diff --git a/landing/app/mira/search/page.tsx b/landing/app/mira/search/page.tsx
index 6a272b0..7154f3b 100644
--- a/landing/app/mira/search/page.tsx
+++ b/landing/app/mira/search/page.tsx
@@ -2,192 +2,136 @@
 
 export const runtime = 'edge';
 
-// Mira 搜尋頁 — Karpathy index pattern 為 primary（leo 2026-05-23）
-// SDD: polaris/mira/.agents/specs/mira-app/design.md §3.5.12.4「Karpathy index pattern（不用 vector embedding）」
-// 三層（C 混合）：
-//   1. Index 即時文字比對：掃 index-entry（entity 名 + 摘要）子字串命中 → 列 entity（零 token）
-//   2. LLM 路由（選用）：整個 index 餵 Claude，問「leo 想找哪些 entity」→ 最貼 Karpathy 本意
-//   3. 向量兜底（折疊）：KBDB /search semantic，SDD 明文「不是 primary，當保險」
+// Mira 搜尋頁 — 大眾化語義入口（leo 2026-07-05 拍板）
+//   default = 語義搜尋（普通用戶記大概意思；語義 normalize：黃仁勳＝皮衣男＝Jensen Huang 歸一）
+//   進階     = 關鍵字（懂的人用）
+// 後端：self-hosted leo21c KBDB（新 ingest 產 triplet/gloss entry，非舊 index-entry blocks）。
+//   過渡期直連 kbdb raw worker（cypher-proxy /kbdb/entries 尚未補，見 mira §1.7）。
+//   舊後端 kbdb.finally.click 已移除（mira 鐵律：禁直打舊 SaaS KBDB 當主儲存）。
+// graph 遍歷之後再補（本次不做）。
 
-import { Suspense, useCallback, useEffect, useMemo, useState } from 'react';
+import { Suspense, useCallback, useEffect, useState } from 'react';
 import { useRouter, useSearchParams } from 'next/navigation';
 import Link from 'next/link';
-import { MarkdownView } from '../_shared/markdown';
 import '../mira.css';
 
-const KBDB_BASE = 'https://kbdb.finally.click';
+// 新 KBDB（self-hosted leo21c）。過渡期直連 raw worker（無 auth、單租戶），
+// 不帶 owner_id（Vectorize owner metadata index 未建，帶了會回 0；總管另修）。
+const KBDB_BASE = 'https://arcrun-kbdb.leo21c.workers.dev';
+// 登入 gate 仍走官方 cypher（session cookie 綁 arcrun.dev；leo21c 對應 URL 未定 → 保留原樣，見回報）。
 const API_BASE = process.env.NEXT_PUBLIC_API_BASE ?? 'https://cypher.arcrun.dev';
-const CLAUDE_API = 'https://claude-api.arcrun.dev';
+
+type SearchMode = 'semantic' | 'keyword';
+
+// KBDB /entries/search 回傳的 entry（gloss：content = "實體：描述"）
+type SearchEntry = {
+  id: string;
+  content: string;
+  entry_type?: string;
+  owner_id?: string;
+  score?: number;
+};
+
+type SearchResponse = {
+  entries?: SearchEntry[];
+  mode?: string;             // 實際採用的 mode（semantic 未開會誠實降級 keyword）
+  capability_hint?: string;  // 降級時的提示
+};
 
 // 繁體異體字正規化（臺→台），讓 query 對得上多用「台」的 KB 內容
 function normalizeQuery(q: string): string {
   return q.replace(/臺/g, '台');
 }
 
-type IndexEntry = {
-  entity: string;       // H1 / page_name 去 index- 前綴
-  pageName: string;     // index-entry 自己的 page_name（index-{entity}）
-  oneLiner: string;     // 「一句話定義」
-  outline: string;      // facet outline 全文（拿來比對 + 餵 LLM）
-  raw: string;          // 完整 content（餵 LLM 用，截斷）
-};
-
-// 解析 index-entry markdown → 結構
-function parseIndexEntry(content: string, pageName: string): IndexEntry {
-  const entity = (content.match(/^#\s+(.+)$/m)?.[1] ?? pageName.replace(/^index-/, '')).trim();
-  const oneLiner = (content.match(/##\s*一句話定義\s*\n+([^\n#]+)/)?.[1] ?? '').trim();
-  const outlineMatch = content.match(/##\s*段落 outline[^\n]*\n([\s\S]*?)(?=\n##|$)/);
-  const outline = (outlineMatch?.[1] ?? '').trim();
-  return { entity, pageName, oneLiner, outline, raw: content.slice(0, 700) };
+// gloss content「實體：描述」→ { title, desc }（無冒號則整串當標題）
+function splitGloss(content: string): { title: string; desc: string } {
+  const m = content.match(/^\s*([^：:]+)[：:]\s*([\s\S]*)$/);
+  if (m) return { title: m[1].trim(), desc: m[2].trim() };
+  return { title: content.trim(), desc: '' };
 }
 
-// entity 名 → wiki page 路由（wiki-{entity}）
-function wikiHref(entity: string): string {
-  return `/mira/wiki/${encodeURIComponent('wiki-' + entity)}`;
-}
-
-// ── 向量兜底型別 ──
-type SearchMatch = {
-  score: number;
-  type: 'block' | 'triplet';
-  metadata?: { entity?: string;[k: string]: unknown };
-  block: { id: string; page_name: string | null; content: string | null; type: string; source: string | null } | null;
-  triplet: { id: string; subject?: string; predicate?: string; object?: string } | null;
-};
-
 function SearchInner() {
   const router = useRouter();
   const params = useSearchParams();
   const initialQ = params.get('q') ?? '';
 
-  const [apiKey, setApiKey] = useState<string | null>(null);
+  const [ready, setReady] = useState(false); // 登入檢查完成
   const [input, setInput] = useState(initialQ);
-  const [query, setQuery] = useState(initialQ.trim());
-  const [index, setIndex] = useState<IndexEntry[] | null>(null);
+  const [mode, setMode] = useState<SearchMode>('semantic');
+  const [results, setResults] = useState<SearchEntry[] | null>(null);
+  const [loading, setLoading] = useState(false);
   const [error, setError] = useState<string | null>(null);
+  const [hint, setHint] = useState<string | null>(null);
 
-  // LLM 路由結果
-  const [llmEntities, setLlmEntities] = useState<{ entity: string; reason: string }[] | null>(null);
-  const [llmLoading, setLlmLoading] = useState(false);
-
-  // 向量兜底
-  const [vecMatches, setVecMatches] = useState<SearchMatch[] | null>(null);
-  const [vecLoading, setVecLoading] = useState(false);
-  const [vecOpen, setVecOpen] = useState(false);
-
-  // 載入：me + 全部 index-entry
+  // 登入 gate（沿用官方 cypher /me；只確認登入，不需 api_key——raw worker 無 auth）
   useEffect(() => {
     (async () => {
       try {
         const meRes = await fetch(`${API_BASE}/me`, { credentials: 'include' });
         if (!meRes.ok) { window.location.href = '/login?redirect=/mira/search'; return; }
-        const me = (await meRes.json()) as { api_key: string };
-        setApiKey(me.api_key);
-        const r = await fetch(`${KBDB_BASE}/blocks?type=index-entry&limit=300`, {
-          headers: { Authorization: `Bearer ${me.api_key}` },
-        });
-        if (!r.ok) { setError(`index 讀取失敗：${r.status}`); return; }
-        const data = (await r.json()) as { blocks?: { content: string; page_name: string }[] };
-        setIndex((data.blocks ?? []).map(b => parseIndexEntry(b.content || '', b.page_name || '')));
+        setReady(true);
       } catch (e) {
         setError(e instanceof Error ? e.message : String(e));
       }
     })();
   }, []);
 
-  // 即時 index 文字比對（純 client，零 token）
-  const indexHits = useMemo(() => {
-    if (!index || !query) return [];
-    const q = normalizeQuery(query).toLowerCase();
-    const terms = q.split(/\s+/).filter(Boolean);
-    const scored = index
-      .map(e => {
-        const hay = normalizeQuery(`${e.entity}\n${e.oneLiner}\n${e.outline}`).toLowerCase();
-        let score = 0;
-        for (const t of terms) {
-          if (e.entity.toLowerCase().includes(t)) score += 10; // entity 名命中權重高
-          else if (hay.includes(t)) score += 3;
-        }
-        return { e, score };
-      })
-      .filter(x => x.score > 0)
-      .sort((a, b) => b.score - a.score)
-      .slice(0, 12);
-    return scored.map(x => x.e);
-  }, [index, query]);
+  // 語義/關鍵字搜尋：GET {KBDB_BASE}/entries/search?q=&mode=&limit=12（不帶 owner_id）
+  const runSearch = useCallback(async (rawQuery: string, m: SearchMode) => {
+    const q = normalizeQuery(rawQuery.trim());
+    if (!q) { setResults(null); setHint(null); return; }
+    setLoading(true);
+    setError(null);
+    setHint(null);
+    try {
+      const url = `${KBDB_BASE}/entries/search?q=${encodeURIComponent(q)}&mode=${m}&limit=12`;
+      const res = await fetch(url);
+      if (!res.ok) { setError(`搜尋失敗：${res.status}`); setResults([]); return; }
+      const data = (await res.json()) as SearchResponse;
+      const entries = (data.entries ?? [])
+        .slice()
+        .sort((a, b) => (b.score ?? 0) - (a.score ?? 0)); // 語義無門檻，照 score 排序顯示前 N
+      setResults(entries);
+      // 誠實降級提示：要了 semantic 卻回 keyword（Vectorize 未開）
+      if (m === 'semantic' && data.mode && data.mode !== 'semantic') {
+        setHint(data.capability_hint ?? '語義索引尚未啟用，暫以關鍵字比對代替。');
+      }
+    } catch (e) {
+      setError(e instanceof Error ? e.message : String(e));
+      setResults([]);
+    } finally {
+      setLoading(false);
+    }
+  }, []);
+
+  // 進站帶 ?q= 且登入完成 → 自動查一次
+  useEffect(() => {
+    if (ready && initialQ.trim()) runSearch(initialQ, mode);
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [ready]);
 
   const submit = (e: React.FormEvent) => {
     e.preventDefault();
     const q = input.trim();
-    setQuery(q);
-    setLlmEntities(null);
-    setVecMatches(null);
-    setVecOpen(false);
     router.replace(`/mira/search?q=${encodeURIComponent(q)}`);
+    runSearch(q, mode);
   };
 
-  // LLM 路由：整個 index 餵 Claude
-  const runLlmRoute = useCallback(async () => {
-    if (!index || !query || llmLoading) return;
-    setLlmLoading(true);
-    setLlmEntities(null);
-    try {
-      const indexDigest = index
-        .map(e => `- ${e.entity}：${e.oneLiner || '(無摘要)'}`)
-        .join('\n');
-      const prompt =
-        `你是 leo 知識庫的索引導航員。以下是所有 wiki entity 的索引（entity：一句話定義）：\n\n` +
-        `${indexDigest}\n\n---\n\n` +
-        `leo 想找：「${query}」\n\n` +
-        `請從上面索引挑出最相關的 entity（最多 6 個，可能 0 個）。` +
-        `只輸出 JSON 陣列，格式 [{"entity":"<完全照抄索引裡的名稱>","reason":"<為何相關，20字內>"}]，不要其他文字。`;
-      const res = await fetch(CLAUDE_API, {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ prompt, timeout_ms: 45000 }),
-      });
-      const data = (await res.json()) as { success?: boolean; data?: { text?: string } };
-      const text = data.data?.text ?? '';
-      const jsonMatch = text.match(/\[[\s\S]*\]/);
-      const parsed = jsonMatch ? JSON.parse(jsonMatch[0]) as { entity: string; reason: string }[] : [];
-      // 只留實際存在於 index 的 entity
-      const known = new Set(index.map(e => e.entity));
-      setLlmEntities(parsed.filter(p => known.has(p.entity)));
-    } catch {
-      setLlmEntities([]);
-    } finally {
-      setLlmLoading(false);
-    }
-  }, [index, query, llmLoading]);
-
-  // 向量兜底
-  const runVecSearch = useCallback(async () => {
-    if (!apiKey || !query || vecLoading) return;
-    setVecOpen(true);
-    setVecLoading(true);
-    try {
-      const res = await fetch(`${KBDB_BASE}/search`, {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${apiKey}` },
-        body: JSON.stringify({ query: normalizeQuery(query), type: 'semantic', topK: 12 }),
-      });
-      const data = (await res.json()) as { matches?: SearchMatch[] };
-      setVecMatches(data.matches ?? []);
-    } catch {
-      setVecMatches([]);
-    } finally {
-      setVecLoading(false);
-    }
-  }, [apiKey, query, vecLoading]);
+  // 切換 default 語義／進階 keyword；已有 query 就重查
+  const switchMode = (m: SearchMode) => {
+    if (m === mode) return;
+    setMode(m);
+    if (input.trim()) runSearch(input, m);
+  };
 
   return (
     <main className="mira-page">
       <div className="mira-content">
         <header style={{ padding: '24px 0 8px' }}>
           <Link href="/mira/feed" style={{ color: '#888', fontSize: 14, textDecoration: 'none' }}>← 河道</Link>
-          <h1 style={{ fontSize: 26, fontWeight: 700, color: '#fff', margin: '8px 0 0' }}>🔍 Wiki 搜尋</h1>
+          <h1 style={{ fontSize: 26, fontWeight: 700, color: '#fff', margin: '8px 0 0' }}>🔍 搜尋</h1>
           <p style={{ color: '#888', fontSize: 12, marginTop: 4 }}>
-            從 {index?.length ?? '…'} 個 wiki 主題的索引找（Karpathy index）
+            會打字就能查——{mode === 'semantic' ? '語義搜尋（記大概意思即可）' : '關鍵字精確比對'}
           </p>
         </header>
 
@@ -196,105 +140,61 @@ function SearchInner() {
             className="mira-search-input"
             value={input}
             onChange={e => setInput(e.target.value)}
-            placeholder="找主題（例：特化、台北大學、本地模型）"
+            placeholder={mode === 'semantic' ? '想找什麼？（例：那個穿皮衣的 AI 老闆、本地模型）' : '關鍵字（精確比對）'}
             autoFocus
           />
-          <button type="submit" className="mira-btn-primary" disabled={!index}>搜尋</button>
+          <button type="submit" className="mira-btn-primary" disabled={!ready}>搜尋</button>
         </form>
 
-        {error && <div className="mira-error" style={{ marginBottom: 10 }}>{error}</div>}
-        {!index && !error && <div className="empty-state">載入索引中…</div>}
+        {/* default 語義／進階 keyword 切換 */}
+        <div className="mira-search-mode" style={{ display: 'flex', gap: 8, marginTop: 10, fontSize: 13 }}>
+          <button
+            type="button"
+            onClick={() => switchMode('semantic')}
+            className={mode === 'semantic' ? 'mira-btn-primary' : 'mira-search-vec-toggle'}
+            style={{ padding: '4px 12px' }}
+          >語義（推薦）</button>
+          <button
+            type="button"
+            onClick={() => switchMode('keyword')}
+            className={mode === 'keyword' ? 'mira-btn-primary' : 'mira-search-vec-toggle'}
+            style={{ padding: '4px 12px' }}
+          >進階：關鍵字</button>
+        </div>
 
-        {index && query && (
-          <>
-            {/* 第 1 層：index 即時命中 */}
-            <section style={{ marginTop: 8 }}>
-              <div className="mira-search-section-head">📇 索引命中（{indexHits.length}）</div>
-              {indexHits.length === 0 ? (
-                <div className="empty-state" style={{ padding: '16px 0' }}>索引裡沒有直接命中的主題。</div>
-              ) : (
-                indexHits.map(e => (
-                  <Link key={e.pageName} href={wikiHref(e.entity)} className="mira-search-result">
-                    <div className="mira-search-result-meta"><span>📚 {e.entity}</span></div>
-                    {e.oneLiner && <div className="mira-search-snippet">{e.oneLiner}</div>}
-                  </Link>
-                ))
-              )}
-            </section>
+        {error && <div className="mira-error" style={{ marginTop: 12 }}>{error}</div>}
+        {hint && <div className="empty-state" style={{ padding: '10px 0', color: 'var(--mira-text-3)' }}>ℹ️ {hint}</div>}
 
-            {/* 第 2 層：LLM 路由（選用） */}
-            <section style={{ marginTop: 18 }}>
-              {!llmEntities && (
-                <button type="button" className="mira-search-llm-btn" onClick={runLlmRoute} disabled={llmLoading}>
-                  {llmLoading ? '🧠 Mira 翻索引中…' : '🧠 找不到？讓 Mira 讀整個索引幫你找'}
-                </button>
-              )}
-              {llmEntities && (
-                <>
-                  <div className="mira-search-section-head">🧠 Mira 從索引挑的（{llmEntities.length}）</div>
-                  {llmEntities.length === 0 ? (
-                    <div className="empty-state" style={{ padding: '12px 0' }}>Mira 也覺得索引裡沒有相關主題。</div>
-                  ) : (
-                    llmEntities.map(p => (
-                      <Link key={p.entity} href={wikiHref(p.entity)} className="mira-search-result">
-                        <div className="mira-search-result-meta"><span>📚 {p.entity}</span></div>
-                        <div className="mira-search-snippet" style={{ color: 'var(--mira-text-3)' }}>{p.reason}</div>
-                      </Link>
-                    ))
-                  )}
-                </>
-              )}
-            </section>
+        {loading && <div className="empty-state" style={{ padding: '16px 0' }}>搜尋中…</div>}
 
-            {/* 第 3 層：向量兜底（折疊） */}
-            <section style={{ marginTop: 18 }}>
-              {!vecOpen ? (
-                <button type="button" className="mira-search-vec-toggle" onClick={runVecSearch}>
-                  ▸ 也試試全文 / 語義搜尋（兜底）
-                </button>
-              ) : (
-                <>
-                  <div className="mira-search-section-head">🧬 語義兜底</div>
-                  {vecLoading && <div className="empty-state" style={{ padding: '12px 0' }}>搜尋中…</div>}
-                  {vecMatches && vecMatches.length === 0 && !vecLoading && (
-                    <div className="empty-state" style={{ padding: '12px 0' }}>沒有更多結果。</div>
-                  )}
-                  {vecMatches && vecMatches.map((m, i) => <VecResult key={i} match={m} />)}
-                </>
-              )}
-            </section>
-          </>
+        {!loading && results && (
+          <section style={{ marginTop: 14 }}>
+            <div className="mira-search-section-head">
+              {mode === 'semantic' ? '🧬 語義結果' : '📇 關鍵字結果'}（{results.length}）
+            </div>
+            {results.length === 0 ? (
+              <div className="empty-state" style={{ padding: '16px 0' }}>沒有找到相關結果，換個說法試試。</div>
+            ) : (
+              results.map(e => <ResultRow key={e.id} entry={e} />)
+            )}
+          </section>
         )}
       </div>
     </main>
   );
 }
 
-function VecResult({ match }: { match: SearchMatch }) {
-  const pct = Math.round((match.score ?? 0) * 100);
-  if (match.triplet) {
-    const { subject, predicate, object } = match.triplet;
-    return (
-      <div className="mira-search-result">
-        <div className="mira-search-result-meta"><span>關係</span>{pct > 0 && <span className="mira-search-score">{pct}%</span>}</div>
-        <div className="mira-search-snippet" style={{ fontFamily: 'monospace' }}>{subject} ﹥﹥ {predicate} ﹥﹥ {object}</div>
-      </div>
-    );
-  }
-  const b = match.block;
-  if (!b) return null;
-  const snippet = (b.content ?? '').replace(/\n+/g, ' ').slice(0, 200);
-  const href = b.type === 'wiki-page' && b.page_name
-    ? `/mira/wiki/${encodeURIComponent(b.page_name)}`
-    : b.page_name ? `/mira/feed#page=${encodeURIComponent(b.page_name)}` : `/mira/feed#raw=${encodeURIComponent(b.id)}`;
+function ResultRow({ entry }: { entry: SearchEntry }) {
+  const { title, desc } = splitGloss(entry.content || '');
+  const pct = entry.score != null ? Math.round(entry.score * 100) : null;
   return (
-    <Link href={href} className="mira-search-result">
+    <div className="mira-search-result">
       <div className="mira-search-result-meta">
-        <span>{b.type === 'wiki-page' ? '📚 Wiki' : '🌊 河道'}</span>
-        {pct > 0 && <span className="mira-search-score">{pct}%</span>}
+        <span>📚 {title}</span>
+        {pct != null && pct > 0 && <span className="mira-search-score">{pct}%</span>}
       </div>
-      <div className="mira-search-snippet"><MarkdownView text={snippet + ((b.content ?? '').length > 200 ? '…' : '')} /></div>
-    </Link>
+      {desc && <div className="mira-search-snippet">{desc}</div>}
+    </div>
   );
 }