The parent-child document architecture adopts a "retrieve small chunks, recall large chunks" strategy: fragmented child documents are used during retrieval to ensure precision, while complete parent documents are used as input to the large model to ensure context integrity.