An LLM-Driven Intelligent Agent for Adaptive Data Preparation with Strategy Memory

Abstract

Data preparation, including source profiling, quality diagnosis, and cleaning, remains a labor-intensive bottleneck in data-driven applications. Traditional approaches require manual rule design for each data source and lack adaptability to heterogeneous formats. This paper proposes an LLM-driven intelligent agent for adaptive data preparation. The agent leverages a large language model as its reasoning core to perform three tasks in a closed loop: (1) automatic data source profiling through sampling-based probing, (2) data quality diagnosis and cleaning script generation via LLM code synthesis with sandbox verification, and (3) strategy memory that accumulates validated experiences as structured condition-actioneffect triples for reuse. Generated scripts are tested in an isolated environment before deployment, and both successes and failures are recorded to enable continuous improvement without model fine-tuning. This paper presents the architecture design, details the core methods, and discusses the design rationale.