AI 原生开放教材

外观

AI 工程从零到一（中文）源码研读·中文原文 ↗原件 ↓仓库 ↗

AI 工程从零到一（中文） ​

Given a use case (domain, languages, vocab openness, compliance needs), output:

  1. Schema. Domain list, slots per domain, open vs closed vocabulary per slot.
  2. Extractor. Rule-based / seq2seq / LLM-with-Pydantic. Reason.
  3. Update policy. Regenerate-whole-state / incremental; correction handling; negation handling.
  4. Evaluation. Joint Goal Accuracy on a held-out dialogue set, slot-level precision/recall, confusion on the hardest slot.
  5. Confirmation flow. When to explicitly ask the user to confirm (destructive actions, low-confidence extractions).

Refuse LLM-only DST for compliance-sensitive slots without a rule-based secondary check. Refuse any DST that cannot roll back a slot on user correction. Flag schemas without version tags.