feat: AI 서비스 MLX 듀얼 백엔드 및 모델 최적화

- MLX(맥미니 27B) 우선 → Ollama(조립컴 9B) fallback 구조
- pydantic-settings 기반 config 전환
- health check에 MLX 상태 추가
- 텍스트 모델 qwen3:8b → qwen3.5:9b-q8_0 변경

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

ai-service/services/embedding_service.py

@@ -37,26 +37,46 @@ def build_metadata(issue: dict) -> dict:
     return meta
 
 
-async def sync_all_issues() -> dict:
-    issues = get_all_issues()
+BATCH_SIZE = 10
+
+
+async def _sync_issues_batch(issues: list[dict]) -> tuple[int, int]:
+    """배치 단위로 임베딩 생성 후 벡터 스토어에 저장"""
     synced = 0
     skipped = 0
+
+    # 유효한 이슈와 텍스트 준비
+    valid = []
     for issue in issues:
         doc_text = build_document_text(issue)
         if not doc_text.strip():
             skipped += 1
             continue
+        valid.append((issue, doc_text))
+
+    # 배치 단위로 임베딩 생성
+    for i in range(0, len(valid), BATCH_SIZE):
+        batch = valid[i:i + BATCH_SIZE]
+        texts = [doc_text for _, doc_text in batch]
         try:
-            embedding = await ollama_client.generate_embedding(doc_text)
-            vector_store.upsert(
-                doc_id=f"issue_{issue['id']}",
-                document=doc_text,
-                embedding=embedding,
-                metadata=build_metadata(issue),
-            )
-            synced += 1
-        except Exception as e:
-            skipped += 1
+            embeddings = await ollama_client.batch_embeddings(texts)
+            for (issue, doc_text), embedding in zip(batch, embeddings):
+                vector_store.upsert(
+                    doc_id=f"issue_{issue['id']}",
+                    document=doc_text,
+                    embedding=embedding,
+                    metadata=build_metadata(issue),
+                )
+                synced += 1
+        except Exception:
+            skipped += len(batch)
+
+    return synced, skipped
+
+
+async def sync_all_issues() -> dict:
+    issues = get_all_issues()
+    synced, skipped = await _sync_issues_batch(issues)
     if issues:
         max_id = max(i["id"] for i in issues)
         metadata_store.set_last_synced_id(max_id)
@@ -83,26 +103,11 @@ async def sync_single_issue(issue_id: int) -> dict:
 async def sync_incremental() -> dict:
     last_id = metadata_store.get_last_synced_id()
     issues = get_issues_since(last_id)
-    synced = 0
-    for issue in issues:
-        doc_text = build_document_text(issue)
-        if not doc_text.strip():
-            continue
-        try:
-            embedding = await ollama_client.generate_embedding(doc_text)
-            vector_store.upsert(
-                doc_id=f"issue_{issue['id']}",
-                document=doc_text,
-                embedding=embedding,
-                metadata=build_metadata(issue),
-            )
-            synced += 1
-        except Exception:
-            pass
+    synced, skipped = await _sync_issues_batch(issues)
     if issues:
         max_id = max(i["id"] for i in issues)
         metadata_store.set_last_synced_id(max_id)
-    return {"synced": synced, "new_issues": len(issues)}
+    return {"synced": synced, "skipped": skipped, "new_issues": len(issues)}
 
 
 async def search_similar_by_id(issue_id: int, n_results: int = 5) -> list[dict]: