For knowledge-intensive vertical applications, maintaining an updatable knowledge base combined with retrieval is more economical and flexible than frequently fine-tuning the model, as fine-tuning requires hours of GPU computation while updating the knowledge base only requires a few minutes of vectorization and indexing operations.