feat(h5): add LLM intent router admin controls and shadow verification.

Expose shadow/canary router policy in ops admin, add FAQ rule fast-path,
and tighten router defaults (1200ms timeout, 0.65 confidence).
Includes verify-h5-llm-router-shadow for production canary rollout.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
john
2026-08-01 16:48:01 +08:00
parent 005612029f
commit 43bc8bbc2b
14 changed files with 1391 additions and 80 deletions
+124
View File
@@ -0,0 +1,124 @@
/**
* H5 Chat Intent Router — FAQ / fast-path rule library.
*
* Add patterns here for messages that should bypass LLM routing latency.
* Consumed by chat-intent-router.mjs → classifyWithRules().
*
* Keep rules conservative: task execution, page generation, and realtime lookup
* must stay on Agent paths (checked via FAQ_EXCLUSION_PATTERNS).
*/
/** Block FAQ match when the message clearly needs Agent / tools. */
export const FAQ_EXCLUSION_PATTERNS = [
/\bpublic\/[^\s"'<>]+\.html\b/i,
/MindSpace\/[^/\s]+\/public\/[^\s"'<>]+\.html/i,
/(?:生成|创建|制作|做|写|设计|发布|搭建).{0,24}(?:H5|h5|HTML|html|网页|页面|活动页|宣传页|落地页|分享页|链接)/u,
/(?:H5|h5|HTML|html|网页|页面).{0,24}(?:生成|创建|制作|做|写|设计|发布|链接)/u,
/(?:写|做|生成|制作|设计).{0,16}主题(?:页面|文章)/u,
/(?:服务号|公众号).{0,16}(?:页面|文章|推文|图文)/u,
/(?:今天|今日|现在|最新|当前|实时).{0,12}(?:新闻|赛况|比分|天气|行情|股价|汇率|热点)/u,
/(?:查(?:一下|查)?|搜索|打听).{0,16}(?:今天|今日|现在|最新|当前)/u,
/(?:世界杯|欧冠|NBA|英超).{0,16}(?:赛况|赛程|比分)/u,
/(?:load_skill|page-data|sqlite|express|后端服务)/iu,
/(?:修|改|测试|回归|verify).{0,12}(?:bug|问卷|policy|public)/iu,
];
/**
* @typedef {{ id: string, reason: string, patterns: RegExp[] }} DirectChatFaqRule
*/
/** @type {DirectChatFaqRule[]} */
export const DIRECT_CHAT_FAQ_RULES = [
{
id: 'concept_what_is',
reason: '概念解释类问答(FAQ',
patterns: [
/^什么是[\s\S]{1,48}[?]?$/u,
/^何为[\s\S]{1,48}[?]?$/u,
/^什么叫[\s\S]{1,48}[?]?$/u,
/^[\s\S]{1,40}是什么[?]?$/u,
/^[\s\S]{1,40}是什么意思[?]?$/u,
/^[\s\S]{1,40}(?:是什么东西|指什么|啥意思)[?]?$/u,
],
},
{
id: 'concept_explain',
reason: '解释说明类问答(FAQ',
patterns: [
/^(?:请)?(?:解释|说明|介绍|讲解|阐述)(?:一下|下)?[\s\S]{1,48}[?]?$/u,
/^帮我(?:解释|说明|介绍|讲解)(?:一下|下)?[\s\S]{1,48}[?]?$/u,
/^能(?:不能|否)(?:解释|说明|介绍)(?:一下|下)?[\s\S]{1,40}[?]?$/u,
],
},
{
id: 'concept_compare',
reason: '概念对比类问答(FAQ',
patterns: [
/^[\s\S]{1,32}(?:和|与|跟)[\s\S]{1,32}(?:的)?(?:区别|差异|不同|对比|比较)[是什么]?[?]?$/u,
/^(?:对比|比较)(?:一下|下)?[\s\S]{1,48}[?]?$/u,
/^[\s\S]{1,24}(?:vs|VS|Vs)[\s\S]{1,24}(?:区别|差异|不同)?[?]?$/u,
],
},
{
id: 'concept_how_works',
reason: '原理/机制类问答(FAQ',
patterns: [
/^[\s\S]{1,40}(?:是怎么|是如何)(?:工作|运行|实现|运作)的[?]?$/u,
/^[\s\S]{1,40}(?:原理|机制|流程)(?:是什么|是怎样的)[?]?$/u,
/^为什么[\s\S]{1,48}[?]?$/u,
],
},
{
id: 'text_transform',
reason: '纯文字处理(总结/翻译/改写)',
patterns: [
/^(?:请)?(?:帮我|帮忙)?(?:总结|概括|归纳|提炼)(?:一下|下)?[\s\S]{1,80}[?]?$/u,
/^(?:请)?(?:帮我|帮忙)?(?:翻译|改写|润色|扩写|缩写|校对)(?:一下|下)?[\s\S]{1,80}[?]?$/u,
],
},
{
id: 'tech_acronym',
reason: '常见技术名词解释(FAQ',
patterns: [
/^(?:什么是|解释一下|介绍一下)?\s*(?:SSE|WebSocket|REST(?:ful)?|GraphQL|JWT|OAuth|RAG|Embedding|向量数据库|Server-Sent Events)[\s\S]{0,24}[?]?$/iu,
/^(?:SSE|WebSocket|REST(?:ful)?|GraphQL|JWT|OAuth|RAG)(?:是什么|什么意思|怎么用)[?]?$/iu,
],
},
{
id: 'learning_advice',
reason: '学习/选型建议(纯文字)',
patterns: [
/^(?:学|学习|入门).{0,24}(?:应该|该|推荐|建议).{0,32}(?:先|从哪|哪个)[?]?$/u,
/^(?:初学者|新手|入门).{0,24}(?:应该|该|推荐|建议).{0,32}[?]?$/u,
/^.{0,24}(?:先学|从哪学|怎么学).{0,32}[?]?$/u,
],
},
{
id: 'casual_opinion',
reason: '主观建议/闲聊(无任务执行)',
patterns: [
/^你觉得.{1,48}[?]?$/u,
/^你认为.{1,48}[?]?$/u,
/^(?:随便|随意)(?:聊|说说|谈谈).{0,24}[?]?$/u,
/^.{0,16}(?:有什么建议|给点建议|怎么看)[?]?$/u,
],
},
];
const MAX_FAQ_TEXT_LENGTH = 200;
/**
* @param {string} text
* @returns {{ id: string, reason: string } | null}
*/
export function matchDirectChatFaqRule(text) {
const normalized = String(text ?? '').trim();
if (!normalized || normalized.length > MAX_FAQ_TEXT_LENGTH) return null;
if (FAQ_EXCLUSION_PATTERNS.some((pattern) => pattern.test(normalized))) return null;
for (const rule of DIRECT_CHAT_FAQ_RULES) {
if (rule.patterns.some((pattern) => pattern.test(normalized))) {
return { id: rule.id, reason: rule.reason };
}
}
return null;
}