A multi-agent support system repeats similar policy questions and identical FAQ answers thousands of times daily. Token spend is high. Which caching strategy design best fits multi-agent optimization on Azure?
Select an answer to reveal the explanation.
Short Explanation
Stack the caches by what is safe to reuse: prompt-prefix caching for stable instructions/tools, semantic cache for near-duplicate asks, response cache for deterministic FAQs — with TTLs and bust-on-policy-change. Option B. No cache (A) wastes money. Caching payment side effects (C) is dangerous. One global answer for all tenants (D) breaks isolation and correctness.
Full Explanation
AI-500 caching strategy includes prompt caching, semantic caching, and response caching. Option B applies each where appropriate and ties invalidation to knowledge changes.
Option A ignores cost/latency optimization requirements.
Option C confuses read-through answer caching with replaying irreversible actions.
Option D violates tenancy and relevance.
Invalidate semantic/response caches when policy documents change; never cache authorized privileged answers across users without ACL-aware keys.
Exam tip: Cache answers and prefixes, not unvalidated side effects.