DataMosaic: An Interactive Demonstration of Constraint-Driven Document-to-Database Construction

Abstract

Large language models (LLMs) have improved document information extraction, but turning extracted facts into relational databases remains fundamentally difficult. The challenge is that extraction is local to text, whereas database construction must satisfy global semantics defined by schemas, keys, and integrity constraints. Consequently, single-pass LLM extraction often produces inconsistent tuples, missing links, and semantically invalid results. We present DataMosaic , an interactive system for constraint-driven document-to-database construction. Given raw documents together with user-specified schemas and constraints, DataMosaic incrementally builds entity and relationship tables through a closed loop of extraction, verification, repair, and targeted re-extraction. The system exposes intermediate table snapshots, detected violations, and repair actions, enabling users to audit how database-level feedback improves the final relational output. In the demo, attendees can configure inputs and constraints, inspect real-time analysis logs, and observe how DataMosaic repairs inconsistencies and materializes query-ready relational tables. The demonstration highlights how integrating extraction with verification and repair enables reliable document-to-database construction.