A customer-service chat assistant has a p95 latency complaint from users, who mostly perceive the wait before any text appears. Usage logs show requests average a long, unchanging system prompt with tool definitions, a short user message, and replies of about 700 output tokens, mostly a polite preamble and a restated summary of the question. Which two changes most directly address the problem? Select two.

Answer: B and D — Stream responses and structure the stable prefix for prompt caching, so the first token reaches users sooner / Ask for a concise reply without the preamble or restated question, and set max_tokens accordingly

(b) and (d) are right. Streaming improves the perceived wait users complain about, and caching a long stable prefix cuts time to first token (see 2.4 for prefix structure). Generation time grows with output length, so removing unneeded preamble shortens total latency and cost. (a) adds cost and doesn't reliably shorten replies. (c) adds a sequential call, increasing latency.