diff --git a/src/lib/db/migrate.ts b/src/lib/db/migrate.ts
index bb69ead..243b7f7 100644
--- a/src/lib/db/migrate.ts
+++ b/src/lib/db/migrate.ts
@@ -331,4 +331,53 @@ export async function runMigrations(): Promise<void> {
     console.error("[migrate] Migration failed:", err);
     throw err;
   }
+
+  // ─── Semantic Cache (pgvector) ──────────────────────────────────────────────
+  // Graceful: if pgvector ext not available, log warning and skip — system still works
+  try {
+    await sql`CREATE EXTENSION IF NOT EXISTS vector`;
+    await sql`
+      CREATE TABLE IF NOT EXISTS semantic_cache (
+        id BIGSERIAL PRIMARY KEY,
+        query_hash TEXT UNIQUE NOT NULL,
+        query TEXT NOT NULL,
+        embedding vector(768),
+        response JSONB NOT NULL,
+        provider TEXT,
+        model TEXT,
+        hit_count INT DEFAULT 0,
+        created_at TIMESTAMPTZ DEFAULT now(),
+        last_used_at TIMESTAMPTZ DEFAULT now()
+      )
+    `;
+    await sql`
+      CREATE INDEX IF NOT EXISTS idx_semantic_cache_embedding
+      ON semantic_cache USING ivfflat (embedding vector_cosine_ops)
+      WITH (lists = 100)
+    `;
+    await sql`CREATE INDEX IF NOT EXISTS idx_semantic_cache_used ON semantic_cache(last_used_at DESC)`;
+    console.log("[migrate] pgvector + semantic_cache ready");
+  } catch (err) {
+    console.warn("[migrate] pgvector not available — semantic cache disabled:", String(err).slice(0, 150));
+  }
+
+  // ─── Performance indexes ─────────────────────────────────────────────────────
+  // Note: CREATE INDEX CONCURRENTLY cannot run inside transaction,
+  // so we use normal CREATE INDEX IF NOT EXISTS here
+  await sql`
+    CREATE INDEX IF NOT EXISTS idx_health_cooldown_active
+    ON health_logs(model_id, cooldown_until)
+    WHERE cooldown_until > now()
+  `.catch((err) => console.warn("[migrate] idx_health_cooldown_active:", String(err).slice(0, 100)));
+
+  await sql`
+    CREATE INDEX IF NOT EXISTS idx_gateway_recent
+    ON gateway_logs(created_at DESC, status)
+  `.catch((err) => console.warn("[migrate] idx_gateway_recent:", String(err).slice(0, 100)));
+
+  await sql`
+    CREATE INDEX IF NOT EXISTS idx_exam_passed_recent
+    ON exam_attempts(model_id, started_at DESC)
+    WHERE passed = true
+  `.catch((err) => console.warn("[migrate] idx_exam_passed_recent:", String(err).slice(0, 100)));
 }
diff --git a/src/lib/semantic-cache.ts b/src/lib/semantic-cache.ts
new file mode 100644
index 0000000..382da0b
--- /dev/null
+++ b/src/lib/semantic-cache.ts
@@ -0,0 +1,140 @@
+/**
+ * Semantic Cache — store responses indexed by message embedding
+ *
+ * Lookup: cosine similarity > threshold → return cached response
+ * Store:  embed query + response, persist to pgvector table
+ *
+ * Embedding source: Ollama local `nomic-embed-text` (free, no external API)
+ * Fallback: if Ollama or pgvector unavailable, operations no-op gracefully
+ */
+import { getSqlClient } from "@/lib/db/schema";
+import { createHash } from "crypto";
+
+const OLLAMA_BASE_URL = process.env.OLLAMA_BASE_URL || "http://localhost:11434";
+const EMBED_MODEL = "nomic-embed-text";
+const DEFAULT_THRESHOLD = 0.92;
+const EMBED_TIMEOUT_MS = 3_000;
+
+export interface SemanticHit {
+  query: string;
+  response: Record<string, unknown>;
+  provider: string | null;
+  model: string | null;
+  similarity: number;
+  hit_count: number;
+}
+
+function hashQuery(query: string): string {
+  return createHash("sha256").update(query).digest("hex").slice(0, 32);
+}
+
+async function embedQuery(query: string): Promise<number[] | null> {
+  try {
+    const res = await fetch(`${OLLAMA_BASE_URL}/api/embeddings`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ model: EMBED_MODEL, prompt: query }),
+      signal: AbortSignal.timeout(EMBED_TIMEOUT_MS),
+    });
+    if (!res.ok) return null;
+    const json = await res.json();
+    const embedding = json.embedding as number[] | undefined;
+    return Array.isArray(embedding) && embedding.length > 0 ? embedding : null;
+  } catch {
+    return null;
+  }
+}
+
+function vectorLiteral(vec: number[]): string {
+  return "[" + vec.join(",") + "]";
+}
+
+/**
+ * Look up cached response by semantic similarity.
+ * Returns null if no match or any error (graceful fallback).
+ */
+export async function getCachedBySimilarity(
+  query: string,
+  threshold: number = DEFAULT_THRESHOLD
+): Promise<SemanticHit | null> {
+  if (!query || query.length < 10) return null;
+  const embedding = await embedQuery(query);
+  if (!embedding) return null;
+  try {
+    const sql = getSqlClient();
+    const vecStr = vectorLiteral(embedding);
+    const rows = await sql<Array<{
+      query: string;
+      response: Record<string, unknown>;
+      provider: string | null;
+      model: string | null;
+      similarity: number;
+      hit_count: number;
+    }>>`
+      SELECT query, response, provider, model, hit_count,
+             1 - (embedding <=> ${vecStr}::vector) AS similarity
+      FROM semantic_cache
+      ORDER BY embedding <=> ${vecStr}::vector
+      LIMIT 1
+    `;
+    const row = rows[0];
+    if (!row || row.similarity < threshold) return null;
+    // Update hit count + last_used_at (fire and forget)
+    sql`
+      UPDATE semantic_cache
+      SET hit_count = hit_count + 1, last_used_at = now()
+      WHERE query = ${row.query}
+    `.catch(() => {});
+    return row;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Store response with embedding. Idempotent via query_hash.
+ */
+export async function storeSemanticResponse(
+  query: string,
+  response: Record<string, unknown>,
+  provider: string | null,
+  model: string | null
+): Promise<void> {
+  if (!query || query.length < 10) return;
+  const embedding = await embedQuery(query);
+  if (!embedding) return;
+  try {
+    const sql = getSqlClient();
+    const hash = hashQuery(query);
+    const vecStr = vectorLiteral(embedding);
+    await sql`
+      INSERT INTO semantic_cache (query_hash, query, embedding, response, provider, model)
+      VALUES (${hash}, ${query.slice(0, 2000)}, ${vecStr}::vector, ${JSON.stringify(response)}::jsonb, ${provider}, ${model})
+      ON CONFLICT (query_hash) DO UPDATE SET
+        last_used_at = now(),
+        hit_count = semantic_cache.hit_count + 1
+    `;
+  } catch {
+    // silent — cache is optional
+  }
+}
+
+export async function getSemanticCacheStats(): Promise<{
+  total: number;
+  hits: number;
+  avgHits: number;
+}> {
+  try {
+    const sql = getSqlClient();
+    const rows = await sql<{ total: number; hits: number; avg: number }[]>`
+      SELECT COUNT(*)::int as total,
+             COALESCE(SUM(hit_count), 0)::int as hits,
+             COALESCE(AVG(hit_count), 0)::float as avg
+      FROM semantic_cache
+    `;
+    const r = rows[0];
+    return { total: r?.total ?? 0, hits: r?.hits ?? 0, avgHits: r?.avg ?? 0 };
+  } catch {
+    return { total: 0, hits: 0, avgHits: 0 };
+  }
+}