@tencent-rtc/mcp 1.7.0 → 1.7.1

package/knowledge/search/url-row-scorer.js

@@ -1,258 +1,337 @@
-const E7_EN_STOPWORDS = new Set([
-    'the', 'and', 'for', 'with', 'from', 'into', 'about', 'using', 'via',
-    'how', 'what', 'where', 'when', 'which', 'please', 'docs', 'documentation',
-    'tencent', 'im', 'rest', 'server', 'webhook', 'webhooks',
-]);
-/** Generic terms — excluded from API-token hits to avoid false positives */
-const E7_GENERIC_TERMS = new Set([
-    'message', 'messages', 'group', 'groups', 'user', 'users', 'account', 'accounts',
-    'send', 'get', 'set', 'delete', 'create', 'update', 'query', 'before', 'after',
-    'webhook', 'friend', 'single', 'multiple', 'history', 'chat', 'sdk', 'api',
-    'profile', 'conversation', 'import', 'recall', 'modify', 'pull', 'mark', 'data',
-    'list', 'online', 'status', 'custom', 'attributes', 'counter', 'member', 'members',
-    'management', 'overview', 'integration', 'initialization', 'billing', 'topic',
-    'topics', 'community', 'robot', 'chatbot', 'streaming', 'broadcast', 'official',
-    'extension', 'receipt', 'receipts', 'global', 'mute', 'ban', 'unban', 'forbid',
-    'ordinary', 'specified', 'historical', 'unread', 'read', 'configure', 'configuration', 'invalidate',
-    'invalidating', 'disband', 'change', 'search', 'cloud', 'local', 'robot', 'bots',
-]);
-const E7_QUERY_NOISE = /\b(rest\s*ful?\s*api|rest\s*api|server\s*api|webhook|webhooks|how\s+to|how\s+do\s+i|what\s+is|where\s+is|please|documentation|docs?)\b/gi;
-const E7_QUERY_NOISE_CN = /怎么|如何|什么|用哪个|查询|服务端/g;
-const E7_PHRASES = [
-    'before send one-to-one message',
-    'after send one-to-one message',
-    'one-to-one message',
-    'group message',
-    'recall group message',
-    'recall one-to-one message',
-    'send one-to-one message',
-    'send single chat message',
-    'getmessagelist',
-    'message history',
-    'add friend',
-    'import single account',
-    'import a single account',
-    'multi-device login',
-    'multi device login',
-    'login policy',
-    'login strategy',
-    'message storage duration',
-    'roaming message storage',
-];
-function extractRestEndpoints(text) {
-    return [...text.matchAll(/v4\/[\w/]+/gi)].map((m) => m[0].toLowerCase());
-}
-function extractE7ApiTokens(text) {
+/**
+ * url-row-scorer.ts
+ *
+ * URL 映射行（chat/urls/...）匹配打分器。
+ *
+ * 设计目标（重构版）：
+ * - **语义词对齐**：按「query 与 title 的高区分度语义特征是否对齐」打分，
+ *   而不是「子串/单字」粗匹配。
+ * - **IDF 自动降权**：高频泛词（`消息`/`message`/`group`）靠 IDF 自动低权重，
+ *   稀有语义词（`消息回应`/`reaction`）高权重，**不依赖枚举词表**。
+ * - **中英通吃、产品无关**：同一套特征提取同时处理中文/英文标题。
+ * - 保留已验证有效的 REST 端点 / API token 高分路径。
+ */
+import { idfWeight, isValidBigram, loadIdf } from './url-row-idf.js';
+/**
+ * 提取 CJK 有效 2-gram（df>0，过滤「送前」「前回」等滑窗噪声）。
+ * 口径与旧版 qBigrams / titleBigramArray **bit-identical**。
+ */
+function extractCjkTokens(text) {
     const tokens = new Set();
-    for (const match of text.matchAll(/\b([a-z][a-z0-9_]{2,})\b/gi)) {
-        const token = match[1].toLowerCase();
-        if (!E7_GENERIC_TERMS.has(token) && !E7_EN_STOPWORDS.has(token))
-            tokens.add(token);
-    }
-    for (const match of text.matchAll(/\b([A-Z][A-Za-z0-9]*(?:\.[A-Za-z]\w*)?)\b/g)) {
-        const token = match[1].toLowerCase();
-        if (!E7_GENERIC_TERMS.has(token))
-            tokens.add(token);
-    }
-    for (const match of text.matchAll(/([a-z]+[A-Z][A-Za-z0-9]*)/g)) {
-        tokens.add(match[1].toLowerCase());
-    }
-    return [...tokens];
-}
-function normalizeE7Text(text) {
-    return text
-        .toLowerCase()
-        .replace(E7_QUERY_NOISE, ' ')
-        .replace(E7_QUERY_NOISE_CN, '')
-        .replace(/[^\w\u4e00-\u9fff/]+/g, ' ')
-        .replace(/\s+/g, ' ')
-        .trim();
-}
-function splitPromptIntoClauses(prompt) {
-    const raw = prompt
-        .replace(/rest\s*api|服务端|server\s*api|api|怎么|如何|什么|用哪个|查询/g, '')
-        .replace(/\s+/g, ' ')
-        .trim();
-    const parts = raw.split(/[、，,；;、\s]+/).filter((p) => p.length >= 2);
-    return parts.length > 0 ? parts : [raw];
-}
-function phraseBonus(query, title) {
-    const q = query.toLowerCase().replace(/\s+/g, ' ');
-    const t = title.toLowerCase();
-    let bonus = 0;
-    for (const phrase of E7_PHRASES) {
-        if (q.includes(phrase) && t.includes(phrase))
-            bonus += 180;
-    }
-    if (/send single chat message/.test(q) && /send one-to-one message/.test(t))
-        bonus += 200;
-    if (/recall group message/.test(q) && /recall group message/.test(t))
-        bonus += 200;
-    if (/before send one-to-one message/.test(q) && /before send one-to-one message/.test(t))
-        bonus += 220;
-    if (/\bbefore\b/.test(q) && t.startsWith('before '))
-        bonus += 120;
-    if (/\bafter\b/.test(q) && t.startsWith('after '))
-        bonus += 120;
-    if (/多端登录/.test(query) && /多端登录|multi[\s-]?device/i.test(title))
-        bonus += 220;
-    if (/登录策略|login\s*(policy|strategy)/i.test(query) && /登录策略|login\s*(policy|strategy)/i.test(title))
-        bonus += 200;
-    return bonus;
-}
-function scoreEnglishStyle(query, title) {
-    const queryEndpoints = extractRestEndpoints(query);
-    const titleEndpoints = extractRestEndpoints(title);
-    for (const qe of queryEndpoints) {
-        if (titleEndpoints.some((te) => te === qe || te.endsWith(qe) || qe.endsWith(te))) {
-            return 1000;
+    for (const seg of text.toLowerCase().match(/[\u4e00-\u9fff]{2,}/g) ?? []) {
+        for (let i = 0; i < seg.length - 1; i += 1) {
+            const bg = seg.slice(i, i + 2);
+            if (isValidBigram(bg))
+                tokens.add(bg);
         }
     }
-    let score = phraseBonus(query, title);
-    const titleLower = title.toLowerCase();
-    const queryApis = extractE7ApiTokens(query);
-    let apiHits = 0;
-    let longApiHit = false;
-    for (const api of queryApis) {
-        if (!titleLower.includes(api))
+    return tokens;
+}
+/**
+ * 提取 EN 有效词（>=4 字母，形态过滤 web/h5/sdk/api/im/tim/app/pc/mini/v\d+）。
+ * 仅在 query **纯 EN**（无 CJK）时启用，避免混合 query 的 EN 翻译
+ * （如「单聊」展开成「one-to-one c2c single chat message」后,无对应 title 命中,
+ * 会稀释 CJK 主干的 coverage 分母,导致「单聊/群聊」互抢位）。
+ */
+function extractEnTokens(text) {
+    const tokens = new Set();
+    for (const m of text.toLowerCase().matchAll(/[a-z][a-z0-9]{2,}/g)) {
+        const w = m[0];
+        if (w.length < 4)
+            continue;
+        if (/^(web|h5|sdk|api|im|tim|app|pc|mini|v\d+)$/.test(w))
             continue;
-        apiHits += 1;
-        if (api.length >= 8)
-            longApiHit = true;
+        tokens.add(w);
     }
-    if (longApiHit)
-        score = Math.max(score, 940);
-    else if (apiHits >= 2)
-        score = Math.max(score, 900);
-    else if (apiHits === 1)
-        score = Math.max(score, 860);
-    const qNorm = normalizeE7Text(query);
-    const tNorm = normalizeE7Text(title);
-    const qTokens = qNorm.split(' ').filter((t) => t.length >= 3 && !E7_GENERIC_TERMS.has(t));
-    const tTokens = tNorm.split(' ').filter((t) => t.length >= 3 && !E7_GENERIC_TERMS.has(t));
-    let overlap = 0;
-    for (const qt of qTokens) {
-        if (tTokens.some((tt) => tt.includes(qt) || qt.includes(tt)))
-            overlap += 1;
+    return tokens;
+}
+/**
+ * 计算单语言路径的语义打分（coverage + meaningfulRatio + titleCoverage 三维度）。
+ * 对纯 CJK 输入，与旧版 4 分支结构 + 公式 **bit-identical**。
+ * 对纯 EN 输入，引入 IDF 调节（旧版只走 650+30*N 无 IDF 导致泛词虚高）。
+ */
+function scoreOneLang(qTokens, tTokens) {
+    const queryTokenCount = qTokens.size;
+    const titleTokenCount = tTokens.size;
+    if (queryTokenCount === 0) {
+        return { coverage: 0, titleCoverage: 0, meaningfulRatio: 0, matchedCount: 0, titleTokenCount };
     }
-    if (overlap >= 3)
-        score = Math.max(score, 720 + overlap * 15);
-    else if (overlap >= 2)
-        score = Math.max(score, 650);
-    else if (overlap >= 1)
-        score = Math.max(score, 520);
-    const qCompact = qNorm.replace(/\s+/g, '');
-    const tCompact = tNorm.replace(/\s+/g, '');
-    if (qCompact.length >= 5 && tCompact.includes(qCompact)) {
-        score = Math.max(score, 560 - Math.min(tCompact.length - qCompact.length, 80));
+    let matchedCount = 0;
+    let meaningfulMatched = 0;
+    for (const t of qTokens) {
+        if (tTokens.has(t)) {
+            matchedCount += 1;
+            if (idfWeight(t) > 0.5)
+                meaningfulMatched += 1;
+        }
     }
-    else if (tCompact.length >= 5 && qCompact.includes(tCompact)) {
-        score = Math.max(score, 500 + Math.min(tCompact.length, 30));
+    const coverage = matchedCount / queryTokenCount;
+    const titleCoverage = titleTokenCount > 0 ? matchedCount / titleTokenCount : 0;
+    const meaningfulRatio = meaningfulMatched / queryTokenCount;
+    return { coverage, titleCoverage, meaningfulRatio, matchedCount, titleTokenCount };
+}
+/** REST 端点（v4/xxx）精确匹配 */
+function restEndpointHit(query, title) {
+    const qe = [...query.matchAll(/v4\/[\w/]+/gi)].map((m) => m[0].toLowerCase());
+    const te = [...title.matchAll(/v4\/[\w/]+/gi)].map((m) => m[0].toLowerCase());
+    if (!qe.length)
+        return 0;
+    for (const q of qe) {
+        if (te.some((t) => t === q || t.endsWith(q) || q.endsWith(t)))
+            return 1000;
     }
-    const qCjk = query.replace(/[^\u4e00-\u9fff]/g, '');
-    const tCjk = title.replace(/[^\u4e00-\u9fff]/g, '');
-    if (qCjk.length >= 2 && tCjk.includes(qCjk)) {
-        let cjkScore = 500;
-        if (/批量|群组|群流|广播|导入|公众号|已读|回执/.test(tCjk) && !/批量|群组|群|已读|回执/.test(qCjk)) {
-            cjkScore -= 250;
+    return 0;
+}
+/**
+ * API token（驼峰 / 点号 / useXxx）精确匹配
+ *
+ * **V2 严格命中（修复 X message row-to-row tie）**：
+ * 旧版 `lower.includes(api)` 会把"共享部分 API"也计入命中，导致所有
+ * `X message (createXMessage → sendMessage)` 标题因共享 `sendMessage` 都得 940，
+ * BM25 决定胜负，倾向"Text message"抢首位（41+ miss）。
+ *
+ * 新版要求 query 和 title 的 **long API（≥8 字符）集合完全一致**（双向 ⊆）
+ * 才算 940。这样：
+ * - query "Send message parameter configuration (sendMessage parameters)" → long={sendMessage}
+ *   - self title: long={sendMessage} → 940
+ *   - "Text message" title: long={createTextMessage, sendMessage} → sizes differ → 0
+ *   - 共享 `sendMessage` 但缺独有 API（如 createTextMessage）的 title 拿 0 分
+ * - query "Image message (createImageMessage → sendMessage)" → long={createImageMessage, sendMessage}
+ *   - self title: long={createImageMessage, sendMessage} → 940
+ *   - "Sound message" title: long={createSoundMessage, sendMessage} → 大小同但内容不同 → 0
+ *
+ * 兼容：
+ * - restapi.md 仍命中（query 和 title 的 v4 API 路径互为完整 token 集合）。
+ * - sdk/* 的 createXMessage/sendMessage 自匹配：query 和 title 双向 ⊆ → 940。
+ */
+function apiTokenHit(query, title) {
+    // 提取 query/title 的 API token（驼峰 / useXxx / 点号）
+    const apiRe = /\b([a-z]+[A-Z][A-Za-z0-9_]*)\b|\b([A-Z][A-Za-z0-9_]*\.[A-Za-z_][A-Za-z0-9_]*)\b|\b(use[A-Z]\w*)\b/g;
+    const extract = (text) => {
+        const out = new Set();
+        for (const m of text.matchAll(apiRe)) {
+            const api = (m[1] ?? m[2] ?? m[3]).toLowerCase();
+            if (api && api.length >= 4)
+                out.add(api);
         }
-        if (/^单发/.test(tCjk) && /发送|单聊/.test(qCjk))
-            cjkScore += 120;
-        score = Math.max(score, cjkScore);
+        return out;
+    };
+    const queryApis = extract(query);
+    const titleApis = extract(title);
+    if (queryApis.size === 0)
+        return 0;
+    // 只看 long API（≥8 字符）的双向 ⊆ 严格命中要求
+    const longQuery = new Set([...queryApis].filter((a) => a.length >= 8));
+    if (longQuery.size === 0) {
+        // 没有 long API 的 query 走 hits 计数（兼容短 token 场景，如 useXxx）
+        let hits = 0;
+        for (const a of queryApis)
+            if (titleApis.has(a))
+                hits += 1;
+        if (hits >= 2)
+            return 900;
+        if (hits === 1)
+            return 860;
+        return 0;
     }
-    if (score <= 0)
+    const longTitle = new Set([...titleApis].filter((a) => a.length >= 8));
+    // 严格双向命中：query.long === title.long（Set 视角）
+    if (longQuery.size !== longTitle.size)
         return 0;
-    const queryL = query.toLowerCase();
-    const titleL = title.toLowerCase();
-    if (/group/.test(queryL) && /one-to-one|c2c/.test(titleL) && !/group/.test(titleL))
-        score -= 400;
-    if (/one-to-one|c2c|single chat/.test(queryL) && /import.*account|add friend/.test(titleL))
-        score -= 400;
-    if (/send.*message|one-to-one message/.test(queryL) && /add friend|import.*account/.test(titleL))
-        score -= 400;
-    if (/getmessagelist|message history/.test(qCompact) && /integration|initialization/.test(titleL) && !/getmessagelist|history/.test(tCompact))
-        score -= 400;
-    if (/extension/.test(titleL) && !/extension/.test(queryL))
-        score -= 350;
-    if (/\bbefore\b/.test(queryL) && !/^before\b/.test(titleL.trim()))
-        score -= 250;
-    if (/\bafter\b/.test(queryL) && !/^after\b/.test(titleL.trim()))
-        score -= 250;
-    if (/batch|multiple/.test(titleL) && !/batch|multiple/.test(queryL))
-        score -= 200;
-    if (/group/.test(titleL) && !/group|channel/.test(queryL) && /one-to-one|c2c|single chat/.test(queryL))
-        score -= 200;
-    return Math.max(score, 0);
+    for (const a of longQuery)
+        if (!longTitle.has(a))
+            return 0;
+    return 940;
 }
-function scoreChineseClauseStyle(query, title) {
-    const clauses = splitPromptIntoClauses(query.toLowerCase());
-    const titleNorm = title.toLowerCase().replace(/\s+/g, '');
-    if (!titleNorm)
-        return 0;
-    let bestScore = 0;
-    const allParts = [...clauses, query.toLowerCase().replace(/\s+/g, '')];
-    const titleTokens = titleNorm.split(/[\/（(（、,，\s]+/).filter(Boolean);
-    const seen = new Set();
-    for (const part of allParts) {
-        const pn = part.replace(/\s+/g, '');
-        if (pn.length < 2 || seen.has(pn))
-            continue;
-        seen.add(pn);
-        let s = 0;
-        if (pn === titleNorm) {
-            s = 1000;
+/**
+ * 语义词对齐打分（双路径版：CJK 路径与 EN 路径独立计算）
+ *
+ * 设计要点（与旧版兼容性 vs 修复 en 端虚高）：
+ * 1. **CJK 路径**：与旧版 4 分支 + 公式 **bit-identical**。
+ *    混合 query（query expansion 注入 EN 翻译后）只走 CJK 路径,EN tokens 不参与
+ *    分母计算,避免「单聊/群聊」互抢位。
+ * 2. **EN 路径**：仅在 query 无 CJK 时启用（即纯 EN query）。
+ *    引入 IDF 调节:旧版 enWordHits 路径「650+30*N」无 IDF,导致 en 端 row 虚高抢位;
+ *    新版用 coverage+meaningfulRatio 与 CJK 路径同口径打分。
+ * 3. **惩罚 1/2**：仅当 score > 0 时生效,纯 EN query 不触发 CJK 整段包含惩罚。
+ *
+ * 关键修复：「修改单聊历史消息 one-to-one c2c single chat message messages」(扩 query)
+ *   旧版: enWordHits=0 → 走 CJK 路径 → 自己 700, 对偶 643
+ *   新版: CJK 路径独立 → 自己 700, 对偶 643 (**bit-identical**)
+ * 纯 EN query: 新版用 coverage+IDF 替代旧 650+30*N,降权泛词。
+ */
+function semanticAlign(query, title) {
+    const qCjk = extractCjkTokens(query);
+    const tCjk = extractCjkTokens(title);
+    const qEn = extractEnTokens(query);
+    const tEn = extractEnTokens(title);
+    let score = 0;
+    if (qCjk.size > 0) {
+        // CJK 路径:与旧版公式 bit-identical
+        const m = scoreOneLang(qCjk, tCjk);
+        if (m.coverage >= 0.5 && m.meaningfulRatio >= 0.5) {
+            score = 500 + Math.round(m.coverage * 120) + Math.round(m.meaningfulRatio * 80);
         }
-        else if (titleNorm.includes(pn)) {
-            const trailing = titleNorm.slice(titleNorm.indexOf(pn) + pn.length);
-            s = 800 - trailing.length * 80;
+        else if (m.titleTokenCount >= 3 && m.titleCoverage >= 0.7) {
+            score = 620 + Math.round(m.titleCoverage * 60);
+        }
+        else if (m.coverage >= 0.3 && m.meaningfulRatio >= 0.3) {
+            score = 400;
+        }
+    }
+    else if (qEn.size > 0) {
+        // EN 路径:仅纯 EN query,引入 IDF 调节(旧版 650+30*N 无 IDF 致虚高)
+        const m = scoreOneLang(qEn, tEn);
+        if (m.coverage >= 0.5 && m.meaningfulRatio >= 0.5) {
+            score = 500 + Math.round(m.coverage * 120) + Math.round(m.meaningfulRatio * 80);
         }
-        else if (pn.includes(titleNorm)) {
-            s = 700 + titleNorm.length;
+        else if (m.titleTokenCount >= 3 && m.titleCoverage >= 0.7) {
+            score = 620 + Math.round(m.titleCoverage * 60);
         }
-        else if (titleTokens.some((token) => pn.includes(token) || token.includes(pn))) {
-            s = 600;
+        else if (m.coverage >= 0.3) {
+            // 弱命中:旧版无此分支,给软基分 300-400
+            score = 300 + Math.round(m.coverage * 200);
         }
-        if (s > bestScore)
-            bestScore = s;
     }
-    if (bestScore <= 0 && clauses.length > 0) {
-        const qn = clauses[0].replace(/发送/g, '发');
-        if (qn.length >= 3 && titleNorm.includes(qn)) {
-            bestScore = 500;
-            if (/批量|群组|群流|广播|导入|公众号|已读|回执/.test(titleNorm)
-                && !/批量|群组|群|已读|回执/.test(clauses[0])) {
-                bestScore -= 250;
+    if (score === 0)
+        return 0;
+    // 惩罚1:CJK 路径下,query 有明确 CJK 语义但覆盖双低(coverage<0.5 && titleCoverage<0.7),
+    // 封顶 400,避免泛词残留抬分。EN 路径不触发。
+    if (qCjk.size > 0) {
+        const qHasCjk3 = /[\u4e00-\u9fff]{3,}/.test(query);
+        if (qHasCjk3) {
+            const m = scoreOneLang(qCjk, tCjk);
+            if (m.coverage < 0.5 && m.titleCoverage < 0.7) {
+                score = Math.min(score, 400);
             }
-            if (/^单发/.test(titleNorm) && /发送|单聊/.test(clauses[0])) {
-                bestScore += 120;
+        }
+    }
+    // 惩罚2:query 的完整 CJK 整段(≥4 字)「完全包含于」title,但 title 明显更长
+    // (多出 ≥3 字独立主题词),说明 title 主题更窄/偏移 → 降权。
+    // 例:query「发送单聊消息」vs title「发送单聊消息已读回执」→ 后者多出「已读回执」。
+    if (qCjk.size > 0) {
+        const qSegments = query.match(/[\u4e00-\u9fff]{3,}/g) ?? [];
+        const titleNorm = title.replace(/\s+/g, '');
+        const qFullCjk = qSegments.filter((s) => s.length >= 4 && titleNorm.includes(s));
+        if (qFullCjk.length > 0) {
+            const titleCjkLen = (title.match(/[\u4e00-\u9fff]/g) ?? []).length;
+            const queryCjkLen = (query.match(/[\u4e00-\u9fff]/g) ?? []).length;
+            const extraCjk = titleCjkLen - queryCjkLen;
+            if (extraCjk >= 3) {
+                score = Math.max(360, score - 120 * Math.floor(extraCjk / 2));
             }
         }
     }
-    return Math.max(bestScore, 0);
+    return score;
+}
+/** 兼容入口：原有 normalize 逻辑保留（供 exact-lookup 调用） */
+function normalizeE7Text(text) {
+    return text
+        .toLowerCase()
+        .replace(/[^\w\u4e00-\u9fff/]+/g, ' ')
+        .replace(/\s+/g, ' ')
+        .trim();
 }
 export function normalizeUrlRowQuery(query) {
     return normalizeE7Text(query);
 }
+/**
+ * 产品无关的 URL 映射行判定：knowledge/<product>/urls/.../xxx.md
+ * 供 exact-lookup / l1-enrich / doc-url-resolver 复用，避免各处硬编码 chat/urls。
+ */
+export function isUrlMappingSource(source) {
+    return /knowledge\/[a-z0-9_-]+\/urls\/.+\.md$/i.test(source.replace(/\\/g, '/'));
+}
+/**
+ * 对 URL 映射行打分。
+ * 优先级：REST 端点 > API token > 语义词对齐。
+ */
 export function scoreUrlRow(query, title) {
-    const enScore = scoreEnglishStyle(query, title);
-    const zhScore = scoreChineseClauseStyle(query, title);
-    if (enScore <= 0 && zhScore <= 0) {
-        return {
-            score: 0,
-            matchedBy: 'none',
-            debug: { enScore, zhScore },
-        };
+    loadIdf();
+    const restScore = restEndpointHit(query, title);
+    const apiScore = restScore > 0 ? 0 : apiTokenHit(query, title);
+    const semanticScore = restScore > 0 || apiScore >= 860 ? 0 : semanticAlign(query, title);
+    if (restScore > 0) {
+        return { score: restScore, matchedBy: 'rest', debug: { restScore, apiScore, semanticScore } };
     }
-    if (zhScore > enScore) {
-        return {
-            score: zhScore,
-            matchedBy: 'zh',
-            debug: { enScore, zhScore },
-        };
+    if (apiScore > 0) {
+        return { score: apiScore, matchedBy: 'api', debug: { restScore, apiScore, semanticScore } };
+    }
+    if (semanticScore > 0) {
+        return { score: semanticScore, matchedBy: 'semantic', debug: { restScore, apiScore, semanticScore } };
+    }
+    return {
+        score: 0,
+        matchedBy: 'none',
+        debug: { restScore, apiScore, semanticScore },
+    };
+}
+/** 从 chunk body 中提取开头的分组标题（chunker 会把 urls 行的 `##` 分组写成 `### 分组标题` 前缀） */
+export function extractSectionTitle(content) {
+    const match = content.match(/^\s*###\s+([^\n]+)/);
+    return match ? match[1].trim() : '';
+}
+/**
+ * 计算一组分组标题的「共有词」（出现在全部分组标题中的单字/2-gram，无区分度）。
+ * 产品无关：对任意 urls 文件的 `##` 分组自动判定，不依赖枚举。
+ */
+export function computeCommonSectionTokens(sectionTitles) {
+    const n = sectionTitles.length;
+    if (n === 0)
+        return { chars: new Set(), bigrams: new Set() };
+    const charFreq = new Map();
+    const bigramFreq = new Map();
+    for (const title of sectionTitles) {
+        const chars = new Set(title.match(/[\u4e00-\u9fff]/g) ?? []);
+        const bigrams = new Set();
+        for (const seg of title.match(/[\u4e00-\u9fff]{2,}/g) ?? []) {
+            for (let i = 0; i < seg.length - 1; i += 1)
+                bigrams.add(seg.slice(i, i + 2));
+        }
+        for (const c of chars)
+            charFreq.set(c, (charFreq.get(c) ?? 0) + 1);
+        for (const b of bigrams)
+            bigramFreq.set(b, (bigramFreq.get(b) ?? 0) + 1);
     }
     return {
-        score: enScore,
-        matchedBy: 'en',
-        debug: { enScore, zhScore },
+        chars: new Set([...charFreq].filter(([, v]) => v === n).map(([k]) => k)),
+        bigrams: new Set([...bigramFreq].filter(([, v]) => v === n).map(([k]) => k)),
     };
 }
+/**
+ * 判定 query 是否命中分组标题（通用分组语义粗筛）。
+ *
+ * 规则（混用单字 + 2-gram，至少 2 个任意词 + 至少 1 个区分词）：
+ * - 提取 query 与分组标题的中文单字集、2-gram 集。
+ * - 排除「共有词」（该文件所有分组都出现的词，如 webhook 里的「回调/相关」），避免其单独触发命中。
+ * - 命中条件：query 与分组标题共享「≥1 个区分词」且「≥2 个任意词（区分+共有）」。
+ *
+ * 例：query「服务端群消息发送前回调」vs 分组「群组系统相关回调」：
+ *   共享单字「群/回/调」，排除共有「回/调」后区分词「群」命中，且任意词 3 个 → 命中。
+ *   而「在线状态相关回调」只共享共有词「回/调」→ 不命中。
+ *
+ * 产品无关：对所有产品的 urls 分组都适用，不依赖具体产品词表。
+ */
+export function queryMatchesSection(query, sectionTitle, commonTokens) {
+    const queryChars = new Set(query.match(/[\u4e00-\u9fff]/g) ?? []);
+    const queryBigrams = new Set();
+    for (const seg of query.match(/[\u4e00-\u9fff]{2,}/g) ?? []) {
+        for (let i = 0; i < seg.length - 1; i += 1)
+            queryBigrams.add(seg.slice(i, i + 2));
+    }
+    const sectionChars = new Set(sectionTitle.match(/[\u4e00-\u9fff]/g) ?? []);
+    const sectionBigrams = new Set();
+    for (const seg of sectionTitle.match(/[\u4e00-\u9fff]{2,}/g) ?? []) {
+        for (let i = 0; i < seg.length - 1; i += 1)
+            sectionBigrams.add(seg.slice(i, i + 2));
+    }
+    // 共享单字 + 共享 2-gram（2-gram 命中算 2 个词）
+    const sharedChar = [...sectionChars].filter((c) => queryChars.has(c));
+    const sharedBigram = [...sectionBigrams].filter((b) => queryBigrams.has(b));
+    // 任意词（共享单字按 1 个词、共享 2-gram 按 1 个词计；2-gram 与单字不重复计）
+    const anyCount = sharedChar.length + sharedBigram.length;
+    // 区分词：共享但不在共有词集合里
+    const distinctCount = sharedChar.filter((c) => !commonTokens.chars.has(c)).length
+        + sharedBigram.filter((b) => !commonTokens.bigrams.has(b)).length;
+    return distinctCount >= 1 && anyCount >= 2;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tencent-rtc/mcp",
-  "version": "1.7.0",
+  "version": "1.7.1",
   "description": "Tencent Cloud MCP Server for SDK",
   "type": "module",
   "main": "index.js",
@@ -20,12 +20,14 @@
     "@types/uuid": "^10.0.0",
     "agentkeepalive": "^4.1.4",
     "axios": "^0.25.0",
+    "cheerio": "^1.2.0",
     "debug": "^4.3.1",
     "express": "^5.1.0",
     "hashring": "^3.2.0",
     "protobufjs": "^6.11.2",
     "semver": "^7.3.5",
     "shx": "^0.3.4",
+    "turndown": "^7.2.4",
     "typescript": "^5.8.2",
     "uuid": "^8.3.2",
     "zod": "^3.24.2"