Structured AI: A Three-Tier Adaptive Framework for Enterprise Data Analytics
Abstract
Traditional Retrieval-Augmented Generation (RAG) systems and SQL-based query engines face significant challenges when processing real-world Excel files containing multiple tables, merged cells, multi-row headers, formulas, pivot-style layouts, and inconsistent schemas. Structured AI is a hybrid three-tier framework that assesses workbook scale, structural cleanliness, and semantic relevance before routing a query to document-centric processing, SQL-optimized processing, semantic sheet selection, or a coordinated hybrid path. The methodology contributes a measurable cleanliness model, query-sheet relevance scoring, cost-aware route selection, formula-preserving conversion, and confidence-based result synthesis. In the reported prototype evaluation, average query accuracy increased from 78% for a static baseline to 89% for the adaptive framework, while median latency on large multi-sheet workbooks decreased from 21.5 s to 12.4 s. These values are reported consistently throughout the manuscript. Because the submitted source did not include the underlying workbook corpus, run-level logs, or sample counts, the revised paper explicitly limits the empirical claim to the reported aggregate evaluation and identifies the evidence required for independent replication.