The "ingestion layer + validation layer" layered design of document review is similar to traditional ETL pipelines, but adds an AI inference step, and processes highly heterogeneous unstructured content, placing higher demands on fault tolerance and interpretability.