Sure, the difference is that the first endpoint would give you back only one single embedding vector for the entire paragraph, while the second endpoint would give you a separate embedding vector for each sentence in the paragraph.
And yes, everything in this code is designed to run on the CPU well on a modest machine and is 100% self-hosted, no API keys needed at all. But if you do have a GPU installed and configured it will automatically use that since it’s powered by llama-cpp which now supports CUDA.