FURI | Spring 2026

Agentic Diagnostic Backtracking for NL2SQL on Federated Databases

Data icon, disabled. Four grey bars arranged like a vertical bar chart.

Natural language to SQL (NL2SQL) is the process of translating English data requests into executable database queries. Modern LLMs handle single-database NL2SQL intuitively given sufficient schema context, but real-world systems often span multiple databases with differing architectures. This research addresses the problem of querying federated DBMS by proposing an agentic framework designed for multi-step code generation problems requiring iterative correction. The approach provides fine-grained stage control while maintaining a self-correcting feedback loop. Strong performance on this task would demonstrate a scalable path toward querying complex data systems and establish a generalizable approach for other multi-step generation problems.