add history fix and calculate

Dockerfile

@@ -18,4 +18,4 @@ RUN python3 -m pip install --no-cache-dir -r requirements.txt
 
 EXPOSE 5000
 
-ENTRYPOINT [ "python3", "/jarvis/start.py"]
+ENTRYPOINT [ "python3", "/start.py"]

jarvis/api.py

@@ -44,9 +44,9 @@ def process_message(message):
 
     # intent_manager.recognise(message['data'], message['uuid'])
     if message['data'] != "":
-        # response = chatgpt_recognise(message['data'], message['uuid'])
+        response = chatgpt_utils.chatgpt_recognise(message['data'], message['uuid'])
-        # text_response = chatgpt_utils.get_answer_from_response(response)
+        text_response = chatgpt_utils.get_answer_from_response(response)
-        text_response = "Tokens are expensive ya know?"
+        # response = "Tokens are expensive ya know?"
 
         chat_utils.send_jarvis_message_to_room(text_response, message['uuid'])
 

jarvis/utils/chatgpt_prompt_2_smaller.txt

@@ -2,16 +2,17 @@ Respond to smart home requests in JSON format with HomeAssistant API terminology
 
 Requests groups:
   - command: change the state of an accessory (properties : location, device_class, device_description, value, comment, scheduleTimeStamp)
-  - query: only for retrieving a smart device state (properties : location, device_class, device_description, property)
+  - query: only for retrieving a smart device state never for general questions (properties : location, device_class, device_description, property)
-  - answer: for any questions (properties : answer)
+  - answer: for questions (properties : answer)
-  - clarify: when you don't understand, ask for details (properties : question)
+  - clarify: when you really don't understand, ask for details but try to guess what the user means (properties : question)
+  - calculate: for math questions, never give the answer (properties : calculation)
 NEVER add other properties
 
-Response:
+Response parameters:
 action: groups just mentioned
 location: room name, unknown if not obvious
 value: wanted state, song name, artist, temperature, etc.
-device_class: homeassistant device class
+device_class: switch, light, button, service
 device_description: information to identify the device later, include room and others identifiers
 
 For queries property "property" should be "state"
@@ -20,5 +21,7 @@ For commands property "scheduleTimeStamp" is for scheduling a command in the fut
 
 The house located at {{location}} and current time is {{timestamp}}.
 
-If questions about you, you are funny smart home AI like Jarvis from Iron Man, be nice and helpful with all topics.
+{{history}}
-Very important for you to only respond with a single valid JSON response and encapsulate every JSON property with double quotes "". Don't add anything and never excuse yourself. Respond to only one request at a time.
+
+You reluctantly answers questions with sarcastic responses.
+Always respond with a single valid JSON response, never excuse yourself.

jarvis/utils/chatgpt_utils.py

@@ -8,34 +8,45 @@ chat_messages = {}
 
 
 def chatgpt_recognise(text, uuid):
-    if len(chat_messages) == 0:
+    # If the chat history is empty, create it
-        chatgpt_init(uuid)
+    if uuid not in chat_messages:
+        chat_messages[uuid] = []
 
+    # Add the user message to the chat history
     chat_messages[uuid].append({"role": "user", "content": text})
 
     # Call ChatGPT API
     start_time = time.time()
     response = openai.ChatCompletion.create(
         model="gpt-3.5-turbo",
-        messages=chat_messages[uuid],
+        messages=get_conversation_as_one_message(uuid),
     )
     end_time = time.time()
-    logging.info("GPT-3 response in " + str(end_time - start_time) + " seconds")
+    logging.info("GPT-3 response in " + str(round(end_time - start_time, ndigits=2)) + " seconds")
 
-    # Check if the response is a "valid" JSON
     try:
+        # Parse the response
         response = json.loads(str(response.choices[0].message.content))
+
+        # Check if the response looks like a "valid" JSON
         if 'action' in response:
+            # Add the response to the chat history
             chat_messages[uuid].append({"role": "assistant", "content": get_answer_from_response(response)})
+
+            # Return the response
             return response
 
     except Exception as e:
+        # If the response is not a JSON, it's probably a plaintext response
         logging.error("Error while parsing ChatGPT response, probably not JSON: " + str(response.choices))
         logging.error(str(e))
+
+        # Try to parse the response as a plaintext response, if unsuccessful, return a "I don't understand" response
         return {"action": "answer", "answer": get_answer_from_response(response)}
 
 
 def clear_chat(uuid):
+    """Clear the chat history for a given uuid"""
     logging.info("Cleared chat for uuid " + uuid)
     chat_messages[uuid] = []
 
@@ -54,15 +65,45 @@ def get_answer_from_response(response):
             return response['device_description']
         elif response['action'] == 'answer':
             return response['answer']
+        elif response['action'] == 'calculate':
+            return response['calculation']
         else:
             return "I don't know how to respond to that..."
 
 
-def chatgpt_init(uuid):
+def get_conversation_as_one_message(uuid):
+    """
+    Prepare the messages to send to OpenAI API
+    We don't use OpenAI way of using an array of messages because it doesn't work well with the prompt and chatgpt
+    breaks out of character.
+    So instead, we add the history to the system prompt and create a new conversation each time.
+    It should not cost more tokens, but I am not sure at 100%.
+    """
+
+    # Load the prompt
     prompt = open("utils/chatgpt_prompt_2_smaller.txt", "r").read()
 
-    prompt.replace("{{timestamp}}", time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
+    # Replace the variables in the prompt
-    prompt.replace("{{location}}", "Lausanne in the canton Vaud of Switzerland")
+    prompt = prompt.replace("{{timestamp}}", time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
+    prompt = prompt.replace("{{location}}", "Lausanne in the canton Vaud of Switzerland")
 
-    chat_messages[uuid] = []
+    # Check if the user has already asked a question (to keep context)
-    chat_messages[uuid].append({"role": "system", "content": prompt})
+    if len(chat_messages[uuid]) > 1:
+        history = ""
+
+        # Add the last 4 messages from the user to the prompt
+        # we don't give the whole history because it adds a lot of tokens.
+        history += "\nFor context, last discussion you had with user:\n"
+        for message in chat_messages[uuid][-4:]:
+            if message['role'] == "user":
+                history += "U: " + message['content'] + "\n"
+            elif message['role'] == "assistant":
+                history += "Y: " + message['content'] + "\n"
+
+        # Replace the {{history}} variable in the prompt with the history
+        prompt = prompt.replace("{{history}}", history)
+    else:
+        # If the user hasn't asked a question yet, remove the history part of the prompt
+        prompt = prompt.replace("{{history}}", "")
+
+    return [{"role": "system", "content": prompt}]

start.py

@@ -2,7 +2,7 @@ import logging
 
 import lingua_franca
 
-import api
+import jarvis.api
 from jarvis.skills.cocktails import CocktailSkill
 from jarvis.skills.intent_services import intent_manager
 from jarvis.utils import whisper_utils
@@ -23,4 +23,4 @@ if __name__ == '__main__':
     whisper_utils.load_model()
 
     # Start the api endpoint
-    api.start_api()
+    jarvis.api.start_api()