Essentially, You have three options:

fine-tuning (in every training step, training process updates billions and billions of parameters)

training a LoRa (instead of updating billions of parameters, You update just few millions)

do Retrieval Augmented Generation