Mental model
Deletion is a lineage operation: find every governed derivative of a source record, expire or rebuild it, and record what was completed or delayed.
Learning outcomes
- Explain the mechanism and ownership boundaries behind AI Retention, Deletion, and Audit.
- Compare the main design alternatives and their operational trade-offs.
- Diagnose common failures and select evidence for a production decision.
Theory
Define retention by data class and purpose, maintain provenance, issue tombstones, rebuild derived indexes, expire caches, manage backup windows, and produce tamper-evident audit events without retaining deleted content.
Trade-offs
Long retention improves debugging and evaluation history but raises privacy and breach exposure. Short retention limits diagnosis unless aggregates and sanitized exemplars are designed deliberately.
Failure modes and misconceptions
Deleting only source rows; leaving vector chunks; immutable logs containing content; backups with no expiry; no proof of deletion; and audit records that reproduce sensitive data.
Decision scenario
A customer invokes deletion after their documents were indexed and used in evaluations. Trace the required actions and evidence across every derivative.
Why is deleting a source document insufficient in a RAG system?
Primary sources
nist-privacy-frameworkw3c-prov-overview