Beyond Tables: Doc2DB-Bench for Relationally Faithful Document-to-Database Construction

Practical AI systems increasingly need to turn long, heterogeneous documents into queryable relational databases, not isolated spreadsheets. In domains such as finance, healthcare, education, transportation, and enterprise operations, downstream workflows rely on normalized schemas, entity identities, keys, cross-table relationships, and integrity constraints for analytics, compliance, auditing, and SQL-backed decision making. Existing Document-to-Table benchmarks are insufficient for this setting: flattening evidence into single tables can duplicate entities, obscure many-to-many relationships, create sparse records, and avoid testing whether extracted facts form a valid database instance. This creates an urgent need to evaluate document understanding as database construction rather than field extraction. The authors introduce Doc2DB-Bench, a benchmark for Document-to-Database construction, containing 203 long-document instances across 42 schemas and seven domain groups, with 117 entity tables, 132 relationship tables, 7,341 rows, and 41,935 cells. Built through a controllable DB-to-Doc synthesis pipeline and organized by a taxonomy of intra-table extraction and inter-table reasoning, the generated documents undergo authenticity verification, proving indistinguishable from real-world references. Doc2DB-Bench thus provides a testbed for reliable, auditable, and relationally faithful LLM-based data systems. The benchmark is publicly available at https://github.com/SetonLiang/Doc2DB-Bench.

The benchmark is constructed through a controllable DB-to-Doc synthesis pipeline. This pipeline begins with a relational database instance and generates a long, heterogeneous document that encodes the same information in natural language. The synthesis is organized by a taxonomy of intra-table extraction and inter-table reasoning, ensuring coverage of both local field-level extraction and global relational inference. The generated documents undergo authenticity verification, demonstrating that they are indistinguishable from real-world references. The benchmark comprises 203 long-document instances across 42 schemas and seven domain groups, yielding 117 entity tables, 132 relationship tables, 7,341 rows, and 41,935 cells. Formally, let a document $D$ be generated from a database instance

with schema . The task is to reconstruct a database instance such that , i.e., satisfies all integrity constraints, keys, and cross-table relationships defined by . The synthesis pipeline can be represented as a mapping , and the evaluation measures the fidelity of the inverse mapping under schema constraints. The taxonomy separates intra-table extraction, where fields are extracted within a single table, from inter-table reasoning, where relationships across tables must be inferred. This design ensures that the benchmark tests not only extraction accuracy but also relational faithfulness.