Recipe Modeler

Recipe modeler

The modeler—used by KitchenPC's “What can I make?” experience—searches for a set of recipes that balances ratings, preferences, restrictions, and efficient use of pantry ingredients.

It requires DBContextCapabilities.RecipeModeler (or All) and loads an in-memory graph. Measure startup and memory with production-scale recipe data before enabling it in a web process.

Anonymous suggestions

Model result = context.Modeler
   .WithAnonymous
   .NumRecipes(5)
   .Scale(50)
   .Generate();

Generate returns recipe IDs, remaining pantry values, and a score. Scale controls the balance used by the modeling session; validate it at the application boundary.

Compile full results

CompiledModel result = context.Modeler
   .WithAnonymous
   .NumRecipes(5)
   .Compile();

Compile adds RecipeBrief values and aggregation data useful for rendering recipes and what the user will need.

Model a user's pantry and preferences

var bananas = context.ParseIngredientUsage("12 bananas");
var carrots = context.ParseIngredientUsage("carrots");

var result = context.Modeler
   .WithProfile(profile => profile
      .AddPantryItem(bananas.Usage)
      .AddPantryItem(carrots.Usage)
      .AddBlacklistedIngredient(Ingredient.FromId(eggId))
      .FavoriteTags(RecipeTag.Dessert | RecipeTag.GlutenFree)
      .AvoidRecipe(previousRecipeId))
   .NumRecipes(5)
   .Compile();

A profile can also provide ratings, favorite ingredients, allowed tags, and a user ID. Implement IUserProfile when preferences come from your own service, or use the fluent ProfileCreator for an ad hoc request.

NLP is not intrinsically required by the modeler: it accepts normalized PantryItem/IngredientUsage data. Enable IngredientParsing too only if your application accepts free-form pantry text.

Operational advice

• The algorithm's usefulness depends on recipe diversity, ingredient metadata, ratings, and realistic quantities.
• A tiny sample with 30 recipes may not satisfy restrictive profiles.
• Treat “no result” as an ordinary outcome and let users relax constraints.
• Do not rebuild the graph per request.
• Consider a dedicated worker or service if production graph size makes web startup undesirable.