this is quite nuanced. in financial markets you have a combination of natural language questions that involve technical keywords / slang / acronym. and for these specific terms, an off-the-shelf embeddings model fails miserably.