Even in production my guess is most teams would be better off just rolling their own embedding model (huggingface) + caching (redis/rocksdb) + FAISS (nearest neighbor) and be good to go. I suppose there is some expertise needed, but working with a vector database vendor has major drawbacks too.