A RAG system is only as current as its index. Stale content produces confident, outdated answers.
Incremental Ingestion
Re-indexing everything regularly is simple but slow and costly for large collections. Instead, detect changes:
- source system change events or webhooks;
- modified timestamps;
- content hashes to spot real changes.
Re-parse, re-chunk and re-embed only changed documents.
Handle Deletions
When a document is removed or unpublished, delete its chunks. Forgotten deletions are a common source of wrong — and sometimes confidential — answers.
Versions and Supersession
Mark superseded documents clearly, prefer current versions in retrieval, and show effective dates in answers.
Embedding Model Changes
Changing the embedding model requires re-embedding the entire collection, because vectors from different models aren't comparable. Plan migrations with a parallel index and switch over when complete.
Monitoring Freshness
- Track the lag between a source change and its appearance in the index.
- Alert when ingestion jobs fail.
- Periodically sample answers and check their sources are current.
Ownership
Assign owners for each content source. RAG quality often depends more on content hygiene — removing duplicates and outdated drafts — than on technology.