from openai import OpenAI import json_repair class ChatbotSimulation: def __init__(self, site_map, page_details, user_state, task, app_name, log_location, openai_api_key, agent='human', max_steps=50, max_tokens=8192, buffer_tokens=500): self.sitemap = site_map self.page_details = page_details self.user_state = user_state self.user_state['current_page'] = 'Home' # Initialize current page self.user_state['task_completed'] = 0 self.task = task self.app_name = app_name self.log_location = log_location self.agent = agent.lower() if self.agent not in ['human', 'llm']: raise ValueError("Invalid agent type. Expected 'Human' or 'llm'.") self.max_steps = max_steps self.max_tokens = max_tokens self.buffer_tokens = buffer_tokens self.conversation = [] # Stores recent conversation snippets self.prompt_count = 0 self.client = OpenAI(api_key=openai_api_key) self.actions = [] def _get_page_uid(self, page_name): """Retrieve the UID of the given page from the sitemap.""" return self.sitemap['pages'].get(page_name, {}).get('uid') def _get_page_details(self, page_name): """Retrieve the page details using its UID.""" uid = self._get_page_uid(page_name) return self.page_details.get(uid, {}) def _generate_system_prompt(self): """Create a dynamic system prompt based on the current state.""" current_page = self.user_state['current_page'] page_info = self._get_page_details(current_page) return f""" You are interacting with the {self.app_name} app. Your task is: {self.task}. You are currently on the {current_page} page. Current user state: {self.user_state}. Page Information: {page_info} - **Features**: Represent available options the user can select on this page. - **User Data**: Represents user-specific data accessible on this page. Provide instructions or request input from the user. If the user provides an invalid action, respond with: "Invalid action. Please select a valid option." ### Instruction Format: You are at the {current_page} page. You have the following options: 1. Feature 1 2. Feature 2 3. Feature 3 4. Feature 4 Please enter your choice as the corresponding number. Rules: - Be sure to display all options that is available in features. - Be robotic and emotionless. Avoid offering any advice to the user. - If a feature requires `input_text`, request input as: "Enter query as: [number]: query" """ def _get_openai_response(self, prompt): """Fetch response from OpenAI API.""" self._trim_conversation() response = self.client.chat.completions.create( model="gpt-4", messages=prompt, max_tokens=1500, # Adjusted max_tokens if needed temperature=1.0, ) return response.choices[0].message.content def _calculate_token_count(self, conversation): """Estimate the token count in the conversation.""" return sum(len(entry['content'].split()) for entry in conversation) def _trim_conversation(self): """Trim the conversation to keep it within the token limit.""" while self._calculate_token_count(self.conversation) > self.max_tokens - self.buffer_tokens: self.conversation.pop(0) # def one_conversation_round(self, user_input): """Conduct one round of conversation between the user and the assistant.""" # User provides input self.conversation.append({"role": "user", "content": user_input}) self.actions.append(user_input) # Update user state using GPT's response update_prompt = f""" Update the user state based on the input: '{user_input}'. Current user state (JSON format): {self.user_state} Sitemap: {self.sitemap} Instructions: 1. If the 'current_page' has changed, update it to a page from the sitemap. 2. If the task is finished, update 'task_completed' to 1. Otherwise, leave it unchanged. 3. If no updates are needed, return the user state exactly as provided, without modification. 4. Preserve the **exact JSON structure** and **format** of the provided user state. 5. The output **must be a single JSON dictionary** representing the updated user state—do not wrap it in a list. 6. Do not change any other fields unless explicitly required by the instructions. Important: - Ensure 'current_page' and 'task_completed' are keys in the returned dictionary. - Return **only the JSON object** without additional output or wrapping. - **AVOID OUTPUT A LIST**, must be JSON! The format for each page should be: {{ "page_name": page_name "features": {{}}, "user_data": {{ "Data Entry 1": "Details of Data Entry 1", "Data Entry 2": "Details of Data Entry 2" }} }} """ self.conversation.append({"role": "assistant", "content": update_prompt}) updated_state = self._get_openai_response(self.conversation) # Parse and update the user state updated_state = json_repair.loads(updated_state) if isinstance(updated_state, list): reformat_prompt = f''' Given the {updated_state}, reformat it into a proper JSON with only 3 keys: page_name, features, user_data Follow the format: {{ "page_name": page_name "features": {{}}, "user_data": {{ "Data Entry 1": "Details of Data Entry 1", "Data Entry 2": "Details of Data Entry 2" }} }} ''' self.conversation.append({"role": "assistant", "content": reformat_prompt}) reformat_state = self._get_openai_response(self.conversation) updated_state = json_repair.loads(reformat_state) if updated_state['task_completed']: return f"Task completed! You took {self.prompt_count} steps." self.user_state = updated_state system_prompt = self._generate_system_prompt() # GPT generates the page instructions self.conversation.append({"role": "system", "content": system_prompt}) gpt_instruction = self._get_openai_response(self.conversation) self.conversation.append({"role": "assistant", "content": gpt_instruction}) return gpt_instruction def _generate_agent_input(self): """ Simulate the agent generating input based on the conversation state. """ agent_prompt = f""" Imagine you are an agent navigate through the Uber environment. Your overarching task is: {self.task}. You may have done some part of the task, or none at all. You will have access to all of your previous actions in the environment, as well as the last message from the assistant giving the current state of the environment. The last message from the assistant was: {self.conversation[-1]['content']} Respond first with a brief "Plan" which suggests what steps you are going to take to accomplish the task, and what your immediate. Then generate an "Action" which is the immediate next step you can take. """ messages = [{"role": "system", "content": agent_prompt}] for action in self.actions: messages.append({"role": "user", "content": action}) messages.append({"role": "assistant", "content": self.conversation[-1]['content']}) agent_response = self.client.chat.completions.create( model="gpt-4", messages=messages, temperature=1.0, ) print(f"LLM-Agent: {agent_response.choices[0].message.content}") return agent_response.choices[0].message.content def start_conversation(self): greeting = f'\n Welcome to {self.app_name} simulator! Your task is: {self.task}. \n' system_prompt = self._generate_system_prompt() # GPT generates the page instructions self.conversation.append({"role": "system", "content": system_prompt}) gpt_instruction = self._get_openai_response(self.conversation) self.conversation.append({"role": "assistant", "content": gpt_instruction}) return greeting + gpt_instruction