AI 智能体 llm-eval 更新于 2026-10-22

LLM Evaluation 工程规则

Eval rules: versioned prompt datasets and regression gates in CI

✓ 应当这么做

✓ Write tests against user-visible behaviour, not private implementation details
✓ Mock at network or process boundaries rather than inside your own modules
✓ Keep every test independent and deterministic

42 篇

Vector search rules: chunking strategy, recall evaluation and index rebuild path

Feature store rules: point-in-time correctness and train/serve parity

Skill design rules: narrow triggers, explicit boundaries and verifiable outputs

Tool-call rules: server-side authorisation and deny-by-default allow-lists

Memory rules: scoped retrieval, expiry and user-visible deletion

RAG rules: citation-grade retrieval and evaluated recall before shipping