use case · practical
Two things you'd actually ship. Redaction masks the people, places and organizations in a document before it's shared or logged — and because it runs on-device, the sensitive original never leaves the tab. Extraction pulls those same entities out into a clean, grouped structure you can hand to the next system.
Choose which entity types to mask:
Every entity, de-duplicated and grouped by type — the who and where, as data.
A statistical NER model won't catch every name, so treat redaction as a strong first pass, not a compliance guarantee — pair it with pattern rules for emails, phone numbers and IDs. The win is that the whole pass happens locally, so the un-redacted text is never transmitted.