Typical benchmarks in the Text-to-SQL field (such as Spider and BIRD) usually provide clean schemas, well-defined single-turn questions, and standard answers, whereas real enterprise scenarios often contain dirty data, multi-turn clarification, permission filtering, cross-database joins, and ambiguity in metric definitions, leading to a systematic gap between benchmark scores and production performance.