Module 09 Summary
What this module established
The strongest privacy control is refusing to index something; everything after ingestion is weaker. Access control is a set intersection applied before ranking - deterministic, testable, and independent of the model behaving. Update, supersede and delete are three different operations, and a delete that leaves the embeddings has deleted nothing.
Carry forward
- Exclusion and restriction solve different problems. Conflating them either leaks data or makes the assistant useless to staff.
retrievedandshown_to_user, with versioned chunk ids, are the fields an incident investigation needs and cannot reconstruct later.- Retention has to reach the document store, the chunk index and the vector store together, and be verified by query.
Before moving on
Move on when one query returns three different result sets at three access levels, and a deleted document is confirmed unreachable by querying for it.