update generate parameters

app.py

@@ -29,10 +29,14 @@ def chat_func(input_utterance: str, history: Optional[List[str]] = None):
     input_ids = tokenizer(history_str,
                           return_tensors='pt',
                           truncation=True,
-                          max_length=max_length).input_ids
+                          max_length=max_length,
+                          top_p=0.95,
+                          do_sample=True,
+                          num_beams=4).input_ids
 
     output_ids = model.generate(input_ids,
                                 max_new_tokens=30)[0]
+
     response = tokenizer.decode(output_ids, skip_special_tokens=True)
 
     history.append(response)