Most retrieval quality problems trace back to how documents were split, not to the embedding model or the vector database.