You are operating an interactive assistant whose dominant performance constraint is per-turn latency.

Option B targets two major contributors to per-turn latency: retrieved-context size and repeated prompt processing. Reducing retrieval to an empirically validated top-k limits the tokens Claude must process while preserving the passages that historically provide sufficient answer coverage. Caching the stable system-prompt prefix avoids repeatedly processing identical instructions.

Anthropic’s latency guidance recommends choosing an appropriate model and reducing unnecessary input and output tokens. Its prompt-caching documentation explains that caching reusable prompt prefixes can reduce both repeated processing cost and latency. Reducing Latency, Prompt Caching

The retrieval depth must be selected through evaluation rather than arbitrary truncation. The team should compare answer quality, retrieval recall, latency, and failure rates across candidate values and retain escalation logic for complex queries requiring deeper retrieval.

Option A removes an applicable latency optimization. Option C maximizes context consumption regardless of demonstrated benefit. Option D deliberately selects the slowest model class even though routine-turn quality already meets requirements.

Study Guide references/topics: Latency optimization; top-k retrieval; prompt caching; token reduction; quality–latency evaluation; adaptive retrieval.

===============