Bounded Semantic Planning and Deterministic Compilation for Reliable Enterprise Text-to-SQL

This work presents Semantic Path Compilation (SPC), an enterprise text-to-SQL architecture. In contrast to text-to-SQL architectures that execute a LLM directly to generate relational queries, SPC fundamentally separates language interpretation from relational query construction. An LLM planner performs bounded semantic grounding and chooses from a set of governed, question-specific options, while deterministic software performs relationship paths, role predicates, aggregation, etc. and generates and validates SQL. Running SPC on the 38-question adjudicated ACME insurance benchmark, with three runs per question, yielded solid success on 37/38 questions (97.4%), compared with 21/38 (55.3%) for direct DDL-to-SQL generation. Across 114 SPC outcomes, there were no adjudicated wrong-but-executed results and one refusal. A conservative paired analysis over all 44 benchmark items also favored SPC, with 17 SPC-only solid successes and one baseline-only solid success. This work is intended to evaluate the complete SPC system and not to investigate the impact of the deterministic compilation in isolation from the semantic artifacts that the SPC system is provided with, as opposed to the direct DDL-to-SQL baseline.