Update app.py

added: client_kwargs={
"max_retries": 3,
"timeout": 30
}

app.py

@@ -16,15 +16,21 @@ from langchain_classic.chains import RetrievalQA
 
 # 1. LLM SETUP (Using Granite via HF API)
 def get_llm():
-    return HuggingFaceEndpoint(
+    token = os.getenv("HF_TOKEN")
+    
+    llm = HuggingFaceEndpoint(
         repo_id="ibm-granite/granite-3.1-8b-instruct", 
         task="text-generation",
         max_new_tokens=512,
         temperature=0.1,
-        huggingfacehub_api_token=os.getenv("HF_TOKEN"),
-        max_retries=3, 
-        timeout=30     
+        huggingfacehub_api_token=token,
+        
+        client_kwargs={
+            "max_retries": 3,
+            "timeout": 30
+        }
     )
+    return llm
 
 # 2. DOCUMENT PROCESSING
 def process_pdf(file_path):