PII

Personal data is any information making a natural person identified or identifiable. Names, identity numbers, contact details, HR records, and health data all qualify; health data is additionally treated as a special category.

When a file containing personal data is attached to an assistant, that data is indexed and can be surfaced as a source in a query result. Trying to protect it with an access rule is more fragile than never attaching the file.

The strongest protection therefore sits at setup: files containing personal data are not brought into any assistant’s scope. A file that is not attached can never appear as a source.

The main document may be clean while its attachments are not. Sample forms at the end of procedure documents can carry real names; meeting minutes include attendee lists; contract templates may have been filled in with real customer details.

The second overlooked place is scanned paperwork. When an image-based PDF goes through text extraction, the identity details inside it enter the index too; being an image is not protection.

A quick review before upload is therefore far cheaper than selective deletion afterwards. This step should not be skipped, especially during bulk uploads.

A deletion request requires removing the data from two places: the source document store and the vector index. Doing only one leaves the data reachable.

That operation needs testing. After taking a document out of scope, query the same content; if an answer still comes back, index removal is not working.

Never indexing personal data in the first place makes this operation unnecessary and is the safest approach. Consult your legal adviser for compliance assessment; this page is not legal advice.