A Schema-Driven Prompt Engineering Approach for Data-Centric LLM Tasks

General-purpose LLMs are increasingly employed in a variety of data-centric tasks, ranging from entity and relationship extraction from plain text to schema-aware data integration and analysis. Their large number of parameters and extensive training corpora enable strong generalization across application domains. However, LLMs lack access to up-to-date knowledge and are prone to hallucinations, limiting their reliability in data management scenarios. To address these issues, prompt engineering techniques have been proposed to specify the context in which a task should be performed. In this paper, we explore the use of conceptual schemas as a foundation for schema-driven prompt engineering, providing structured and reusable contexts for grounding LLM behavior in data-centric applications. We present SchemaLink, an intelligent web-based system for the graphical design and enhancement of conceptual schemas and their exploitation for LLM-based tasks. We demonstrate the applicability of our approach to knowledge extraction from plain text and to the discovery of joinable columns in data lakes, showing how schema-driven prompting improves grounding and consistency across heterogeneous data sources.