Search papers, labs, and topics across Lattice.
This paper introduces a semantic-layer-mediated agent that enhances natural language-to-SQL (NL2SQL) capabilities over complex enterprise databases by utilizing a curated semantic layer to separate semantic intent from physical SQL execution. By employing a compact intermediate representation known as the Semantic Model Query (SMQ), the agent effectively translates user queries into dialect-specific SQL through a deterministic compiler, achieving a remarkable 94.15% execution accuracy on the challenging Spider2-snow benchmark. The results indicate significant improvements over traditional schema-only approaches, demonstrating the agent's ability to handle the intricacies of real-world enterprise data environments.
Achieving 94.15% execution accuracy, this NL2SQL agent leverages a semantic layer to tackle the complexities of enterprise databases, outperforming traditional methods.
Natural language-to-SQL (NL2SQL) over real-world enterprise databases remains significantly more challenging than on academic benchmarks. Enterprise schemas often contain hundreds of physical tables with cryptic column names, heterogeneous SQL dialects, and complex analytical workloads requiring nested aggregations, temporal reasoning, and multi-table joins. We present a semantic-layer-mediated NL2SQL agent that decouples semantic intent from physical SQL execution. Rather than generating SQL directly over raw schemas, the agent reasons over a curated semantic layer through a compact intermediate representation called the Semantic Model Query (SMQ). A deterministic compiler translates each SMQ into dialect-specific SQL, providing verified building blocks that the agent composes into the final query. The system employs a constrained think-act loop, supports SQLite, BigQuery, and Snowflake backends, and is integrated into an end-to-end evaluation framework. Using Gemini 3 Pro, the system achieves 94.15% execution accuracy on the 547-task Spider2-snow benchmark, ranking third on the official leaderboard and substantially outperforming schema-only approaches. We describe the system architecture, SMQ representation, agent workflow, evaluation results, and discuss semantic-layer quality and the trade-off between improved grounding and overfitting.
