Kaikaku.AI has released Epicure, a set of three ingredient-embedding models trained on 4.14 million multilingual recipes and FlavorDB data. The models separate recipe co-occurrence from chemical similarity, producing different food recommendations depending on what kind of “fit” is being measured.
Kaikaku.AI has introduced Epicure, a family of three AI ingredient-embedding models that distinguish between foods that commonly appear together in recipes and foods that are chemically related.
According to the arXiv paper “Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings” by Radzikowski and Chen, Epicure consists of three sibling models trained on 4.14 million recipes across multiple languages, alongside information from the FlavorDB flavor database.
The paper describes three variants. Cooc is trained on recipe co-occurrence, meaning it learns which ingredients tend to appear together in cooking contexts. Chem is trained using compound-based metapaths from FlavorDB, focusing on relationships between ingredients through their chemical flavor compounds. Core blends the two approaches.
That separation matters because “what goes with chicken” can mean different things. A recipe-based model may return ingredients that cooks frequently combine with chicken, while a chemistry-based model may surface ingredients with related flavor compounds even if they are less common in recipes.
The Hugging Face model card for Kaikaku’s Epicure Cooc model gives a concrete example: for “chicken,” the nearest neighbors include garlic, onion, black pepper, turkey, and carrot. That list reflects common recipe context rather than a purely molecular definition of similarity.
The Decoder reports that Epicure is presented as the first set of AI models to clearly separate whether an ingredient fits a recipe or is chemically related. Decrypt similarly reports that Kaikaku.AI released three models trained on 4.14 million recipes, with Cooc, Chem, and Core producing different answers depending on whether recipe co-occurrence or flavor chemistry is emphasized.
The distinction is useful because culinary compatibility is not one thing. A chef may care whether ingredients are conventional together, whether they share volatile compounds, or whether a recommendation balances familiarity with novelty. Epicure’s design makes those assumptions more explicit by assigning them to separate models.
According to The Decoder’s summary of the work, the chemistry-based model also performs better than the recipe-based alternatives at classifying taste and nutritional values, even though it was not directly trained on that information. The available source describes this as an emergent result of the chemical representation, not as evidence that the model has human-like culinary understanding.
Decrypt characterizes the release as compressing a large body of cooking data into small embedding models, while noting the same core training set of 4.14 million recipes. The arXiv paper frames the work around ingredient embeddings: numerical representations that place ingredients near one another according to learned relationships.
Those embeddings can support search and recommendation tasks. A recipe app, for example, could use a co-occurrence model to suggest familiar substitutions or complementary ingredients. A food research tool could use the chemical model to explore less obvious pairings based on shared compounds. The blended Core model offers a middle ground between culinary convention and chemical similarity.
The sources do not establish whether Epicure has been adopted in commercial kitchens or consumer cooking products. They do show a research and model release intended to make food recommendation systems more transparent about the type of similarity they use.
Many AI recommendation systems collapse different meanings of “similar” into a single answer. Epicure’s contribution, as described by Radzikowski and Chen and reported by The Decoder and Decrypt, is to separate those meanings for ingredients: what people cook together, what molecules connect, and what results from combining both.
That could make culinary AI tools easier to interpret. If a system recommends garlic with chicken, users may want to know whether that is because millions of recipes pair them, because their flavor chemistry overlaps, or because both signals point in the same direction.
Epicure does not replace culinary judgment, but it gives researchers and developers a clearer vocabulary for a basic food question: does this ingredient belong because cooks use it, because molecules suggest it, or because both are true?
Kaikaku.AI has introduced Epicure, a family of three AI ingredient embedding models that distinguish between foods that commonly appear together in recipes and foods that are chemically related.
Cooc is trained on recipe co occurrence, meaning it learns which ingredients tend to appear together in cooking contexts.
Chem is trained using compound based metapaths from FlavorDB, focusing on relationships between ingredients through their chemical flavor compounds.
Continue reading
