Update src/simple_rag.py

src/simple_rag.py

@@ -114,8 +114,12 @@ def ask_question(query_text: str, k: int = 3):
             extra_body={
                 "chat_template_kwargs": {
                     "thinking_mode": "off"
-                }
+                },
+            "cache": {
+                "no-cache": True
+            }
             },
+            
             max_tokens=512
         )
         answer = completion.choices[0].message.content.strip()