what are the best practices while creating a knowledge base that maximise the nlp efficiency?