Untargeted metabolomics experiments suffer from large proportions of unannotated molecules. Using a reference data-driven approach, we increase the spectral annotation rate by assigning potential sources to molecular features. We have applied this approach using a food reference database to generate diet readouts from clinical samples.