diff --git a/README.md b/README.md
index c3f5972..45ef1be 100644
--- a/README.md
+++ b/README.md
@@ -25,7 +25,7 @@ Use the default model to continue the conversation from given messages.
 ```python
 from towhee import ops
 
-chat = ops.LLM.Llama_2('llama-2-13b-chat', max_tokens=512)
+chat = ops.LLM.Llama_2('llama-2-13b-chat', n_ctx=4096, max_tokens=200)
 
 message = [
     {'system': 'You are a very helpful assistant.'},
diff --git a/llama2.py b/llama2.py
index f8b5fcc..ababbde 100644
--- a/llama2.py
+++ b/llama2.py
@@ -39,36 +39,32 @@ class LlamaCpp(PyOperator):
             self.model_path = model_name_or_file
         assert os.path.isfile(self.model_path), f'Invalid model path: {self.model_path}'
 
-        self.model = Llama(model_path=self.model_path)
+        init_kwargs = {}
+        for k in vars(Llama).keys() & self.kwargs.keys():
+            init_kwargs[k] = self.kwargs.pop(k)
+        self.model = Llama(model_path=self.model_path, **init_kwargs)
 
     def __call__(self, messages: List[dict]):
-        prompt = self.parse_inputs(messages)
-        resp = self.model(prompt, **self.kwargs)
+        messages = self.parse_inputs(messages)
+        resp = self.model.create_chat_completion(messages, **self.kwargs)
         answer = self.parse_outputs(resp)
         return answer
         
     def parse_inputs(self, messages: List[dict]):
         assert isinstance(messages, list), \
             'Inputs must be a list of dictionaries with keys from ["system", "question", "answer"].'
-        prompt = ''
-        question = messages.pop(-1)
-        assert len(question) == 1 and 'question' in question.keys()
-        question = question['question']
+        new_messages = []
         for m in messages:
             for k, v in m.items():
                 if k == 'system':
-                    prompt += f'''[INST] <<SYS>> {v} <</SYS>>\n'''
+                    new_messages.append({'role': 'system', 'content': v})
                 elif k == 'question':
-                    prompt += f''' {v} [/INST]\n'''
+                    new_messages.append({'role': 'user', 'content': v})
                 elif k == 'answer':
-                    prompt += f''' {v} '''
+                    new_messages.append({'role': 'assistant', 'content': v})
                 else:
                     raise KeyError(f'Invalid key of message: {k}')
-        if len(prompt) > 0:
-            prompt = '<s> ' + prompt + ' </s>' + f'<s> [INST] {question} [/INST]'
-        else:
-            prompt = question
-        return prompt
+        return new_messages
     
     def parse_outputs(self, response):
         return response['choices'][0]['text']