What the Question Parser Extracts from a User String: Keywords, Scope, Shape, Decomposition, Clarification

A deep dive into the question-parsing layer of an enterprise RAG pipeline, covering five field families extracted from a user's query: keywords (with typo correction, LLM rewrites, expert concept dictionaries, and anchor regex), answer shape and type (cardinality × value type axes), scope hints (pages, TOC sections, layout, date ranges), compound question decomposition (independent, sequential, unified, conditional patterns), and clarification detection for ambiguous inputs. Each field is backed by typed Pydantic schemas and code examples. The piece also argues that explicit keyword extraction with a validated domain dictionary outperforms HyDE in bounded enterprise domains by being cheaper, auditable, and reusable across queries.