Decrease tokens by sending less messages

functions.py

@@ -23,7 +23,7 @@ def get_main_data():
 
     authors = ['Sofia', 'Eliza', 'Sindy', 'Carlos', 'Andres', 'Adriana', 'Carolina', 'Valeria']
     model = "gpt-4"
-    temperature_values = [0.8, 1.0, 1.2]
+    temperature_values = [0.8, 1.0]
 
     return scores_parameters, authors, model, temperature_values
 
@@ -106,13 +106,19 @@ def handle_call(msg_history: gr.State, temperature: gr.State):
 def get_ai_answer(msg: str, msg_history: gr.State, temperature: gr.State):
     """
     Returns the response given by the model, all the message history so far and the seconds
-    the api took to retrieve such response.
+    the api took to retrieve such response. It also removes some messages in the message history
+    so only the last n (keep) are used (costs are cheaper)
     """
     msg_history.append({"role": "user", "content": msg})
     response, needed_time = handle_call(msg_history, temperature)
     AI_response = response["choices"][0]["message"]["content"]
     msg_history.append({'role': 'assistant', 'content': AI_response})
 
+    keep = 3  # Number of messages to keep
+    if len(msg_history) > (2 * keep) + 1:  # last +1 is due to the system message
+        msg_history.pop(1)
+        msg_history.pop(1)
+
     return AI_response, msg_history, needed_time
 
 

prompt.txt

@@ -67,7 +67,6 @@ Manejo de situaciones no contempladas:
 
 Informacion importante:
 - Antes de responder, asegúrate de que tu mensaje:
-  * Tenga 50 palabras o menos.
   * Refleje la personalidad de Roomie.
   * Esté alineado con los escenarios proporcionados.
   * Sea coherente con el tono y estilo de comunicación definidos anteriormente.