calculating LLM embeddings integrate data sources built with first support tensors fastest & cheapest for offline inference