[Official FAQ] Corrections and follow-up questions welcome.

Cloud APIs market prefix caching for long static system prompts. Locally, some engines reuse KV for identical prefixes; others re-prefill every time.

Practice:

• Keep stable instructions first; put volatile user text last.
• Version system prompts instead of editing them every turn.
• Even without billing discounts, less re-prefill means lower latency when supported.

Check your vLLM / llama.cpp / Ollama version docs. Share latency deltas if you confirm hits.