Using the new gpt-3.5-turbo-0125

app.py

@@ -394,7 +394,7 @@ def main():
                 
                                 # let's use Llama2, and fall back on GPT3.5 if it fails
                                 reworded_query = reword_query(query, guest, 
-                                                            model_name='llama2-13b-chat')
+                                                            model_name='gpt-3.5-turbo-0125')
                                 new_query = reworded_query['rewritten_question']
                                 
                                 if reworded_query['status'] != 'error': # or reworded_query['changed']: 
@@ -566,21 +566,21 @@ def reword_query(query, guest, model_name='llama2-13b-chat', response_processing
     prompt_fields = {
         "you_are":f"You are an expert in linguistics and semantics, analyzing the question asked by a user to a vector search system, \
                     and making sure that the question is well formulated and that the system can understand it.",
-                    
+
         "your_task":f"Your task is to detect if the name of the guest ({guest}) is mentioned in the user's question, \
                     and if that is not the case, rewrite the question using the guest name, \
                     without changing the meaning of the question. \
                     Most of the time, the user will have used a pronoun to designate the guest, in which case, \
                     simply replace the pronoun with the guest name.",
-                    
-        "question":f"If the user mentions the guest name, ie {query}, just return his question as is. \
+
+        "question":f"If the user mentions the guest name ({guest}) in the following question '{query}', just return his question as is. \
                     If the user does not mention the guest name, rewrite the question using the guest name.",
-        
-        "final_instruction":f"Only regerate the requested rewritten question or the original, WITHOUT ANY COMMENT OR REPHRASING. \
+
+        "final_instruction":f"Only regenerate the requested rewritten question or the original, WITHOUT ANY COMMENT OR REPHRASING. \
                     Your answer must be as close as possible to the original question, \
                     and exactly identical, word for word, if the user mentions the guest name, i.e. {guest}.",
     }
-    
+
     # prompt created by chatGPT :-) 
     # and Llama still outputs the original question and precedes the answer with 'rewritten question' 
     prompt_fields2 = {
@@ -596,7 +596,7 @@ def reword_query(query, guest, model_name='llama2-13b-chat', response_processing
     "question": (
         f"Original question: '{query}'. "
         "Rewrite this question to include the guest's FULL name if it's not already mentioned."
-        "The Only thing you can and MUST add is a question mark if it's missing."
+        "Add a question mark if it's missing, nothing else."
     ),
     "final_instruction": (
         "Create a rewritten question or keep the original question as is. "
@@ -652,7 +652,7 @@ def reword_query(query, guest, model_name='llama2-13b-chat', response_processing
             # st.write(f"Error in LLM response: 'error':{reworded_query['error']}")
             # st.write("The LLM could not connect to the server. Please try again later.")
             # st.stop()
-            return reword_query(query, guest, model_name='gpt-3.5-turbo-0613')
+            return reword_query(query, guest, model_name='gpt-3.5-turbo-0125')
             
         if response_processing:
             if isinstance(response, list) and isinstance(response[0], dict) and 'generated_text' in response[0]:
@@ -666,7 +666,7 @@ def reword_query(query, guest, model_name='llama2-13b-chat', response_processing
                 return result | {'status': 'success'}
             else:
                 print("Found no answer")
-                return reword_query(query, guest, model_name='gpt-3.5-turbo-0613')
+                return reword_query(query, guest, model_name='gpt-3.5-turbo-0125')
                 # return {'original_question': query, 'rewritten_question': query, 'changed': False, 'status': 'no properly formatted answer' }
         else:
             return response
@@ -677,36 +677,51 @@ def reword_query(query, guest, model_name='llama2-13b-chat', response_processing
         # return response[0]['generated_text'] .split("Rewritten question: '")[-1][:-1]
     
     else:
-        # assume openai 
-        model_ids = ['gpt-3.5-turbo-16k', 'gpt-3.5-turbo-0613']
-        model_name = model_ids[1]
-        GPTllm = GPT_Turbo(model=model_name, 
-                            api_key=openai_api_key)
+        # we assume / force openai 
+        model_ids = ['gpt-3.5-turbo-0125', 'gpt-3.5-turbo-16k', 'gpt-3.5-turbo-0613']
+        if model_name not in model_ids:
+            model_name = model_ids[0]
+        GPTllm = GPT_Turbo(model=model_name, api_key=openai_api_key)
         
         openai_prompt = """
-        {your_task}\n
+        {your_task} \n
+        {final_instruction} /n 
         ```
         \n\n
         Question: {question}\n
-        {final_instruction} 
 
         Answer:
         """
         prompt = openai_prompt.format(**prompt_fields)
-        
+
+        openai_prompt2 = """
+        {your_task}\n
+        ```
+        \n\n
+        {final_instruction} 
+        """
+        prompt2 = openai_prompt2.format(**{'your_task':prompt_fields['your_task'], 
+                                          'final_instruction':prompt_fields['final_instruction']})
+
         try:
-            resp = GPTllm.get_chat_completion(prompt=openai_prompt,
+            # https://platform.openai.com/docs/guides/text-generation/chat-completions-api
+            resp = GPTllm.get_chat_completion(prompt=prompt,
                                             system_message=prompt_fields['you_are'],
+                                            user_message = None, #prompt_fields['question'],
                                             temperature=0.01,
-                                            max_tokens=1500, # it's a question...
+                                            max_tokens=1500, # it's a long question...
                                             show_response=True,
                                             stream=False)
-            return {'rewritten_question': resp.choices[0].delta.content,
-                    'changed': True, 'status': 'success'}
+
+            if resp.choices[0].finish_reason == 'stop':
+                return {'rewritten_question': resp.choices[0].message.content,
+                        'changed': True, 'status': 'success'}
+            else:
+                raise Exception("LLM did not stop")  # to go to the except block
         except Exception:
             return {'rewritten_question': query, 'changed': False, 'status': 'not success'}
-         
+
 
 if __name__ == '__main__':
     main()
-# %%
+    # streamlit run app.py --server.allowRunOnSave True

openai_interface.py

@@ -21,6 +21,7 @@ class GPT_Turbo:
     def get_chat_completion(self, 
                             prompt: str, 
                             system_message: str='You are a helpful assistant.',
+                            user_message: str=None,
                             temperature: int=0, 
                             max_tokens: int=500,
                             stream: bool=False,
@@ -30,6 +31,8 @@ class GPT_Turbo:
             {'role': 'system', 'content': system_message},
             {'role': 'assistant', 'content': prompt}
                     ]
+        if user_message is not None:
+            messages.append({'role': 'user', 'content': user_message})
         
         response = self.client.chat.completions.create( model=self.model,
                                                         messages=messages,