@mafdet/mcp 1.4.18 → 1.4.32

package/dist/search-docs.js

@@ -1,4 +1,32 @@
 import { getJson, MafdetClientError } from "./client.js";
+const CJK = /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}]/u;
+/**
+ * Split a query into terms. Whole-phrase matching alone made every multi-word
+ * question fail: "transcribe audio" appears verbatim on no page, so an agent
+ * asking the obvious question got nothing while the answer sat one word away.
+ *
+ * Chinese needs more than splitting on spaces, because it has none:
+ * "怎么转录音频" arrives as a single token that appears nowhere, and the docs are
+ * bilingual. Any run of CJK is therefore also broken into two-character
+ * shingles, which is enough to surface 转录 and 音频 out of that question. Some
+ * shingles are meaningless (么转); they simply match nothing and cost nothing,
+ * since ranking is by how many terms hit.
+ */
+function terms(query) {
+    const out = [];
+    for (const token of query
+        .toLowerCase()
+        .split(/[^\p{L}\p{N}]+/u)
+        .filter((t) => t.length > 0)) {
+        out.push(token);
+        if (CJK.test(token) && token.length > 2) {
+            for (let i = 0; i + 2 <= token.length; i += 1) {
+                out.push(token.slice(i, i + 2));
+            }
+        }
+    }
+    return [...new Set(out)];
+}
 let cache;
 async function loadIndex(docsUrl) {
     if (cache && cache.docsUrl === docsUrl) {
@@ -24,23 +52,46 @@ export async function searchDocs(input) {
         return [];
     }
     const base = input.docsBaseForUrl ?? input.docsUrl;
+    const words = terms(needle);
     const scored = [];
     for (const doc of docs) {
         const title = doc.title.toLowerCase();
+        const headings = (doc.headings ?? []).join(" ").toLowerCase();
         const text = doc.text.toLowerCase();
-        // Title matches rank above body matches; skip non-matches entirely.
+        // The exact phrase still ranks hardest — someone who typed it means it.
         let score = 0;
         if (title.includes(needle))
-            score += 2;
+            score += 20;
         if (text.includes(needle))
-            score += 1;
+            score += 10;
+        // Then per term, so a question phrased in the reader's words still lands.
+        let matched = 0;
+        for (const word of words) {
+            let hit = 0;
+            if (title.includes(word))
+                hit += 2;
+            if (headings.includes(word))
+                hit += 1.5;
+            if (text.includes(word))
+                hit += 1;
+            if (hit > 0)
+                matched += 1;
+            score += hit;
+        }
         if (score > 0) {
-            scored.push({ doc, score });
+            scored.push({ doc, score, matched });
         }
     }
+    // Pages carrying every term first. Falling back to partial matches rather
+    // than returning nothing: one unknown word in a question should not hide the
+    // page that answers the rest of it.
+    const complete = scored.filter((s) => s.matched === words.length);
+    const ranked = complete.length > 0 ? complete : scored;
     // Stable ordering: score desc, then id asc (docs are pre-sorted by id).
-    scored.sort((a, b) => b.score - a.score || a.doc.id.localeCompare(b.doc.id));
-    return scored.slice(0, input.limit).map(({ doc }) => ({
+    ranked.sort((a, b) => b.matched - a.matched ||
+        b.score - a.score ||
+        a.doc.id.localeCompare(b.doc.id));
+    return ranked.slice(0, input.limit).map(({ doc }) => ({
         title: doc.title,
         url: new URL(doc.url, base).toString(),
         summary: doc.summary,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mafdet/mcp",
-  "version": "1.4.18",
+  "version": "1.4.32",
   "description": "Mafdet AI MCP server — read-only account tools (models, balance, usage, subscription, docs) for Claude Code, Cursor, Codex, OpenCode, and Claude Desktop.",
   "type": "module",
   "license": "MIT",