That again depends on the specific use case - stnadard embeddings i.e. have serious shortcomings when questions are about a product variety as the product name is not sensibly (if at all) part of the embedding. ElasitSearch has a whole talk about how one must combine vectors and semantic search and possibly tagging to get a good result in a LOT of cases. Seriously, there is no easy answer with no context - and you provide no context.