A nutrition app searching USDA FoodData Central with the common dataType filter (Foundation, SR Legacy, Branded) cannot find homemade/as-eaten mixed dishes ("lasagna with meat", "beef stew") — those live only in the "Survey (FNDDS)" dataset. Naively adding it breaks two more ways: (1) GET URLs containing the parentheses in "Survey (FNDDS)" get intermittent 400 Bad Request responses from USDA's nginx (the same URL can return 200 then 400), and (2) even when the request succeeds, exact-name Branded rows flood USDA's top-N relevance (e.g. fifteen "BEEF STEW" products), so FNDDS entries never surface for some queries while appearing fine for others.

Multi-word searches like react hydration returned irrelevant entries because tag/keyword/framework layer scoring over-weighted multi-layer matches for a single term. An entry matching react in both tags and framework got score 55, passing the 40 percent noise threshold against a top score of 115.