diff --git a/app/create_app.py b/app/create_app.py
index 7ed9bae..6972ec1 100644
--- a/app/create_app.py
+++ b/app/create_app.py
@@ -10,6 +10,7 @@ from starlette.staticfiles import StaticFiles
 from app.middlewares.app_middlewares import add_app_middlewares
 from app.utils.db_utils import check_database_connection
 from app.utils.milvus_utils import milvus_service
+from app.utils.nltk_utils import check_nltk
 from app.utils.postgres_checkpointer import check_postgres_connection, close_postgres_connection
 from app.utils.redis_utils import redis_utils
 
@@ -23,6 +24,7 @@ def create_app():
       asyncio.create_task(check_postgres_connection()),
       asyncio.create_task(milvus_service.check_milvus_connection()),
       asyncio.create_task(redis_utils.check_redis_connection()),
+      asyncio.create_task(check_nltk()),
     )
     async_engine = async_results[0]
     yield
diff --git a/app/utils/milvus_utils.py b/app/utils/milvus_utils.py
index af9d8b3..62b4adc 100644
--- a/app/utils/milvus_utils.py
+++ b/app/utils/milvus_utils.py
@@ -72,10 +72,6 @@ class MilvusService:
       embed_model=self.embeddings,
     )
 
-    # 预先加载NLTK语料库以避免多线程环境中的竞争条件
-    # 否则在多线程环境下，parser.get_nodes_from_documents 可能会出现报错信息：'WordListCorpusReader' object has no attribute '_LazyCorpusLoader__args'
-    load_nltk()
-
     from llama_index.core.node_parser import SimpleNodeParser
     # parser = HierarchicalNodeParser.from_defaults(chunk_sizes=[2048, 512, 128])
     parser = SimpleNodeParser()
diff --git a/app/utils/nltk_utils.py b/app/utils/nltk_utils.py
index 12f4ffd..a22b306 100644
--- a/app/utils/nltk_utils.py
+++ b/app/utils/nltk_utils.py
@@ -1,14 +1,35 @@
 # 预先加载NLTK语料库以避免多线程环境中的竞争条件
+import asyncio
+
 import nltk
-
-load_flag = False
+from nltk.data import find
 
 
+# 判断nltk是否已经存在
+def is_nltk_package_downloaded(package_name):
+  try:
+    find(f'tokenizers/{package_name}')
+    return True
+  except LookupError:
+    return False
+
+
+# 下载nltk所需要文件
 def load_nltk():
-  global load_flag
-  if not load_flag:
+  if not is_nltk_package_downloaded("punkt"):
+    print("ℹ️ 下载nltk")
     nltk.download('punkt')
     nltk.download('averaged_perceptron_tagger')
     nltk.download('maxent_ne_chunker')
     nltk.download('words')
-    load_flag = True
+    print("✅ nltk下载成功")
+  else:
+    print("✅ nltk已经存在，无需下载")
+
+
+# 检查nltk是否已经下载，用于llama-index做文本分割
+async def check_nltk():
+  print("nltk_path:", nltk.data.path)
+  # 预先加载NLTK语料库以避免多线程环境中的竞争条件
+  # 否则在多线程环境下，parser.get_nodes_from_documents 可能会出现报错信息：'WordListCorpusReader' object has no attribute '_LazyCorpusLoader__args'
+  await asyncio.wait_for(asyncio.to_thread(load_nltk), timeout=10)