How would you keep a RAG system fresh without reindexing everything constantly?

Instruction: Describe how you would preserve freshness while controlling ingestion cost. Document D v4 has 10 chunks. A content edit creates v5 with 11 chunks; the same update event arrives twice. The system identifies chunks by source/version/chunk index. What should be staged, activated and removed, and why is appending only the new chunk insufficient?

Context: Checks whether the candidate can explain the core concept clearly and connect it to real production decisions. Describe how you would preserve freshness while controlling ingestion cost.

Updated

Official answer available

Read the opening below, then unlock the full answer and practical guidance.

The way I'd approach it in an interview is this: I keep RAG fresh by separating change detection from full reindexing. Every document should have a stable identity, a version or content hash, and clear timestamps...

Upgrade to view official answer

Related Questions