1. What is the primary tradeoff when choosing between a hosted LLM API and self-hosting an open-weight model?
2. What is the purpose of a fallback chain in an LLM application?
3. Why is streaming output preferable for long LLM responses in user-facing applications?
4. What architectural pattern separates the LLM's role from application business logic?
5. What is 'context window management' and why does it matter?
6. What does 'latency vs. throughput' mean as a scaling tradeoff for LLM inference?
7. What is 'graceful degradation' in an AI system?
8. Why should idempotency be a design goal for AI agent actions?