Embeddings that always produce vectors of the same length regardless of input length, which limits how much detail can be captured for very long documents.
Performance retention over long documents and conversations