MongoDB Atlas

Atlas Search now offers StableTfl, a new scoring function that computes relevance from only the query term, its frequency, and document length, removing dependence on corpus-wide statistics like BM25 uses. This means paginated search and RAG retrieval no longer produce duplicate or skipped results when requests hit different replicas, since every replica now returns the same score for a given query and document without needing sticky routing, synchronized statistics, or index snapshots.

Matters for: Teams running paginated search or RAG retrieval pipelines on distributed, multi-replica Atlas Search deployments.