EasyDeepLearn

How do you handle content updates and stale embeddings in production RAG?

medium

Answer

  • (1) Track document versions; re-embed changed docs and upsert.
  • (2) Delete stale vectors — tombstone or hard delete.
  • (3) Backfill jobs when embedding model changes (full re-embed of the corpus, staged rollout).
  • (4) Version namespaces: keep old + new embeddings in parallel, canary rollout with a small % of traffic on the new model.
  • (5) Idempotent IDs (stable across re-embeddings) simplify updates.
  • Avoid full-corpus re-embeds for minor changes.
Check yourself — multiple choice
  • Never update
  • Version + upsert changed docs, delete stale, backfill on model changes with canary, stable idempotent IDs
  • Random reindex
  • Only append

RAG updates: version + upsert + tombstone + canary backfills on model changes.

#vector-db#production

Practise LLMs & GenAI

214 interview questions in this topic.

Related questions