Why is metadata important in RAG?
Strong answer: metadata controls filtering, permissions, citations, freshness, authority, evaluation slices, and debugging.
Follow-up: give an example where semantic similarity alone is unsafe.
Interview Questions: Metadata Design
Use these questions to practice explaining metadata as a production RAG control plane.
What fields would you require for a policy chunk?
Strong answer: source id, chunk id, URI, section, authority, owner, access level, allowed roles, region, version, effective date, source position, schema version.
Follow-up: which fields are needed for citations?
How do you enforce access control with metadata?
Strong answer: store tenant, role, access, sensitivity, and entitlement fields, then filter before retrieval results enter the prompt.
Follow-up: what should be logged when a source is blocked?
How do you handle stale documents?
Strong answer: store effective_from, expires_at, last_reviewed_at, indexed_at, authority, and retired state, then filter or downgrade stale sources.
Follow-up: what metric detects stale retrieval?
Why use controlled vocabularies?
Strong answer: filters need stable values. Free-text regions, roles, and authority labels cause missed retrieval and inconsistent behavior.
Follow-up: how would you migrate old values?
How does metadata help evaluation?
Strong answer: it lets teams slice metrics by region, source type, role, version, authority, and metadata completeness.
Follow-up: what eval would catch a region filter bug?
What is schema_version for?
Strong answer: it tracks which metadata contract a chunk follows, enabling migrations, validation, and debugging across index versions.
Follow-up: what happens if schema versions mix in one index?
What is the risk of missing source positions?
Strong answer: citations become vague, audits become weak, and reviewers cannot inspect exact evidence.
Follow-up: how would you store source position for tables?
How should authority metadata affect ranking?
Strong answer: approved current sources should outrank drafts, historical notes, and untrusted user-generated content.
Follow-up: when should historical content still be retrieved?
How do you measure metadata quality?
Strong answer: track completeness, validity, unknown-value rate, stale-source rate, filter miss rate, blocked-source rate, and schema drift.
Follow-up: which metric would you put on a release dashboard?