Bring your own embeddings. PyTorch and TensorFlow packages are 2GB+ each (don't quote me on that), which is unnecessary if you're making a network call to your favorite embedding service.