Your long-form answers are grounded for two sentences and then invent specifics. You already retrieve on the question. What do you change?

A strong answer identifies the structural cause: the information need for later sentences is not expressible from the initial question, meaning no amount of initial top-k retrieval will reach the necessary context. Instead, you should propose retrieving mid-generation based on a signal the model already produces, such as thresholding the minimum token probability of a drafted sentence. A weak answer treats this as a recall problem and simply increases k, which adds latency and distractors without addressing the drifted topic.