SF-MAS: Schema-Free Multi-Agent System for Text-to-SQL

Text-to-SQL systems assume complete database schemas are available, yet in production, schemas are often missing, outdated, or access-restricted—after database migration, schema evolution, or in security sandboxes. We present SF-MAS, the first NL2SQL system that operates with zero schema knowledge. SF-MAS combines six specialized agents with a three-channel schema inference engine that recovers column semantics from data values, system catalog metadata, and LLM commonsense knowledge, reconciled by a deterministic arbitration rule. A SQLite-backed semantic cache cuts schema discovery overhead by 2,786x (0.07s per column). For cross-table reasoning, LLM-based disambiguation resolves semantically indistinguishable table names by jointly weighing name similarity, column semantics, and query intent (+13.1 pp). A column-level revision feedback mechanism—checking entity coverage, aggregation matching, and table selection during SQL repair—adds another +3.0 pp. On BIRD dev (1,534 samples), SF-MAS reaches 52.54% execution accuracy without any schema input, surpassing DIN-SQL + GPT-4 (50.72%), which consumes complete schemas and a larger model. Ablations confirm LLM disambiguation as a necessary condition for schema-free table selection, with column feedback providing consistent refinement. Our results demonstrate that competitive NL2SQL is achievable with zero schema knowledge, opening the door to schema-constrained applications in dynamic data environments.