In the previous chapter, we explored ways to enhance the quality of model responses, ensuring they adhere to the desired format and provide the expected answers. However, there is a limitation that is difficult to overcome—the model cannot answer what it does not know. Indeed, many models were trained on datasets that are two or three years old. A model might be unaware of recent facts related to sports, politics, economics, and so on. As a result, the model’s answers can only be considered up-to-date as of the time it was trained. A seemingly obvious solution is to retrain the model regularly on new data, ensuring it remains informed about recent events. But this is time-consuming and expensive. To implement such an approach, the training process would need to be continuous, with model checkpoints updated weekly, which is impractical. Additionally, there are tasks where the model must provide answers based on private, non-public information such as internal company documentation, specific regulations, or employee databases. Solving such a problem would require collecting a specialized dataset based on private data and retraining the model from scratch.