I would guess that the embedded summary is better, but for many tasks where you use embeddings (like document search), summarizing every document with an LLM is too expensive and slow.