Rosetta: Recovering Column Semantics in Undocumented Data Warehouses

A recent study published on arXiv (2608.07946) presents Rosetta, a system aimed at deciphering column semantics in undocumented data warehouses. In contrast to text-to-SQL benchmarks that feature schemas with clear column names, many production warehouses utilize obscure identifiers with minimal documentation. Rosetta tackles this issue by integrating a language model within a verification framework: a deterministic profiler gathers structural data, such as value fingerprints, a library of 26 patterns, and checksum assessments. Based on this data, the model suggests semantics, ensuring each fact is linked to its source and assigned a confidence score based on its evidence category. In evaluations against human documentation on 680 paired columns from eleven BIRD databases, the harness achieved 0.475 accuracy on 42% of the columns it analyzed, while the same model used directly yielded 0.223 on 94%. For the 283 columns where both methods provided results, the harness did not outperform the baseline. The full paper can be accessed at https://arxiv.org/abs/2608.07946.