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
You stack the caches by what is safe to reuse: prompt-prefix caching for stable instructions/tools, semantic cache for near-duplicate asks, and response cache for deterministic FAQs with TTLs and bust-on-policy-change. Caching nothing wastes money, caching payment side effects is dangerous, and one global answer for all tenants breaks isolation and correctness.
Full Explanation
AI-500 caching strategy includes prompt caching, semantic caching, and response caching. The combined strategy applies each where appropriate and ties invalidation to knowledge changes. Caching nothing and recomputing every prompt and tool path ignores cost and latency optimization requirements. Caching all tool side effects, including payments, for 30 days confuses read-through answer caching with replaying irreversible actions. Storing only the last user question globally and reusing one answer for every tenant violates tenancy and relevance. Invalidate semantic and 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.