How do you design a reliable schema extraction pipeline using LLMs when the input document layout is highly variable and non-standard at Reducto?