Guide · Data
RAG Corpus Ownership, Classification, and Retention
If nobody owns the corpus, classification is informal, and retention is forever, retrieval will leak and rot. APRF Data Governance requires ownership, labels, and lifecycle—not only ACLs at query time.
ACLs alone are not enough—APRF Data Governance requires corpus owners, classification, and proven retention. Primary control: Data Governance
A corpus is a dataset with an owner—or it shouldn't be indexed
RAG data governance means each source has a named owner, classification, allowed purpose, retention/purge, and change control. Privacy/AuthZ enforce access; governance decides what may exist and for how long. One team indexed "everything in SharePoint"; HR drafts mixed with FAQs, no owner, no purge. After someone left, their drafts still answered customers. Fix: per-site owners, `classification` on chunks, 90-day draft purge, tickets for new connectors.
Rules before the next crawl
Inventory corpora (source, purpose, owner, class, retention). No owner → no index. Classify at ingest (public/internal/confidential/restricted) and filter in retrieval. Define TTLs and prove delete from index + object store. Change-control new connectors (purpose, legal basis, blast radius). Pair with memory/session retention so chat state doesn't outlive the corpus policy.
APRF Data Governance owns the register
This is the operational half of RAG leakage prevention—labels and purge jobs, not only redaction prompts.
Assess against APRF Core
Run the Core Profile quiz — gated pass/fail blockers for AI production readiness, not a vanity score.
Related
Frequently asked questions
- How is this different from the RAG leakage guide?
- Leakage focuses on query-time ACLs and exfiltration. This guide covers ownership, classification, and retention—what may be in the corpus at all.
- Do we need a DPO for every corpus?
- You need a named accountable owner and a path to privacy/legal review for personal or regulated data. Scale the formality with risk.
- What about memory stores?
- Session and long-term memory need the same ownership and TTL discipline—see Memory Management.