Trustworthy LLM-Agent-Oriented Enterprise ETL Orchestration via Neural Execution Verification

Abstract

To address the challenges of heterogeneous data sources, complex business semantics, dense task dependencies, and insufficient execution reliability in enterprise ETL processes, this paper proposes an automated orchestration and reliable execution method for enterprise ETL based on LLM-Agent and neural verifier. This method first models data tables, fields, transformation relationships, quality rules, and task dependencies into a unified ETL state diagram, fusing structural features, semantic descriptions, and quality profiles to form a reasonable data state representation. Then, LLM-Agent is used to decompose enterprise data processing requirements into tasks, select operations, generate parameters, and plan dependencies, generating candidate ETL orchestration schemes with executable structures. To reduce logical biases and constraint omissions in the automatically generated process, a neural validator is further designed to judge the semantic consistency, structural legality, constraint satisfaction, and execution reliability of candidate operations, and an execution gating mechanism is used to complete operation screening, process correction, and reliable control. Comparative experimental results show that the proposed method exhibits good overall performance in terms of process planning accuracy, verification, and discrimination ability, constraint satisfaction degree, and execution success rate. It can effectively improve the quality of automated orchestration and reliable execution of enterprise ETL tasks, and provide effective methodological support for the intelligent governance of complex enterprise data pipelines.