@tencentcloud/sdk-mcp 1.7.7 → 1.7.8

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

@@ -54,7 +54,14 @@ 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 };
+        return {
+            coverage: 0,
+            titleCoverage: 0,
+            meaningfulRatio: 0,
+            matchedCount: 0,
+            meaningfulMatched: 0,
+            titleTokenCount,
+        };
     }
     let matchedCount = 0;
     let meaningfulMatched = 0;
@@ -68,7 +75,7 @@ function scoreOneLang(qTokens, tTokens) {
     const coverage = matchedCount / queryTokenCount;
     const titleCoverage = titleTokenCount > 0 ? matchedCount / titleTokenCount : 0;
     const meaningfulRatio = meaningfulMatched / queryTokenCount;
-    return { coverage, titleCoverage, meaningfulRatio, matchedCount, titleTokenCount };
+    return { coverage, titleCoverage, meaningfulRatio, matchedCount, meaningfulMatched, titleTokenCount };
 }
 /** REST 端点（v4/xxx）精确匹配 */
 function restEndpointHit(query, title) {
@@ -169,10 +176,19 @@ function semanticAlign(query, title) {
     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) {
+        if (m.coverage >= 0.5 && m.meaningfulRatio >= 0.5 && m.matchedCount >= 2) {
+            // 收紧(方案 A′):第一分支要求 matchedCount ≥ 2,避免 query 端 isValidBigram 过滤后
+            // 分母极小(如 qCjk={本地,日志} 仅含 2 个词,匹配 1 个即达 coverage=0.5)造成虚高。
+            // 场景:query「web im sdk 怎么关闭浏览器本地日志」→ qCjk 经 isValidBigram 过滤仅剩
+            //   {本地,日志}(关闭/浏览/览器/地日 在 urls 语料 df=0 被剔除);
+            //   vs title「V3 版本日志 / Changelog」共享 {日志} → matchedCount=1 → 不再走此分支;
+            //   → 降级到「else if coverage≥0.3 && meaningfulRatio≥0.3」分支拿 400 底分。
+            // 兼容:多 token 强对齐场景(如「屏幕分享」vs「屏幕分享 / Screen Share」)matchedCount 天然 ≥2。
             score = 500 + Math.round(m.coverage * 120) + Math.round(m.meaningfulRatio * 80);
         }
-        else if (m.titleTokenCount >= 3 && m.titleCoverage >= 0.7) {
+        else if (m.titleTokenCount >= 3 && m.titleCoverage >= 0.7 && m.meaningfulMatched >= 1) {
+            // 收紧(方案 A):titleCoverage 高覆盖分支要求至少命中 1 个高区分度词(idf>0.5),
+            // 避免共享 token 全是跨语料泛词但恰好覆盖短标题绝大部分造成伪高覆盖。
             score = 620 + Math.round(m.titleCoverage * 60);
         }
         else if (m.coverage >= 0.3 && m.meaningfulRatio >= 0.3) {
@@ -182,10 +198,12 @@ function semanticAlign(query, title) {
     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) {
+        if (m.coverage >= 0.5 && m.meaningfulRatio >= 0.5 && m.matchedCount >= 2) {
+            // 同 CJK 第一分支的收紧:纯 EN 端 query 过滤后短 token 集合同样有虚高风险。
             score = 500 + Math.round(m.coverage * 120) + Math.round(m.meaningfulRatio * 80);
         }
-        else if (m.titleTokenCount >= 3 && m.titleCoverage >= 0.7) {
+        else if (m.titleTokenCount >= 3 && m.titleCoverage >= 0.7 && m.meaningfulMatched >= 1) {
+            // 同 CJK 分支的收紧:纯 EN 短标题也不允许仅靠泛词(如 message/group)撑满 titleCoverage 拿高分。
             score = 620 + Math.round(m.titleCoverage * 60);
         }
         else if (m.coverage >= 0.3) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tencentcloud/sdk-mcp",
-  "version": "1.7.7",
+  "version": "1.7.8",
   "description": "Tencent Cloud MCP Server for SDK",
   "type": "module",
   "main": "index.js",
@@ -12,6 +12,7 @@
   "dependencies": {
     "@grpc/grpc-js": "^1.5.1",
     "@grpc/proto-loader": "^0.6.6",
+    "@huggingface/transformers": "^4.3.0",
     "@modelcontextprotocol/sdk": "1.27.1",
     "@types/express": "^5.0.1",
     "@types/hashring": "^3.2.1",
@@ -29,6 +30,7 @@
     "turndown": "^7.2.4",
     "typescript": "^5.8.2",
     "uuid": "^8.3.2",
+    "vectra": "^0.15.0",
     "zod": "^3.24.2"
 },
   "author": "Tencent Cloud Terminal R&D Center",