#multi-vector

DeepMind demonstrates a mathematical limit on fixed-size dense embeddings that causes retrieval failures in RAG systems at scale, and the LIMIT benchmark exposes this ceiling even on small toy tasks.