Good to see the article touching on the performance impact of a niche vs general embedding model/aggressive subword tokenization