Assemble Each RAG Generation Prompt from a Base Prompt Plus the Rules Each Question Needs

A dispatcher approach is proposed for assembling prompts in Retrieval-Augmented Generation (RAG) systems to avoid complex, monolithic mega-prompts. This method dynamically composes a prompt at call time by combining a fixed, shape-agnostic base prompt with specific rule fragments tailored to the query. The process is guided by a structured `ParsedQuestion` object, which details the expected answer shape, keywords, and other constraints. This modular design enables more maintainable, debuggable, and efficient interactions with a Large Language Model for enterprise document intelligence tasks.