Preprint reports lossless migration of 21 databases for MongoDB query benchmark

Researchers report in an arXiv preprint that their conversion pipeline migrated all 21 BIRD benchmark databases from SQLite into document databases without data loss. Earlier conversion methods failed on 6 databases, dropped up to 25.9% of rows on others, and produced queries that ran over an order of magnitude slower as data scaled. The pipeline uses coding agents with human verification to organize data around expected queries and rewrite those queries for MongoDB. The resulting AptMQL-Bench contains 3,186 natural-language requests. Claude Opus 4.5 achieved 57.38% accuracy without external knowledge evidence and 70.34% with it, showing that reliable query generation remains difficult.