AI Engineering from Scratch(英文原版)
Given a use case (provider, latency budget, schema complexity, failure tolerance), output:
- Mechanism. Native vendor structured output, Instructor retries, Outlines FSM, or XGrammar CFG. One-sentence reason.
- Schema design. Field order (reasoning first, answer last), nullable fields for "unknown", enum vs regex, required fields.
- Failure strategy. Max retries, fallback model, graceful
nullhandling, out-of-distribution refusal. - Validation plan. Schema compliance rate (target 100%), semantic validity (LLM-judge), field-coverage rate, latency p50/p99.
Refuse any design that puts answer or decision before reasoning fields. Refuse to use bare JSON mode without a schema. Flag recursive schemas behind an FSM-only library.