The installment of the Enterprise Document Intelligence series dissects the question‑parsing brick that sits between raw user input and a RAG pipeline. It shows how a single string, such as “What is the maximum coverage amount?”, becomes a typed row with topic, answer shape, scope hint, negative cue and layout hint. The article walks each of the five field families the question parser fills.

Five column families capture the full intent of a query. Keywords aggregate tokens from the user, LLM rewrites, concept dictionaries and regex anchors to drive retrieval. Answer shape splits cardinality—single, list, table, tree, nested_json—and value type such as amount or date. Scope hints point to pages, sections or layouts, while decomposition breaks compound questions into sub‑queries and clarification flags vague inputs for follow‑up.

These columns populate question_df, a DataFrame consumed by downstream retrieval and generation bricks. Dispatch logic then selects chunk size, model and activation set based on answer type and document profile, enabling use cases from insurance policy numbers to medical patient IDs. Teams can drop unused fields or extend the schema as failure modes emerge, keeping the pipeline lean.