@cachly-dev/mcp-server 0.5.36 → 0.5.37

package/dist/index.d.ts

@@ -6,10 +6,33 @@
  * Keeps the index small and scores meaningful.
  */
 declare const STOPWORDS: Set<string>;
+/**
+ * Convert a katakana string to Hepburn romaji.
+ * Handles digraphs, geminate consonants (ッ), and long vowel marks (ー).
+ */
+declare function katakanaToRomaji(kata: string): string;
+/**
+ * Arabic light stemmer — strips definite article and common prefix particles.
+ * Handles: ال (al-), و (wa-), ب (bi-), ل (li-), ف (fa-), ك (ka-).
+ * Only strips when the result is still ≥ 3 chars to avoid over-stemming.
+ */
+declare function arabicLightStem(word: string): string;
+declare const CROSS_LINGUAL_MAP: Map<string, string[]>;
+/**
+ * Expand a token to its cross-lingual synonyms.
+ * Returns synonyms to be added to the token stream (pre-tokenized).
+ * CJK synonyms are NOT pre-bigram'd here — caller handles that.
+ */
+declare function expandCrossLingual(token: string): string[];
 /**
  * Tokenize text into meaningful keywords.
- * Handles accented Latin characters and CJK (Chinese, Japanese, Korean).
- * CJK text is split into character bigrams (the standard for indexing logographic languages).
+ * Handles accented Latin, RTL (Arabic, Hebrew), and CJK (Chinese, Japanese, Korean).
+ *
+ * Features:
+ *  • CJK → character bigrams
+ *  • Katakana → additionally emits Hepburn romaji tokens (enables romaji queries)
+ *  • Arabic/Hebrew → word tokenization with light Arabic stemming
+ *  • All scripts → cross-lingual synonym expansion (EN↔JA↔ZH↔KO↔AR↔HE)
  */
 declare function tokenize(text: string): string[];
 /**
@@ -33,4 +56,4 @@ declare function levenshtein(a: string, b: string): number;
 declare function extractTimestamp(content: string): number | undefined;
 /** Recency multiplier: 1.0 for now, 0.5 after half-life days, exponential decay. */
 declare function recencyBoost(timestampMs: number | undefined): number;
-export { tokenize, splitMultiQuery, levenshtein, recencyBoost, extractTimestamp, STOPWORDS };
+export { tokenize, splitMultiQuery, levenshtein, recencyBoost, extractTimestamp, STOPWORDS, katakanaToRomaji, arabicLightStem, expandCrossLingual, CROSS_LINGUAL_MAP };

package/dist/index.js

@@ -288,8 +288,11 @@ function embedProviderHint() {
 //   • Recency boost — newer entries rank higher (exponential decay, 7-day half-life)
 //   • Multi-query splitting — numbered lists, semicolons, conjunctions
 //   • Fuzzy matching with Levenshtein distance ≤ 2
-//   • Multilingual stopwords (EN, DE, FR, ES, IT, PT, ZH, JA, KO)
+//   • Multilingual stopwords (EN, DE, FR, ES, IT, PT, ZH, JA, KO, AR, HE)
 //   • CJK character bigram extraction — handles Chinese, Japanese, Korean
+//   • RTL language support (Arabic, Hebrew) — word tokenization + light stemming
+//   • Romanization matching — katakana → romaji tokens for romaji queries
+//   • Cross-language retrieval — tech term synonyms EN↔JA↔ZH↔KO↔AR↔HE
 //   • Pipeline Redis reads for performance
 //
 /**
@@ -386,14 +389,330 @@ const STOPWORDS = new Set([
     '이', '가', '은', '는', '을', '를', '의', '에', '와', '과', '도',
     '로', '에서', '한', '하', '있', '없', '것', '수', '않', '들',
     '이다', '하다', '이고', '이며', '라고', '에게', '에서', '으로',
+    // ── Arabic — high-frequency function words & particles ──
+    'في', 'من', 'إلى', 'على', 'عن', 'مع', 'هذا', 'هذه', 'ذلك', 'تلك',
+    'الذي', 'التي', 'الذين', 'اللواتي', 'هو', 'هي', 'هم', 'هن', 'نحن',
+    'أنت', 'أنتم', 'أنا', 'كان', 'كانت', 'كانوا', 'يكون', 'تكون',
+    'أن', 'إن', 'لأن', 'لكن', 'أو', 'بل', 'ثم', 'حتى', 'إذا', 'كي',
+    'قد', 'لقد', 'لم', 'لن', 'ما', 'لا', 'ليس', 'غير', 'بعض', 'كل',
+    'جميع', 'أي', 'كيف', 'متى', 'أين', 'لماذا', 'ماذا', 'هل', 'ال',
+    'و', 'ف', 'ب', 'ل', 'ك', 'يا', 'أم', 'إلا', 'عند', 'بين',
+    // ── Hebrew — common particles, pronouns, conjunctions ──
+    'של', 'את', 'אל', 'על', 'עם', 'לא', 'הוא', 'היא', 'הם', 'הן',
+    'אני', 'אתה', 'אנחנו', 'אתם', 'כי', 'אם', 'אבל', 'גם', 'כבר',
+    'רק', 'עוד', 'יש', 'אין', 'מה', 'זה', 'זאת', 'אלה', 'כל', 'כן',
+    'לו', 'לה', 'להם', 'בו', 'בה', 'בהם', 'שם', 'כך', 'כן', 'מי',
+    'אשר', 'אחרי', 'לפני', 'תחת', 'בין', 'מאז', 'עד', 'כמו', 'אז',
 ]);
 // CJK Unicode ranges used for bigram extraction
 const CJK_RE = /[\u3040-\u30ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uac00-\ud7af]/;
 const SEGMENT_RE = /([\u3040-\u30ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uac00-\ud7af]+|[^\u3040-\u30ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uac00-\ud7af]+)/g;
+// RTL Unicode ranges: Arabic (U+0600–U+06FF + extended), Hebrew (U+0590–U+05FF)
+const RTL_RE = /[\u0590-\u05ff\u0600-\u06ff\u0750-\u077f\u08a0-\u08ff\ufb1d-\ufb4f\ufb50-\ufdff\ufe70-\ufeff]/;
+// ── Katakana → Romaji conversion table (Hepburn system) ──────────────────────
+// Digraphs must be listed before single chars so they match first.
+const KATA_ROMAJI_TABLE = [
+    // Special combinations for loanwords
+    ['ファ', 'fa'], ['フィ', 'fi'], ['フェ', 'fe'], ['フォ', 'fo'],
+    ['ティ', 'ti'], ['ディ', 'di'], ['トゥ', 'tu'], ['ドゥ', 'du'],
+    ['ウィ', 'wi'], ['ウェ', 'we'], ['ウォ', 'wo'],
+    ['チェ', 'che'], ['ジェ', 'je'], ['シェ', 'she'],
+    ['イェ', 'ye'], ['ヴァ', 'va'], ['ヴィ', 'vi'], ['ヴェ', 've'], ['ヴォ', 'vo'],
+    // Digraphs (2-char → romaji)
+    ['キャ', 'kya'], ['キュ', 'kyu'], ['キョ', 'kyo'],
+    ['シャ', 'sha'], ['シュ', 'shu'], ['ショ', 'sho'],
+    ['チャ', 'cha'], ['チュ', 'chu'], ['チョ', 'cho'],
+    ['ニャ', 'nya'], ['ニュ', 'nyu'], ['ニョ', 'nyo'],
+    ['ヒャ', 'hya'], ['ヒュ', 'hyu'], ['ヒョ', 'hyo'],
+    ['ミャ', 'mya'], ['ミュ', 'myu'], ['ミョ', 'myo'],
+    ['リャ', 'rya'], ['リュ', 'ryu'], ['リョ', 'ryo'],
+    ['ギャ', 'gya'], ['ギュ', 'gyu'], ['ギョ', 'gyo'],
+    ['ジャ', 'ja'], ['ジュ', 'ju'], ['ジョ', 'jo'],
+    ['ビャ', 'bya'], ['ビュ', 'byu'], ['ビョ', 'byo'],
+    ['ピャ', 'pya'], ['ピュ', 'pyu'], ['ピョ', 'pyo'],
+    // Single chars
+    ['ア', 'a'], ['イ', 'i'], ['ウ', 'u'], ['エ', 'e'], ['オ', 'o'],
+    ['カ', 'ka'], ['キ', 'ki'], ['ク', 'ku'], ['ケ', 'ke'], ['コ', 'ko'],
+    ['サ', 'sa'], ['シ', 'shi'], ['ス', 'su'], ['セ', 'se'], ['ソ', 'so'],
+    ['タ', 'ta'], ['チ', 'chi'], ['ツ', 'tsu'], ['テ', 'te'], ['ト', 'to'],
+    ['ナ', 'na'], ['ニ', 'ni'], ['ヌ', 'nu'], ['ネ', 'ne'], ['ノ', 'no'],
+    ['ハ', 'ha'], ['ヒ', 'hi'], ['フ', 'fu'], ['ヘ', 'he'], ['ホ', 'ho'],
+    ['マ', 'ma'], ['ミ', 'mi'], ['ム', 'mu'], ['メ', 'me'], ['モ', 'mo'],
+    ['ヤ', 'ya'], ['ユ', 'yu'], ['ヨ', 'yo'],
+    ['ラ', 'ra'], ['リ', 'ri'], ['ル', 'ru'], ['レ', 're'], ['ロ', 'ro'],
+    ['ワ', 'wa'], ['ヲ', 'wo'], ['ン', 'n'],
+    // Voiced
+    ['ガ', 'ga'], ['ギ', 'gi'], ['グ', 'gu'], ['ゲ', 'ge'], ['ゴ', 'go'],
+    ['ザ', 'za'], ['ジ', 'ji'], ['ズ', 'zu'], ['ゼ', 'ze'], ['ゾ', 'zo'],
+    ['ダ', 'da'], ['ヂ', 'ji'], ['ヅ', 'zu'], ['デ', 'de'], ['ド', 'do'],
+    ['バ', 'ba'], ['ビ', 'bi'], ['ブ', 'bu'], ['ベ', 'be'], ['ボ', 'bo'],
+    ['パ', 'pa'], ['ピ', 'pi'], ['プ', 'pu'], ['ペ', 'pe'], ['ポ', 'po'],
+    ['ヴ', 'v'],
+    // Long vowel / small chars
+    ['ー', ''], ['ァ', 'a'], ['ィ', 'i'], ['ゥ', 'u'], ['ェ', 'e'], ['ォ', 'o'],
+    ['ッ', ''], // handled separately (doubles next consonant)
+];
+const KATA_MAP = new Map(KATA_ROMAJI_TABLE);
+/**
+ * Convert a katakana string to Hepburn romaji.
+ * Handles digraphs, geminate consonants (ッ), and long vowel marks (ー).
+ */
+function katakanaToRomaji(kata) {
+    let result = '';
+    let i = 0;
+    while (i < kata.length) {
+        // Geminate consonant: ッ doubles the following consonant
+        if (kata[i] === 'ッ' && i + 1 < kata.length) {
+            const next = KATA_MAP.get(kata[i + 1]) ?? KATA_MAP.get(kata[i + 1] + kata[i + 2]) ?? '';
+            if (next.length > 0)
+                result += next[0]; // double first consonant
+            i++;
+            continue;
+        }
+        // Try 2-char digraph first
+        if (i + 1 < kata.length) {
+            const two = kata[i] + kata[i + 1];
+            const r2 = KATA_MAP.get(two);
+            if (r2 !== undefined) {
+                result += r2;
+                i += 2;
+                continue;
+            }
+        }
+        // Single char
+        const r1 = KATA_MAP.get(kata[i]);
+        if (r1 !== undefined)
+            result += r1;
+        i++;
+    }
+    return result;
+}
+/**
+ * Arabic light stemmer — strips definite article and common prefix particles.
+ * Handles: ال (al-), و (wa-), ب (bi-), ل (li-), ف (fa-), ك (ka-).
+ * Only strips when the result is still ≥ 3 chars to avoid over-stemming.
+ */
+function arabicLightStem(word) {
+    if (word.startsWith('ال') && word.length > 4)
+        return word.slice(2);
+    if (word.length > 3 && 'وبلفك'.includes(word[0]) && RTL_RE.test(word[1]))
+        return word.slice(1);
+    return word;
+}
+// ── Cross-lingual tech term dictionary ────────────────────────────────────────
+// Maps tokens (English, Japanese katakana, Chinese, Korean, Arabic, Hebrew)
+// to their equivalents in other languages. Used to expand tokens at search time
+// so "deploy" finds documents containing "デプロイ" / "部署" / "배포", and vice versa.
+//
+// Keys must be lowercase. CJK values will be bigram-expanded at use time.
+const CROSS_LINGUAL_MAP = new Map([
+    // English → all
+    ['deploy', ['デプロイ', '部署', '배포', 'deployment', 'deploying']],
+    ['deployment', ['デプロイ', '部署', '배포', 'deploy']],
+    ['container', ['コンテナ', '容器', '컨테이너', 'docker']],
+    ['server', ['サーバー', 'サーバ', '服务器', '서버', 'سيرفر', 'שרת']],
+    ['database', ['データベース', '数据库', '데이터베이스', 'db', 'قاعدة البيانات']],
+    ['cache', ['キャッシュ', '缓存', '캐시', 'كاش', 'מטמון']],
+    ['error', ['エラー', '错误', '오류', 'خطأ', 'שגיאה', 'exception', 'err']],
+    ['bug', ['バグ', '缺陷', '버그', 'issue', 'defect', 'خلل']],
+    ['fix', ['修正', '修复', '수정', 'bugfix', 'patch', 'hotfix', 'إصلاح', 'תיקון']],
+    ['build', ['ビルド', '构建', '빌드', 'بناء', 'בנייה']],
+    ['test', ['テスト', '测试', '테스트', 'اختبار', 'בדיקה']],
+    ['auth', ['認証', '认证', '인증', 'authentication', 'login', 'oauth', 'مصادقة', 'אימות']],
+    ['kubernetes', ['クベルネテス', 'k8s', 'kube']],
+    ['network', ['ネットワーク', '网络', '네트워크', 'شبكة', 'רשת']],
+    ['timeout', ['タイムアウト', '超时', '타임아웃', 'مهلة']],
+    ['memory', ['メモリ', '内存', '메모리', 'ram', 'ذاكرة', 'זיכרון']],
+    ['config', ['設定', '配置', '설정', 'configuration', 'settings', 'conf', 'إعدادات', 'הגדרות']],
+    ['install', ['インストール', '安装', '설치', 'تثبيت', 'התקנה']],
+    ['update', ['アップデート', '更新', '업데이트', 'upgrade', 'تحديث', 'עדכון']],
+    ['log', ['ログ', '日志', '로그', 'logging', 'سجل', 'יומן']],
+    ['port', ['ポート', '端口', '포트', 'منفذ', 'פורט']],
+    ['file', ['ファイル', '文件', '파일', 'ملف', 'קובץ']],
+    ['image', ['イメージ', '镜像', '이미지', 'صورة', 'תמונה']],
+    ['volume', ['ボリューム', '卷', '볼륨']],
+    ['cluster', ['クラスター', '集群', '클러스터']],
+    ['node', ['ノード', '节点', '노드']],
+    ['service', ['サービス', '服务', '서비스', 'svc', 'خدمة', 'שירות']],
+    ['certificate', ['証明書', '证书', '인증서', 'cert', 'ssl', 'tls', 'شهادة', 'תעודה']],
+    ['password', ['パスワード', '密码', '비밀번호', 'passwd', 'pwd', 'كلمة المرور', 'סיסמה']],
+    ['token', ['トークン', '令牌', '토큰', 'jwt', 'secret', 'رمز', 'אסימון']],
+    ['health', ['ヘルス', '健康', '헬스', 'healthcheck', 'probe', 'صحة']],
+    ['migration', ['マイグレーション', '迁移', '마이그레이션', 'migrate', 'هجرة']],
+    ['backup', ['バックアップ', '备份', '백업', 'نسخ احتياطي', 'גיבוי']],
+    ['monitor', ['モニター', '监控', '모니터링', 'monitoring', 'metrics', 'alert', 'مراقبة', 'ניטור']],
+    ['performance', ['パフォーマンス', '性能', '성능', 'latency', 'throughput']],
+    ['connection', ['接続', '连接', '연결', 'conn', 'socket', 'اتصال', 'חיבור']],
+    ['queue', ['キュー', '队列', '큐', 'طابور']],
+    ['redis', ['レディス', '레디스']],
+    ['nginx', ['エンジンエックス']],
+    ['linux', ['リナックス', 'لينكس']],
+    ['api', ['エーピーアイ', '接口', 'endpoint', 'واجهة برمجية', 'ממשק']],
+    ['healthcheck', ['ヘルスチェック', 'health', 'probe', 'فحص الصحة']],
+    ['loadbalancer', ['ロードバランサー', '负载均衡', '로드밸런서', 'lb']],
+    ['ssl', ['tls', 'https', 'certificate', 'cert']],
+    ['tls', ['ssl', 'https', 'certificate', 'cert']],
+    ['docker', ['container', 'コンテナ', '容器', '컨테이너', 'دوكر']],
+    ['git', ['version control', 'repo', 'repository', 'commit', 'push', 'pull']],
+    ['ci', ['pipeline', 'github actions', 'gitlab', 'jenkins', 'build']],
+    ['cd', ['deploy', 'deployment', 'release']],
+    ['debug', ['デバッグ', '调试', '디버그', 'debugging', 'breakpoint', 'تصحيح', 'איתור באגים']],
+    ['crash', ['クラッシュ', '崩溃', '크래시', 'panic', 'segfault', 'انهيار']],
+    ['restart', ['再起動', '重启', '재시작', 'reboot', 'إعادة تشغيل', 'הפעלה מחדש']],
+    ['permission', ['権限', '权限', '권한', 'access', 'acl', 'chmod', 'صلاحية', 'הרשאה']],
+    // Japanese katakana → English
+    ['デプロイ', ['deploy', 'deployment']],
+    ['コンテナ', ['container', 'docker']],
+    ['サーバー', ['server']],
+    ['サーバ', ['server']],
+    ['データベース', ['database', 'db']],
+    ['キャッシュ', ['cache']],
+    ['エラー', ['error', 'err']],
+    ['バグ', ['bug', 'issue']],
+    ['ビルド', ['build']],
+    ['テスト', ['test']],
+    ['認証', ['auth', 'authentication']],
+    ['設定', ['config', 'configuration']],
+    ['インストール', ['install']],
+    ['アップデート', ['update', 'upgrade']],
+    ['ログ', ['log', 'logs']],
+    ['ポート', ['port']],
+    ['ファイル', ['file']],
+    ['イメージ', ['image']],
+    ['クラスター', ['cluster']],
+    ['ノード', ['node']],
+    ['サービス', ['service']],
+    ['パスワード', ['password']],
+    ['トークン', ['token']],
+    ['バックアップ', ['backup']],
+    ['モニター', ['monitor', 'monitoring']],
+    ['接続', ['connection']],
+    ['キュー', ['queue']],
+    ['ヘルスチェック', ['healthcheck', 'health check', 'health']],
+    ['ロードバランサー', ['loadbalancer', 'load balancer']],
+    ['デバッグ', ['debug']],
+    ['クラッシュ', ['crash', 'panic']],
+    ['再起動', ['restart', 'reboot']],
+    ['権限', ['permission', 'access']],
+    // Chinese → English
+    ['部署', ['deploy', 'deployment']],
+    ['容器', ['container', 'docker']],
+    ['服务器', ['server']],
+    ['数据库', ['database']],
+    ['缓存', ['cache']],
+    ['错误', ['error']],
+    ['修复', ['fix', 'bugfix']],
+    ['构建', ['build']],
+    ['测试', ['test']],
+    ['认证', ['auth']],
+    ['配置', ['config']],
+    ['安装', ['install']],
+    ['更新', ['update']],
+    ['日志', ['log']],
+    ['端口', ['port']],
+    ['镜像', ['image']],
+    ['集群', ['cluster']],
+    ['节点', ['node']],
+    ['服务', ['service']],
+    ['密码', ['password']],
+    ['令牌', ['token']],
+    ['备份', ['backup']],
+    ['监控', ['monitor']],
+    ['连接', ['connection']],
+    ['队列', ['queue']],
+    ['性能', ['performance']],
+    ['内存', ['memory']],
+    ['调试', ['debug']],
+    ['崩溃', ['crash']],
+    ['重启', ['restart']],
+    ['权限', ['permission']],
+    // Korean → English
+    ['배포', ['deploy', 'deployment']],
+    ['컨테이너', ['container']],
+    ['서버', ['server']],
+    ['데이터베이스', ['database']],
+    ['캐시', ['cache']],
+    ['오류', ['error']],
+    ['수정', ['fix']],
+    ['빌드', ['build']],
+    ['테스트', ['test']],
+    ['인증', ['auth']],
+    ['설정', ['config']],
+    ['설치', ['install']],
+    ['업데이트', ['update']],
+    ['로그', ['log']],
+    ['포트', ['port']],
+    ['이미지', ['image']],
+    ['클러스터', ['cluster']],
+    ['노드', ['node']],
+    ['서비스', ['service']],
+    ['토큰', ['token']],
+    ['백업', ['backup']],
+    ['모니터링', ['monitor']],
+    ['연결', ['connection']],
+    ['큐', ['queue']],
+    ['성능', ['performance']],
+    ['메모리', ['memory']],
+    ['디버그', ['debug']],
+    ['크래시', ['crash']],
+    ['재시작', ['restart']],
+    ['권한', ['permission']],
+    // Arabic → English
+    ['خطأ', ['error']],
+    ['إصلاح', ['fix']],
+    ['بناء', ['build']],
+    ['اختبار', ['test']],
+    ['مصادقة', ['auth']],
+    ['إعدادات', ['config']],
+    ['تثبيت', ['install']],
+    ['تحديث', ['update']],
+    ['سجل', ['log']],
+    ['صورة', ['image']],
+    ['خدمة', ['service']],
+    ['نسخ احتياطي', ['backup']],
+    ['مراقبة', ['monitor']],
+    ['اتصال', ['connection']],
+    ['ذاكرة', ['memory']],
+    ['تصحيح', ['debug']],
+    ['شبكة', ['network']],
+    // Hebrew → English
+    ['שגיאה', ['error']],
+    ['תיקון', ['fix']],
+    ['בנייה', ['build']],
+    ['בדיקה', ['test']],
+    ['אימות', ['auth']],
+    ['הגדרות', ['config']],
+    ['התקנה', ['install']],
+    ['עדכון', ['update']],
+    ['יומן', ['log']],
+    ['שרת', ['server']],
+    ['שירות', ['service']],
+    ['גיבוי', ['backup']],
+    ['ניטור', ['monitor']],
+    ['חיבור', ['connection']],
+    ['זיכרון', ['memory']],
+    ['רשת', ['network']],
+    ['תמונה', ['image']],
+    ['סיסמה', ['password']],
+    ['הרשאה', ['permission']],
+]);
+/**
+ * Expand a token to its cross-lingual synonyms.
+ * Returns synonyms to be added to the token stream (pre-tokenized).
+ * CJK synonyms are NOT pre-bigram'd here — caller handles that.
+ */
+function expandCrossLingual(token) {
+    return CROSS_LINGUAL_MAP.get(token) ?? [];
+}
 /**
  * Tokenize text into meaningful keywords.
- * Handles accented Latin characters and CJK (Chinese, Japanese, Korean).
- * CJK text is split into character bigrams (the standard for indexing logographic languages).
+ * Handles accented Latin, RTL (Arabic, Hebrew), and CJK (Chinese, Japanese, Korean).
+ *
+ * Features:
+ *  • CJK → character bigrams
+ *  • Katakana → additionally emits Hepburn romaji tokens (enables romaji queries)
+ *  • Arabic/Hebrew → word tokenization with light Arabic stemming
+ *  • All scripts → cross-lingual synonym expansion (EN↔JA↔ZH↔KO↔AR↔HE)
  */
 function tokenize(text) {
     const tokens = [];
@@ -413,9 +732,37 @@ function tokenize(text) {
                 if (!STOPWORDS.has(last))
                     tokens.push(last);
             }
+            // Whole-segment cross-lingual lookup (for known tech terms like デプロイ, 部署, 배포)
+            // This ensures CJK full-words map to their English equivalents
+            const segSyns = expandCrossLingual(seg);
+            for (const syn of segSyns) {
+                if (!CJK_RE.test(syn)) {
+                    syn.split(/\s+/).filter(w => w.length > 1 && !STOPWORDS.has(w)).forEach(w => tokens.push(w));
+                }
+            }
+            // Romanization: extract katakana-only portion and emit romaji tokens.
+            // This allows users to search "kontena" or "depuroi" to find katakana docs.
+            const kataOnly = seg.replace(/[^\u30a0-\u30ff]/g, '');
+            if (kataOnly.length >= 2) {
+                const romaji = katakanaToRomaji(kataOnly);
+                if (romaji.length >= 2 && romaji.length <= 40 && !STOPWORDS.has(romaji)) {
+                    tokens.push(romaji);
+                }
+            }
+        }
+        else if (RTL_RE.test(seg)) {
+            // RTL segment (Arabic / Hebrew) — space-separated words with light stemming
+            const words = seg.trim().split(/\s+/);
+            for (const w of words) {
+                if (w.length <= 1 || STOPWORDS.has(w))
+                    continue;
+                const stemmed = arabicLightStem(w);
+                if (stemmed.length > 1 && !STOPWORDS.has(stemmed))
+                    tokens.push(stemmed);
+            }
         }
         else {
-            // Latin / other — existing logic
+            // Latin / other — existing logic preserved exactly
             const words = seg
                 .replace(/[^a-záàâãäåæçéèêëíìîïñóòôõöúùûüýÿßœ0-9:_\-./]/g, ' ')
                 .split(/\s+/)
@@ -423,6 +770,44 @@ function tokenize(text) {
             tokens.push(...words);
         }
     }
+    // ── Cross-lingual expansion ──────────────────────────────────────────────
+    // For each token, look up synonyms in other languages and add them.
+    // CJK synonyms are bigram-expanded inline; Latin/RTL synonyms added as-is.
+    const expansions = [];
+    const seen = new Set(tokens);
+    for (const tok of tokens) {
+        const syns = expandCrossLingual(tok);
+        for (const syn of syns) {
+            if (seen.has(syn))
+                continue;
+            seen.add(syn);
+            if (CJK_RE.test(syn)) {
+                // CJK synonym → expand to character bigrams
+                for (let i = 0; i < syn.length - 1; i++) {
+                    const bg = syn[i] + syn[i + 1];
+                    if (!STOPWORDS.has(syn[i]) && !STOPWORDS.has(bg))
+                        expansions.push(bg);
+                }
+                if (syn.length >= 1) {
+                    const last = syn[syn.length - 1];
+                    if (!STOPWORDS.has(last))
+                        expansions.push(last);
+                }
+            }
+            else if (RTL_RE.test(syn)) {
+                // RTL synonym
+                const stemmed = arabicLightStem(syn);
+                if (stemmed.length > 1 && !STOPWORDS.has(stemmed))
+                    expansions.push(stemmed);
+            }
+            else {
+                // Latin synonym — split in case of multi-word (e.g. "load balancer")
+                const parts = syn.split(/\s+/).filter(w => w.length > 1 && !STOPWORDS.has(w));
+                expansions.push(...parts);
+            }
+        }
+    }
+    tokens.push(...expansions);
     return tokens;
 }
 /**
@@ -696,7 +1081,7 @@ async function keywordSearch(redis, patterns, query, topK = 10) {
     return sorted.slice(0, topK);
 }
 // ── Exported for testing ──────────────────────────────────────────────────────
-export { tokenize, splitMultiQuery, levenshtein, recencyBoost, extractTimestamp, STOPWORDS };
+export { tokenize, splitMultiQuery, levenshtein, recencyBoost, extractTimestamp, STOPWORDS, katakanaToRomaji, arabicLightStem, expandCrossLingual, CROSS_LINGUAL_MAP };
 // ── Connection pool ───────────────────────────────────────────────────────────
 /** Reuse Redis connections across tool calls (keyed by instance_id). */
 const pool = new Map();