diff --git a/src/backend/bisheng/database/models/message.py b/src/backend/bisheng/database/models/message.py
index 65c37aed5..bafe1f5c5 100644
--- a/src/backend/bisheng/database/models/message.py
+++ b/src/backend/bisheng/database/models/message.py
@@ -50,6 +50,12 @@ class ChatMessage(MessageBase, table=True):
     id: Optional[int] = Field(default=None, primary_key=True)
     receiver: Optional[Dict] = Field(default=None, sa_column=Column(JSON))
 
+    # 关键：设置表级别字符集为 utf8mb4
+    __table_args__ = {
+        "mysql_charset": "utf8mb4",
+        "mysql_collate": "utf8mb4_unicode_ci"
+    }
+
 
 class ChatMessageRead(MessageBase):
     id: Optional[int] = None
diff --git a/src/backend/bisheng/database/service.py b/src/backend/bisheng/database/service.py
index 251c7e5d1..1b4a0c164 100644
--- a/src/backend/bisheng/database/service.py
+++ b/src/backend/bisheng/database/service.py
@@ -29,17 +29,19 @@ class DatabaseService(Service):
 
     def _create_engine(self):
         connect_args = {'check_same_thread': False} if self.database_url.startswith("sqlite") else {}
+        connect_args['charset'] = "utf8mb4"  # Ensure utf8mb4 charset for MySQL
         return create_engine(
             self.database_url,
             connect_args=connect_args,
             pool_size=100,
             max_overflow=20,
             pool_timeout=3,
-            pool_pre_ping=True
+            pool_pre_ping=True,
         )
 
     def _create_async_engine(self):
         connect_args = {'check_same_thread': False} if self.async_database_url.startswith("sqlite") else {}
+        connect_args['charset'] = "utf8mb4"  # Ensure utf8mb4 charset for MySQL
         return create_async_engine(
             self.async_database_url,
             connect_args=connect_args,
diff --git a/src/backend/bisheng/worker/knowledge/rebuild_knowledge_worker.py b/src/backend/bisheng/worker/knowledge/rebuild_knowledge_worker.py
index 2a56f23c0..e46f12e33 100644
--- a/src/backend/bisheng/worker/knowledge/rebuild_knowledge_worker.py
+++ b/src/backend/bisheng/worker/knowledge/rebuild_knowledge_worker.py
@@ -86,6 +86,10 @@ def rebuild_knowledge_celery(knowledge_id: int, new_model_id: str) -> str:
 
             # 5. 更新knowledge状态
             if failed_files:
+
+                # 删除es索引和milvus集合，避免数据不一致
+                _delete_es_files(knowledge, failed_files)
+
                 knowledge.state = KnowledgeState.FAILED.value
                 logger.error(f"knowledge_id={knowledge_id} rebuild failed, failed_files={failed_files}")
             else:
@@ -110,24 +114,32 @@ def rebuild_knowledge_celery(knowledge_id: int, new_model_id: str) -> str:
             raise e
 
 
-def _delete_milvus_collection(knowledge: Knowledge):
-    """删除Milvus集合"""
+
+def _delete_es_files(knowledge: Knowledge, file_ids: List[int]):
+    """删除ES中的文件数据"""
     try:
+        index_name = knowledge.index_name or knowledge.collection_name
         embeddings = FakeEmbedding()
-        vector_client = decide_vectorstores(knowledge.collection_name, "Milvus", embeddings)
+        es_client = decide_vectorstores(index_name, "ElasticKeywordsSearch", embeddings)
 
-        if isinstance(vector_client.col, Collection):
-            logger.info(f"Deleting milvus collection: {knowledge.collection_name}")
-            vector_client.col.drop(timeout=10)
-            logger.info(f"Successfully deleted milvus collection: {knowledge.collection_name}")
-        else:
-            logger.warning(f"Milvus collection not found: {knowledge.collection_name}")
+        if not es_client.client.indices.exists(index=index_name):
+            logger.warning(f"ES index {index_name} does not exist, skipping deletion")
+            return
 
-        vector_client.close_connection(vector_client.alias)
+        for file_id in file_ids:
+            delete_query = {
+                "query": {
+                    "match": {
+                        "metadata.file_id": file_id
+                    }
+                }
+            }
+            response = es_client.client.delete_by_query(index=index_name, body=delete_query)
+            deleted = response.get("deleted", 0)
+            logger.info(f"Deleted {deleted} documents from ES for file_id={file_id}")
 
     except Exception as e:
-        logger.warning(f"Failed to delete milvus collection {knowledge.collection_name}: {str(e)}")
-        # 即使删除失败也继续执行，因为可能集合本来就不存在
+        logger.exception(f"Failed to delete ES files for knowledge_id={knowledge.id}: {str(e)}")
 
 
 def _rebuild_embeddings(knowledge: Knowledge, files: List[KnowledgeFile], new_model_id: str) -> tuple[