I would assume the embedding model isn't trained on code and specific words that are industry/company specific.