RAG Wins Come From Data Hygiene, Not Bigger Embeddings
Retrieval quality still hinges on chunking, deduplication, and fresh indexes—not embedding dimension alone.
Teams that struggle with RAG usually have messy knowledge bases: duplicate docs, stale wikis, and PDFs without structure.
Hybrid search—keyword plus semantic—remains essential for proper nouns, SKUs, and internal codenames embeddings miss.
Re-ranking with a small cross-encoder or LLM step often beats jumping to the largest embedding model available.