RAS: Retrieval-And-Structuring for Knowledge-Intensive LLM Generation