The trigger fires. Do you retrieve with the drafted sen...

Using the raw drafted sentence is a failure because the flagged span is the model’s guess, and embedding it retrieves neighbors of that guess, effectively manufacturing support for the model’s own error. Masking low-confidence tokens is free but often degenerates if the removed span is critical, such as the subject. Generating a question costs one short decode but produces an interrogative query that matches how the query encoder was trained, making it the most effective approach.