appears
- Over-redaction: the share of words that should have stayed but were removed
- Every leaking example listed with the words that leaked
What you need
Each example's original text, and the same text with the sensitive details replaced.