Once the vector search is done for a user query, for each and every result row, a single llm is done to get raking etc. details. I wanted to understand the design considerations for the same instead of doing a batch ranking. Is it because of putting quality + precision over performance + token usage ?