Update app.py

app.py

@@ -110,10 +110,10 @@ retriever = db.as_retriever(
 # Set up the LLM
 llm = ChatOpenAI(
     base_url="https://api-inference.huggingface.co/v1/",
-    temperature=0,
+    temperature=0.1,
     api_key=HF_TOKEN,
     model="Qwen/Qwen2.5-Coder-32B-Instruct",
-    stream=True
+    max_new_tokens=4096
     
 )