ATLAS: Adaptive Text-to-SQL with Lifecycle-Aware Self-Maintaining Context

Abstract

Deploying Text-to-SQL in production is hampered by context-window limits on large schemas, metadata that goes stale as schemas evolve, and infrastructure sprawl from external vector stores. We demonstrate ATLAS , which addresses all three by co-locating schema metadata, semantic annotations, and vector embeddings entirely within a single RDBMS with native vector-index support. ATLAS contributes: (1) Unified in-database storage with native vector indexes, requiring no external engine; (2) Two-stage adaptive schema linking that uses vector retrieval to narrow hundreds of tables to a compact candidate set before LLM refinement; (3) Rich Context life-cycle that automatically generates semantic descriptions, synonyms, sample values, business rules, and value mappings, bridging raw schema and LLM understanding; (4) Agent-driven self-maintenance via a coordinator-executor loop that detects DDL changes and regenerates annotations without manual intervention. We showcase three live scenarios on an enterprise database with 517 tables.