A retail company is using a foundation model trained on general internet text. They now want it to generate product descriptions tailored to their unique catalog and brand tone. The model must learn from existing product descriptions without retraining from scratch. Which approach should they use?