フリー問題

問題文

A team wants both citations and the lowest possible latency. Why can these not both be maximized?

選択肢

1. Citations increase the max output tokens requirement beyond any model's limit, since every claim in the answer has to carry the passage it came from
2. Citations require a provisioned deployment, which is slower because the request has to be routed to the reserved capacity rather than the nearest available instance
3. Citations require fine-tuning, which increases per-request latency because the customized weights have to be loaded before the request can be processed
4. Producing citations requires a retrieval step whose latency cannot be removed without giving up the citations, so only its speed and count can shrink