Power customer service data is predominantly stored in structured databases using relational schemas, yet existing language models primarily focus on textual data parsing while lacking effective mining capabilities for structured databases, consequently constraining the advancement of intelligent recommendation methods for power service expertise. To address this limitation, this paper proposes a Text2SQL benchmark-constrained intelligent recommendation analysis method for power customer service data, which takes relational metadata as parsing targets to construct a business constraint rule base mapped to the power service lexicon. It extends domain-specific terms through attribute similarity correlation analysis. Building association dependency models via joint probability distribution calculations oriented by service invocation scenarios. Then defines data module frequency matrices and computes object dependency weights based on co-occurrence distributions. The proposed method performs topological subgraph segmentation to create optimized energy-domain retrieval models. The framework integrates local power-domain digital object repositories with natural language processing through a graph neural network recommendation system, establishing Text2SQL-based NL-to-SQL conversion constraints with feedback calibration mechanisms and knowledge selection templates to enhance recommendation accuracy. This approach demonstrates superior performance over traditional text-based mining methods by enabling precise natural language-to-query conversion constraints for power service data recommendations.