diff --git a/services/hermes_analyst_service.py b/services/hermes_analyst_service.py
index 8977401..b4be5a7 100644
--- a/services/hermes_analyst_service.py
+++ b/services/hermes_analyst_service.py
@@ -124,6 +124,7 @@ class HermesAnalystService:
                 caller='hermes_analyst',
                 threshold=0.85,
                 request_id=rid,
+                mark_saved_call=True,
             )
             if rag.has_high_confidence:
                 logger.info(
diff --git a/services/openclaw_strategist_service.py b/services/openclaw_strategist_service.py
index 847c24c..f76a669 100644
--- a/services/openclaw_strategist_service.py
+++ b/services/openclaw_strategist_service.py
@@ -180,6 +180,7 @@ def generate_strategy_response(query: str, context: Optional[Dict[str, Any]] = N
             rag = rag_service.query(
                 text=q, caller='openclaw_qa',
                 threshold=0.85, request_id=request_id,
+                mark_saved_call=True,
             )
             if rag.has_high_confidence:
                 logger.info(
diff --git a/services/rag_service.py b/services/rag_service.py
index cb702a1..be1a608 100644
--- a/services/rag_service.py
+++ b/services/rag_service.py
@@ -301,6 +301,7 @@ class RAGService:
         threshold: float = RAG_DEFAULT_THRESHOLD,
         request_id: Optional[str] = None,
         insight_type: Optional[str] = None,
+        mark_saved_call: bool = False,
     ) -> RAGResult:
         """執行 RAG 召回。
 
@@ -311,6 +312,10 @@ class RAGService:
             threshold:    cosine similarity 門檻（0-1，預設 0.85）
             request_id:   與 ai_calls.request_id 串鏈
             insight_type: 限制 ai_insights.insight_type（None = 全類型）
+            mark_saved_call:
+                True 時，若本次結果 has_high_confidence，rag_query_log.saved_call
+                會寫 True。只給「命中後直接跳過 LLM」的 RAG-first caller 使用，
+                觀測台相似案例查詢等輔助用途應保留 False。
 
         Returns:
             RAGResult。失敗時 hits=[] + duration_ms 仍記錄。
@@ -381,9 +386,9 @@ class RAGService:
             embedding_signature=signature,
             hits=hits,
             threshold=threshold,
-            saved_call=False,  # caller 確認後再設（_call_after_decision API 預留）
             duration_ms=duration_ms,
         )
+        result.saved_call = bool(mark_saved_call and result.has_high_confidence)
 
         # ── 路徑 5：fire-and-forget rag_query_log ──
         self._async_log(
@@ -393,6 +398,7 @@ class RAGService:
             top_k=top_k,
             threshold=threshold,
             hits=hits,
+            saved_call=result.saved_call,
             request_id=request_id,
         )
 
@@ -548,7 +554,8 @@ class RAGService:
         top_k: int,
         threshold: float,
         hits: List[Dict[str, Any]],
-        request_id: Optional[str],
+        request_id: Optional[str] = None,
+        saved_call: bool = False,
     ) -> None:
         """放到 daemon thread 寫入 rag_query_log，主流程不阻塞。
 
@@ -563,7 +570,7 @@ class RAGService:
 
         threading.Thread(
             target=self._write_log,
-            args=(caller, text, query_vec, top_k, threshold, hits, request_id),
+            args=(caller, text, query_vec, top_k, threshold, hits, request_id, saved_call),
             name=f"rag-query-log-{caller}",
             daemon=True,
         ).start()
@@ -576,7 +583,8 @@ class RAGService:
         top_k: int,
         threshold: float,
         hits: List[Dict[str, Any]],
-        request_id: Optional[str],
+        request_id: Optional[str] = None,
+        saved_call: bool = False,
     ) -> None:
         """try/except 全包；DB 掛了只 log warning 不爆炸。"""
         try:
@@ -625,7 +633,7 @@ class RAGService:
                         'threshold':    round(float(threshold), 3),
                         'hit_count':    len(hits),
                         'used_results': used_results if used_results else None,
-                        'saved_call':   False,  # caller 確認後另寫；INSERT 階段固定 False
+                        'saved_call':   bool(saved_call),
                         'request_id':   (request_id or None),
                     },
                 )
diff --git a/tests/test_rag_service.py b/tests/test_rag_service.py
index dc70381..182af32 100644
--- a/tests/test_rag_service.py
+++ b/tests/test_rag_service.py
@@ -126,6 +126,44 @@ class TestRagEnabledHits:
         assert result.has_high_confidence is True
         assert '本週業績漲' in result.synthesize()
 
+    def test_mark_saved_call_only_when_requested_and_confident(self, rag_enabled, monkeypatch):
+        """RAG-first caller 明確標記時，高信心命中才計入 saved_call。"""
+        from services import rag_service as rs
+
+        monkeypatch.setattr(
+            'services.ollama_service.ollama_service.generate_embedding',
+            lambda text, model="bge-m3:latest": _fake_embedding(),
+        )
+        monkeypatch.setattr(
+            rs.rag_service,
+            '_select_hits',
+            lambda **_kw: [{'id': 101, 'content': '命中內容', 'score': 0.95}],
+        )
+        captured = {}
+
+        def _capture_async_log(**kwargs):
+            captured.update(kwargs)
+
+        monkeypatch.setattr(rs.rag_service, '_async_log', _capture_async_log)
+
+        result = rs.rag_service.query(
+            "本週業績趨勢",
+            caller='openclaw_qa',
+            mark_saved_call=True,
+        )
+
+        assert result.saved_call is True
+        assert captured['saved_call'] is True
+
+        result = rs.rag_service.query(
+            "本週業績趨勢",
+            caller='admin_quality_trend',
+            mark_saved_call=False,
+        )
+
+        assert result.saved_call is False
+        assert captured['saved_call'] is False
+
     def test_low_confidence_no_hit(self, rag_enabled, monkeypatch):
         """所有結果 distance>0.15 → similarity<0.85 → SQL 已過濾，回空 → has_high_confidence False。"""
         from services import rag_service as rs
@@ -292,10 +330,12 @@ class TestFireAndForgetLog:
             threshold=0.85,
             hits=[{'id': 101}],
             request_id='req-1',
+            saved_call=True,
         )
 
         assert "embedding_signature" in captured["sql"]
         assert captured["params"]["embedding_signature"] == rs.get_embedding_signature()
+        assert captured["params"]["saved_call"] is True
         fake_session.commit.assert_called_once()
 
     def test_embedding_failure_falls_back_to_empty(self, rag_enabled, monkeypatch):