So you’re right. MetaCAT uses bi-LSTM but it can be done with BERT as well.
As for the main NER pipeline. You can consider it as a custom pipeline. It is basically a vocab-based look up with Context-based disambiguation to link words or phrases to the most appropriate concept.