The use of machine learning to recommend foods that are both healthy and tasty is an open problem. Fundamentally, it is challenging to balance health goals with preferences in taste, while offering users a large diversity of options. Representing recipes via embedding vectors trained on large-scale food datasets can capture the implicit semantics of a recipe. We utilize pre-Trained embeddings to perform recipe search and compare our search results with a keyword based search. We compare the health score, nutritional content and recipe titles returned using both search approaches. Our exploratory experiments show that recipe search via embeddings can return more diverse recipe titles in contrast to keyword based search.