Looks like it uses Googles Langextract, which uses only LLMs for NLP, while OP is using a small NER model that runs locally.