ARTFEED — Contemporary Art Intelligence

Schema-Aware Localisation (SAL) for Oracle NL2SQL

ai-technology · 2026-07-29

A recent publication on arXiv (2607.22572) presents Schema-Aware Localisation (SAL), a middleware solution designed for Oracle NL2SQL that mitigates hallucination and schema grounding issues prevalent in large language models. SAL constructs a dynamic schema map by querying Oracle's USER_TAB_COLUMNS catalog, identifies pertinent subsets of tables for each question, and incorporates accurate context into LLM prompts. The SQL generated is assessed by the Hallucination Index (Hidx), which verifies alias.column references against the current catalog and automatically corrects any invalid references. This method does not necessitate model retraining and specifically addresses ORA-00904 invalid-identifier errors frequently encountered in enterprise Oracle databases.

Key facts

  • Paper arXiv:2607.22572 introduces Schema-Aware Localisation (SAL)
  • SAL is a lightweight middleware layer for Oracle NL2SQL
  • No model retraining is required
  • SAL queries Oracle's USER_TAB_COLUMNS catalog to build a live schema map
  • It selects relevant table subsets per question, falling back to full schema for multi-table queries
  • Hallucination Index (Hidx) validates alias.column references against live catalog
  • Hidx automatically rewrites invalid references
  • Targets ORA-00904 invalid-identifier errors in enterprise Oracle databases

Entities

Institutions

  • arXiv

Sources