Meaning-based search: documents are matched by similarity, not by exact words. First request downloads the embedding model.