fix(kb): 修复 metadata 覆写数据库图谱配置，改用文件级持久化

- update_kb_config 不再触发全量 _save_metadata，避免内存 metadata 覆盖数据库已有图谱配置
- 补全文件大小时改用 _persist_file 逐文件写入
- delete_file_chunks_only 通过 chunk_repo 检查文件是否有图谱数据再决定是否清理

backend/package/yuxi/knowledge/base.py

@@ -1242,8 +1242,6 @@ class KnowledgeBase(ABC):
         if update_llm_model_spec:
             self.databases_meta[db_id]["llm_model_spec"] = llm_model_spec
 
-        asyncio.create_task(self._save_metadata())
-
         return self.get_database_info(db_id)
 
     def get_retrievers(self) -> dict[str, dict]:
@@ -1394,7 +1392,9 @@ class KnowledgeBase(ABC):
 
         if updated:
             logger.info(f"Filled {updated}/{len(files_to_update)} missing file sizes from MinIO for {self.kb_type}")
-            await self._save_metadata()
+            for file_id, file_size in results:
+                if file_size is not None:
+                    await self._persist_file(file_id)
 
     async def _save_metadata(self) -> None:
         from yuxi.repositories.evaluation_repository import EvaluationRepository
@@ -1421,19 +1421,6 @@ class KnowledgeBase(ABC):
             }
             if existing is None:
                 await kb_repo.create(payload)
-            else:
-                await kb_repo.update(
-                    db_id,
-                    {
-                        "name": payload["name"],
-                        "description": payload["description"],
-                        "kb_type": payload["kb_type"],
-                        "embedding_model_spec": payload["embedding_model_spec"],
-                        "llm_model_spec": payload["llm_model_spec"],
-                        "query_params": payload["query_params"],
-                        "additional_params": payload["additional_params"],
-                    },
-                )
 
         for file_id, meta in self.files_meta.items():
             db_id = meta.get("database_id")

backend/package/yuxi/knowledge/implementations/milvus.py

@@ -633,7 +633,7 @@ class MilvusKB(KnowledgeBase):
                 request_params=params,
             )
             self.files_meta[file_id]["processing_params"] = params
-            await self._save_metadata()
+            await self._persist_file(file_id)
             logger.debug(f"[index_file] file_id={file_id}, processing_params={params}")
 
         # Add to processing queue
@@ -1189,15 +1189,15 @@ class MilvusKB(KnowledgeBase):
 
     async def delete_file_chunks_only(self, db_id: str, file_id: str) -> None:
         """仅删除文件的chunks数据，保留元数据（用于更新操作）"""
-        graph_config = (self.databases_meta.get(db_id, {}).get("metadata") or {}).get("graph_build_config")
-        if graph_config:
+        chunk_repo = KnowledgeChunkRepository()
+        if await chunk_repo.count_graph_indexed_by_file_id(file_id):
             from yuxi.knowledge.graphs.milvus_graph_service import MilvusGraphService
 
             try:
                 await MilvusGraphService().delete_file_graph(db_id, file_id)
             except Exception as e:
                 logger.error(f"Failed to delete graph data for file {file_id}: {e}")
-        await KnowledgeChunkRepository().delete_by_file_id(file_id)
+        await chunk_repo.delete_by_file_id(file_id)
         collection = await self._get_milvus_collection(db_id)
 
         if collection:

backend/package/yuxi/repositories/knowledge_chunk_repository.py

@@ -97,6 +97,15 @@ class KnowledgeChunkRepository:
     async def count_graph_indexed_by_db_id(self, db_id: str) -> int:
         return await self._count_by_db_id(db_id, KnowledgeChunk.graph_indexed.is_(True))
 
+    async def count_graph_indexed_by_file_id(self, file_id: str) -> int:
+        async with pg_manager.get_async_session_context() as session:
+            result = await session.execute(
+                select(func.count())
+                .select_from(KnowledgeChunk)
+                .where(KnowledgeChunk.file_id == file_id, KnowledgeChunk.graph_indexed.is_(True))
+            )
+            return int(result.scalar() or 0)
+
     async def count_graph_pending_by_db_id(self, db_id: str) -> int:
         return await self._count_by_db_id(db_id, KnowledgeChunk.graph_indexed.is_not(True))