When an Astra DB collection is set up with a Vectorize embedding integration, your application must not also attach its own embedding model: the collection generates vectors itself, and double-embedding wastes calls and can corrupt similarity results. The Langflow Astra DB component docs say: "Attach an embedding model component to generate embeddings. If the collection has a vectorize integration, don't attach an embedding model component." Related detail from the same docs: the similarity metric (cosine, dot_product, euclidean) is a collection-level setting, so it must be chosen when the collection is created, not per query.