Chatting With Your Data: LLM-Enabled Data Transformation for Enterprise Text to SQL

Recent advances in large language models (LLMs) have shown remarkable success in code generation across general-purpose programming languages. However, the translation of natural language to SQL in real-world enterprise settings remains a challenge. While state-of-the-art models achieve over 80\% execution accuracy on academic datasets like Spider and BIRD, their performance drops dramatically in the face of large, heterogeneous enterprise schemas. This gap stems not from SQL’s syntactic complexity but from the implicit business knowledge and schema-level irregularity embedded in real enterprise databases. In this work, we present MAIA, a Management Abstraction and Intelligence Algorithm that transforms fragmented enterprise data models into semantically enriched logical representations. These logical data models (LDMs) are embedded into prompts that guide LLM reasoning through schema abstraction, synonym resolution, and business logic alignment. Our method reframes Text-to-SQL as a knowledge representation problem and introduces a sequential agent-based framework that orchestrates aspects like object and variable selection, condition inference, and query assembly. We evaluate our approach on a real world benchmark derived from a technology firm managing software development tickets. Our framework significantly outperforms standard prompting baselines on models like Phi-4-14B and LLaMA-3-70B-Instruct, especially on queries involving joins, nested logic, and ambiguous semantics. This work highlights the importance of schema intelligence and suggests that the most impactful innovations in industry Text-to-SQL systems may lie not in code synthesis, but in making the underlying structured data representation more logical and explainable.