This is an interesting and complicated set of problems. It's similar to problems in the LLM & social network spaces. You have several dimensions, like recipes, ingredients, nutrition, cultural context, etc, and you have gradients to help determine recommendations, which points to a graph DB. But you also have rules when it comes to substitutions (not every substitution is appropriate for every situation), and you might have rules on other dimensions as well. That points to a layer of business logic. But to serve recommendations performantly, you might want to run rules upon writes or on a schedule to persist the results to the DB. Either that or do some sophisticated caching.