fix(auto-repair): preserve exact playbook candidates

apps/api/src/services/playbook_service.py

@@ -14,6 +14,7 @@ Phase 3 ADR-030: RAG 向量搜尋整合
 - 封裝所有業務邏輯
 """
 
+import re as _re
 from typing import Protocol
 
 import structlog
@@ -32,13 +33,11 @@ from src.models.playbook import (
 )
 from src.repositories.interfaces import IPlaybookRepository
 from src.repositories.playbook_repository import get_playbook_repository
-from src.services.playbook_rag import get_playbook_rag_service
+from src.services.playbook_rag import PlaybookMatch, get_playbook_rag_service
 from src.utils.timezone import now_taipei
 
 logger = structlog.get_logger(__name__)
 
-import re as _re
-
 
 def _parse_ssh_command(ssh_cmd: str) -> tuple[str, str]:
     """
@@ -275,16 +274,16 @@ class PlaybookService:
 
         payload = KMWritePayload(
             path_type="playbook_extract",
-            entry_create_kwargs=dict(
-                title=f"[Playbook] {playbook.name}",
-                content=body,
-                entry_type=EntryType.INCIDENT_CASE,
-                category="auto_repair",
-                tags=[*playbook.tags, "playbook", "auto_extracted", playbook.status.value],
-                source=EntrySource.AI_EXTRACTED,
-                related_incident_id=incident.incident_id,
-                created_by="playbook_service",
-            ),
+            entry_create_kwargs={
+                "title": f"[Playbook] {playbook.name}",
+                "content": body,
+                "entry_type": EntryType.INCIDENT_CASE,
+                "category": "auto_repair",
+                "tags": [*playbook.tags, "playbook", "auto_extracted", playbook.status.value],
+                "source": EntrySource.AI_EXTRACTED,
+                "related_incident_id": incident.incident_id,
+                "created_by": "playbook_service",
+            },
             incident_id=incident.incident_id,
         )
         result = await km_write_with_flag(payload)
@@ -348,6 +347,17 @@ class PlaybookService:
                     vector_weight=0.6,
                     jaccard_weight=0.4,
                 )
+                hybrid_by_id = {match.playbook_id: match for match in hybrid_matches}
+                for playbook_id, jaccard_score in jaccard_results:
+                    if playbook_id in hybrid_by_id:
+                        continue
+                    hybrid_matches.append(
+                        PlaybookMatch(
+                            playbook_id=playbook_id,
+                            similarity_score=jaccard_score,
+                            match_type="jaccard",
+                        )
+                    )
 
                 # 補充 playbook_map (RAG 可能找到 Jaccard 沒找到的)
                 for match in hybrid_matches:
@@ -404,9 +414,9 @@ class PlaybookService:
                 )
             )
 
-        # Step 4: 按綜合分數排序 (similarity * success_rate)
+        # Step 4: 先保住 exact signal，避免精準 Playbook 被語意近似項擠掉。
         recommendations.sort(
-            key=lambda r: r.similarity_score * (0.5 + 0.5 * r.playbook.success_rate),
+            key=lambda r: self._recommendation_priority(r, symptoms),
             reverse=True,
         )
 
@@ -821,6 +831,25 @@ class PlaybookService:
 
         return matched
 
+    @staticmethod
+    def _normalized_overlap(left: list[str], right: list[str]) -> bool:
+        left_values = {value.casefold() for value in left if value}
+        right_values = {value.casefold() for value in right if value}
+        return bool(left_values & right_values)
+
+    def _recommendation_priority(
+        self,
+        recommendation: PlaybookRecommendation,
+        symptoms: SymptomPattern,
+    ) -> tuple[bool, bool, float]:
+        pattern = recommendation.playbook.symptom_pattern
+        alert_exact = self._normalized_overlap(symptoms.alert_names, pattern.alert_names)
+        service_exact = self._normalized_overlap(symptoms.affected_services, pattern.affected_services)
+        quality_score = recommendation.similarity_score * (
+            0.5 + 0.5 * recommendation.playbook.success_rate
+        )
+        return (alert_exact, service_exact, quality_score)
+
     def _generate_recommendation_reason(
         self,
         playbook: Playbook,

apps/api/tests/test_playbook_service.py

@@ -25,6 +25,7 @@ from src.models.playbook import (
     RiskLevel,
     SymptomPattern,
 )
+from src.services.playbook_rag import PlaybookMatch
 from src.services.playbook_service import PlaybookService
 from src.utils.timezone import now_taipei
 
@@ -282,6 +283,92 @@ class TestPlaybookService:
         # Should be empty or have very low similarity
         assert len(recommendations) == 0 or recommendations[0].similarity_score < 0.4
 
+    @pytest.mark.asyncio
+    async def test_get_recommendations_prioritizes_exact_alert(self, service, mock_repo):
+        """Exact alert matches should win over fuzzy service matches."""
+        exact_alert = create_test_playbook(
+            playbook_id="PB-EXACT-ALERT",
+            success_count=0,
+            failure_count=0,
+        )
+        exact_alert.symptom_pattern.alert_names = ["AwoooPT16E"]
+        exact_alert.symptom_pattern.affected_services = ["different-service"]
+        competing_service = create_test_playbook(
+            playbook_id="PB-FUZZY-SERVICE",
+            success_count=20,
+            failure_count=0,
+        )
+        competing_service.symptom_pattern.alert_names = ["SentryDown"]
+        competing_service.symptom_pattern.affected_services = [
+            "awoooi-auto-repair-canary-livefire"
+        ]
+        await mock_repo.create(exact_alert)
+        await mock_repo.create(competing_service)
+
+        symptoms = SymptomPattern(
+            alert_names=["AwoooPT16E"],
+            affected_services=["awoooi-auto-repair-canary-livefire"],
+            severity_range=["P2"],
+        )
+
+        recommendations = await service.get_recommendations(
+            symptoms,
+            top_k=1,
+            use_rag=False,
+        )
+
+        assert recommendations[0].playbook.playbook_id == "PB-EXACT-ALERT"
+
+    @pytest.mark.asyncio
+    async def test_get_recommendations_preserves_jaccard_candidates(
+        self,
+        service,
+        mock_repo,
+    ):
+        """RAG hybrid top-k must not drop exact Jaccard candidates."""
+        exact_alert = create_test_playbook(
+            playbook_id="PB-EXACT-JACCARD",
+            success_count=0,
+            failure_count=0,
+        )
+        exact_alert.symptom_pattern.alert_names = ["AwoooPT16F"]
+        exact_alert.symptom_pattern.affected_services = ["different-service"]
+        competing_vector = create_test_playbook(
+            playbook_id="PB-VECTOR-ONLY",
+            success_count=20,
+            failure_count=0,
+        )
+        competing_vector.symptom_pattern.alert_names = ["SentryDown"]
+        competing_vector.symptom_pattern.affected_services = [
+            "awoooi-auto-repair-canary-livefire"
+        ]
+        await mock_repo.create(exact_alert)
+        await mock_repo.create(competing_vector)
+
+        class FakeRagService:
+            async def hybrid_search(self, **_kwargs):
+                return [
+                    PlaybookMatch(
+                        playbook_id="PB-VECTOR-ONLY",
+                        similarity_score=0.99,
+                        match_type="vector",
+                    )
+                ]
+
+        async def fake_rag_service():
+            return FakeRagService()
+
+        service._get_rag_service = fake_rag_service
+        symptoms = SymptomPattern(
+            alert_names=["AwoooPT16F"],
+            affected_services=["awoooi-auto-repair-canary-livefire"],
+            severity_range=["P2"],
+        )
+
+        recommendations = await service.get_recommendations(symptoms, top_k=1)
+
+        assert recommendations[0].playbook.playbook_id == "PB-EXACT-JACCARD"
+
     @pytest.mark.asyncio
     async def test_approve_playbook(self, service, mock_repo):
         """Test approving a draft playbook"""

scripts/ops/awooop-seed-auto-repair-canary-playbook.py

@@ -27,7 +27,7 @@ for _api_root in (_IMAGE_API_ROOT, _REPO_API_ROOT):
         sys.path.insert(0, str(_api_root))
         break
 
-from src.models.playbook import (
+from src.models.playbook import (  # noqa: E402
     ActionType,
     Playbook,
     PlaybookSource,
@@ -36,9 +36,9 @@ from src.models.playbook import (
     RiskLevel,
     SymptomPattern,
 )
-from src.core.redis_client import close_redis_pool, init_redis_pool
-from src.repositories.playbook_repository import get_playbook_repository
-from src.utils.timezone import now_taipei
+from src.core.redis_client import close_redis_pool, init_redis_pool  # noqa: E402
+from src.repositories.playbook_repository import get_playbook_repository  # noqa: E402
+from src.utils.timezone import now_taipei  # noqa: E402
 
 
 DEFAULT_ALERTNAME = "AwoooPAutoRepairCanaryT16"
@@ -63,10 +63,11 @@ class SeedResult:
 def _playbook_id_for_alertname(alertname: str) -> str:
     if alertname == DEFAULT_ALERTNAME:
         return "PB-AWOOOP-T16-CANARY"
+    prefix = "PB-AWOOOP-CANARY-"
     suffix = re.sub(r"[^A-Z0-9]+", "-", alertname.upper()).strip("-")
     suffix = suffix.replace("AWOOOP-AUTO-REPAIR-CANARY-", "")
-    suffix = suffix[:18] or "T16"
-    return f"PB-AWOOOP-CANARY-{suffix}"
+    suffix = suffix[: 32 - len(prefix)] or "T16"
+    return f"{prefix}{suffix}"
 
 
 async def seed_canary_playbook(