AI Just Turned 128,000 Recipes Into a Searchable Nutrition Database

Researchers have built a framework that turns recipes — usually free text, readable by people but not directly computable — into structured data. Their compilation, RecipeDB2, covers 128,942 recipes and 35,474 ingredients from 32 regions and 99 countries. Each recipe is decomposed into typed ingredient entities, grounded in a reference nutritional database, and annotated with geo-cultural and dietary context: ingredient phrases are parsed into seven culinary attributes with a transformer-based named-entity model, ingredients are linked to USDA reference tables using a BERT embedding strategy (F1 = 87.90 on a manually adjudicated set of the 200 most frequent ingredients), yielding 148 nutritional parameters per mapped ingredient, a Random Forest classifier propagates 34 ingredient categories across the full vocabulary, and a deterministic, conservative rule set assigns each recipe a dietary style. The authors present this as a scalable framework for turning culinary heritage — long treated as an artistic rather than a quantitative object — into a data-driven analysis.