How is pgvector with Sentence Transformers, a CPU-only embedding model, compared to a model hosted by OpenAI?