Just one tool in your toolbox. I think it depends on the task and size of the data. It used to be: Use embeddings for every extraction task.
Now it is: choose a harness with a cheap/fast model or classifier, or embeddings depending on your extraction task
