import os import subprocess import random from typing import List, Dict, Tuple from datetime import datetime import logging import gradio as gr from huggingface_hub import InferenceClient # --- Configuration --- MODEL_NAME = "mistralai/Mixtral-8x7B-Instruct-v0.1" # Hugging Face model for text generation MAX_HISTORY_TURNS = 5 # Number of previous turns to include in the prompt MAX_TOKENS_PER_TURN = 2048 # Maximum number of tokens to generate per turn VERBOSE_LOGGING = True # Enable verbose logging for debugging DEFAULT_AGENT = "WEB_DEV" # Default agent to use # --- Logging Setup --- logging.basicConfig( filename="app.log", # Name of the log file level=logging.INFO, # Set the logging level (INFO, DEBUG, etc.) format="%(asctime)s - %(levelname)s - %(message)s", ) # --- Agent Definitions --- class Agent: """Base class for all agents.""" def __init__(self, name: str, description: str): self.name = name self.description = description def handle_action(self, action: str, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles an action from the user. Args: action: The action name. action_input: The input for the action. history: The conversation history. task: The current task. Returns: A tuple containing the next action name, action input, updated history, and updated task. """ raise NotImplementedError("Agent subclasses must implement handle_action.") def get_prompt(self, message: str, history: List[Tuple[str, str]], task: str) -> str: """Generates a prompt for the language model. Args: message: The user's message. history: The conversation history. task: The current task. Returns: The prompt string. """ now = datetime.now() date_time_str = now.strftime("%Y-%m-%d %H:%M:%S") prompt = f""" {date_time_str} Agent: {self.name} Task: {task} History: {self.format_history(history)} Message: {message} """ return prompt def format_history(self, history: List[Tuple[str, str]]) -> str: """Formats the conversation history for the prompt.""" formatted_history = "" for user_message, agent_response in history[-MAX_HISTORY_TURNS:]: formatted_history += f"[INST] {user_message} [/INST]\n{agent_response}\n" return formatted_history class WebDevAgent(Agent): """Agent for web development tasks.""" def __init__(self): super().__init__(name="WEB_DEV", description="Agent specialized in web development tasks.") def handle_action(self, action: str, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: if action == "SEARCH": return self._handle_search_action(action_input, history, task) elif action == "GENERATE_HTML": return self._handle_generate_html_action(action_input, history, task) elif action == "GENERATE_CSS": return self._handle_generate_css_action(action_input, history, task) elif action == "GENERATE_JS": return self._handle_generate_js_action(action_input, history, task) elif action == "COMPLETE": return "COMPLETE", "COMPLETE", history, task else: return "MAIN", None, history, task def _handle_search_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the SEARCH action.""" if VERBOSE_LOGGING: logging.info(f"Calling SEARCH action with input: {action_input}") try: if "http" in action_input: if "<" in action_input: action_input = action_input.strip("<") if ">" in action_input: action_input = action_input.strip(">") response = i_s(action_input) # Use i_search for web search history.append(("observation: search result is:", response)) else: history.append(("observation: I need a valid URL for the SEARCH action.", "")) except Exception as e: history.append(("observation:", str(e))) return "MAIN", None, history, task def _handle_generate_html_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the GENERATE_HTML action.""" if VERBOSE_LOGGING: logging.info(f"Calling GENERATE_HTML action with input: {action_input}") # Simulate OpenAI's code generation capabilities using Hugging Face prompt = self.get_prompt(f"Generate HTML code for a web page that {action_input}", history, task) response = run_gpt(prompt, stop_tokens=["```", "```html"], max_tokens=MAX_TOKENS_PER_TURN) history.append(("observation: generated HTML code:", response)) return "MAIN", None, history, task def _handle_generate_css_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the GENERATE_CSS action.""" if VERBOSE_LOGGING: logging.info(f"Calling GENERATE_CSS action with input: {action_input}") # Simulate OpenAI's code generation capabilities using Hugging Face prompt = self.get_prompt(f"Generate CSS code for a web page that {action_input}", history, task) response = run_gpt(prompt, stop_tokens=["```", "```css"], max_tokens=MAX_TOKENS_PER_TURN) history.append(("observation: generated CSS code:", response)) return "MAIN", None, history, task def _handle_generate_js_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the GENERATE_JS action.""" if VERBOSE_LOGGING: logging.info(f"Calling GENERATE_JS action with input: {action_input}") # Simulate OpenAI's code generation capabilities using Hugging Face prompt = self.get_prompt(f"Generate JavaScript code for a web page that {action_input}", history, task) response = run_gpt(prompt, stop_tokens=["```", "```js"], max_tokens=MAX_TOKENS_PER_TURN) history.append(("observation: generated JavaScript code:", response)) return "MAIN", None, history, task class AiSystemPromptAgent(Agent): """Agent for generating system prompts.""" def __init__(self): super().__init__(name="AI_SYSTEM_PROMPT", description="Agent specialized in generating system prompts.") def handle_action(self, action: str, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: if action == "GENERATE_PROMPT": return self._handle_generate_prompt_action(action_input, history, task) elif action == "COMPLETE": return "COMPLETE", "COMPLETE", history, task else: return "MAIN", None, history, task def _handle_generate_prompt_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the GENERATE_PROMPT action.""" if VERBOSE_LOGGING: logging.info(f"Calling GENERATE_PROMPT action with input: {action_input}") # Simulate OpenAI's prompt generation capabilities using Hugging Face prompt = self.get_prompt(f"Generate a system prompt for a language model that {action_input}", history, task) response = run_gpt(prompt, stop_tokens=["```", "```json"], max_tokens=MAX_TOKENS_PER_TURN) history.append(("observation: generated system prompt:", response)) return "MAIN", None, history, task class PythonCodeDevAgent(Agent): """Agent for Python code development tasks.""" def __init__(self): super().__init__(name="PYTHON_CODE_DEV", description="Agent specialized in Python code development tasks.") def handle_action(self, action: str, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: if action == "GENERATE_CODE": return self._handle_generate_code_action(action_input, history, task) elif action == "RUN_CODE": return self._handle_run_code_action(action_input, history, task) elif action == "COMPLETE": return "COMPLETE", "COMPLETE", history, task else: return "MAIN", None, history, task def _handle_generate_code_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the GENERATE_CODE action.""" if VERBOSE_LOGGING: logging.info(f"Calling GENERATE_CODE action with input: {action_input}") # Simulate OpenAI's code generation capabilities using Hugging Face prompt = self.get_prompt(f"Generate Python code that {action_input}", history, task) response = run_gpt(prompt, stop_tokens=["```", "```python"], max_tokens=MAX_TOKENS_PER_TURN) history.append(("observation: generated Python code:", response)) return "MAIN", None, history, task def _handle_run_code_action(self, action_input: str, history: List[Tuple[str, str]], task: str) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the RUN_CODE action.""" if VERBOSE_LOGGING: logging.info(f"Calling RUN_CODE action with input: {action_input}") # Simulate OpenAI's code execution capabilities using Hugging Face prompt = self.get_prompt(f"Run the following Python code and provide the output: {action_input}", history, task) response = run_gpt(prompt, stop_tokens=["```", "```python"], max_tokens=MAX_TOKENS_PER_TURN) history.append(("observation: code output:", response)) return "MAIN", None, history, task # --- Action Handlers --- def handle_main_action(action: str, action_input: str, history: List[Tuple[str, str]], task: str, agent: Agent) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the MAIN action, which is the default action.""" if VERBOSE_LOGGING: logging.info(f"Calling MAIN action with input: {action_input}") prompt = agent.get_prompt(action_input, history, task) response = run_gpt(prompt, stop_tokens=["observation:", "task:", "action:", "thought:"], max_tokens=MAX_TOKENS_PER_TURN) if VERBOSE_LOGGING: logging.info(f"Response from model: {response}") history.append((action_input, response)) lines = response.strip().strip("\n").split("\n") for line in lines: if line == "": continue if line.startswith("thought: "): history.append((line, "")) if VERBOSE_LOGGING: logging.info(f"Thought: {line}") elif line.startswith("action: "): action_name, action_input = parse_action(line) history.append((line, "")) if VERBOSE_LOGGING: logging.info(f"Action: {action_name} - {action_input}") if "COMPLETE" in action_name or "COMPLETE" in action_input: task = "END" return action_name, action_input, history, task else: return action_name, action_input, history, task else: history.append((line, "")) if VERBOSE_LOGGING: logging.info(f"Other Output: {line}") return "MAIN", None, history, task def handle_update_task_action(action: str, action_input: str, history: List[Tuple[str, str]], task: str, agent: Agent) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the UPDATE-TASK action, which updates the current task.""" if VERBOSE_LOGGING: logging.info(f"Calling UPDATE-TASK action with input: {action_input}") prompt = agent.get_prompt(action_input, history, task) task = run_gpt(prompt, stop_tokens=[], max_tokens=64).strip("\n") history.append(("observation: task has been updated to:", task)) return "MAIN", None, history, task def handle_search_action(action: str, action_input: str, history: List[Tuple[str, str]], task: str, agent: Agent) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the SEARCH action, which performs a web search.""" if VERBOSE_LOGGING: logging.info(f"Calling SEARCH action with input: {action_input}") try: if "http" in action_input: if "<" in action_input: action_input = action_input.strip("<") if ">" in action_input: action_input = action_input.strip(">") response = i_s(action_input) # Use i_search for web search history.append(("observation: search result is:", response)) else: history.append(("observation: I need a valid URL for the SEARCH action.", "")) except Exception as e: history.append(("observation:", str(e))) return "MAIN", None, history, task def handle_complete_action(action: str, action_input: str, history: List[Tuple[str, str]], task: str, agent: Agent) -> Tuple[str, str, List[Tuple[str, str]], str]: """Handles the COMPLETE action, which ends the current task.""" if VERBOSE_LOGGING: logging.info(f"Calling COMPLETE action.") task = "END" return "COMPLETE", "COMPLETE", history, task # --- Action Mapping --- ACTION_HANDLERS: Dict[str, callable] = { "MAIN": handle_main_action, "UPDATE-TASK": handle_update_task_action, "SEARCH": handle_search_action, "COMPLETE": handle_complete_action, } # --- Utility Functions --- def run_gpt(prompt: str, stop_tokens: List[str], max_tokens: int) -> str: """Runs the language model and returns the generated text.""" if VERBOSE_LOGGING: logging.info(f"Prompt: {prompt}") client = InferenceClient(MODEL_NAME) resp = client.text_generation(prompt, max_new_tokens=max_tokens, stop_sequences=stop_tokens, temperature=0.7, top_p=0.8, repetition_penalty=1.5) if VERBOSE_LOGGING: logging.info(f"Response: {resp}") return resp def parse_action(line: str) -> Tuple[str, str]: """Parses an action line to get the action name and input.""" parts = line.split(":", 1) if len(parts) == 2: action_name = parts[0].replace("action", "").strip() action_input = parts[1].strip() else: action_name = parts[0].replace("action", "").strip() action_input = "" return action_name, action_input def run_agent(purpose: str, history: List[Tuple[str, str]], agent: Agent) -> List[Tuple[str, str]]: """Runs the agent and returns the updated conversation history.""" task = None directory = "./" action_name = "UPDATE-TASK" if task is None else "MAIN" action_input = None while True: if VERBOSE_LOGGING: logging.info(f"---") logging.info(f"Purpose: {purpose}") logging.info(f"Task: {task}") logging.info(f"---") logging.info(f"History: {history}") logging.info(f"---") if VERBOSE_LOGGING: logging.info(f"Running action: {action_name} - {action_input}") try: if "RESPONSE" in action_name or "COMPLETE" in action_name: action_name = "COMPLETE" task = "END" return history if action_name not in ACTION_HANDLERS: action_name = "MAIN" if action_name == "" or action_name is None: action_name = "MAIN" action_handler = ACTION_HANDLERS[action_name] action_name, action_input, history, task = action_handler(action_name, action_input, history, task, agent) yield history if task == "END": return history except Exception as e: history.append(("observation: the previous command did not produce any useful output, I need to check the commands syntax, or use a different command", "")) logging.error(f"Error in run_agent: {e}") return history # --- Gradio Interface --- def main(): with gr.Blocks() as demo: gr.Markdown("## FragMixt: Your No-Code Development Powerhouse") gr.Markdown("### Agents w/ Agents: Mastering No-Code Development") # Chat Interface chatbot = gr.Chatbot(show_label=False, show_share_button=False, show_copy_button=True, likeable=True, layout="panel") # Input Components message = gr.Textbox(label="Enter your message", placeholder="Ask me anything!") purpose = gr.Textbox(label="Purpose", placeholder="What is the purpose of this interaction?") agent_name = gr.Dropdown(label="Agents", choices=[agent.name for agent in [WebDevAgent(), AiSystemPromptAgent(), PythonCodeDevAgent()]], value=DEFAULT_AGENT, interactive=True) sys_prompt = gr.Textbox(label="System Prompt", max_lines=1, interactive=True) temperature = gr.Slider(label="Temperature", value=0.9, minimum=0.0, maximum=1.0, step=0.05, interactive=True, info="Higher values produce more diverse outputs") max_new_tokens = gr.Slider(label="Max new tokens", value=1048*10, minimum=0, maximum=1048*10, step=64, interactive=True, info="The maximum numbers of new tokens") top_p = gr.Slider(label="Top-p (nucleus sampling)", value=0.90, minimum=0.0, maximum=1, step=0.05, interactive=True, info="Higher values sample more low-probability tokens") repetition_penalty = gr.Slider(label="Repetition penalty", value=1.2, minimum=1.0, maximum=2.0, step=0.05, interactive=True, info="Penalize repeated tokens") # Button to submit the message submit_button = gr.Button(value="Send") # Project Explorer Tab with gr.Tab("Project Explorer"): project_path = gr.Textbox(label="Project Path", placeholder="/home/user/app/current_project") explore_button = gr.Button(value="Explore") project_output = gr.Textbox(label="File Tree", lines=20) # Chat App Logic Tab with gr.Tab("Chat App"): history = gr.State([]) examples = [ ["What is the purpose of this AI agent?", "I am designed to assist with no-code development tasks."], ["Can you help me generate a Python function to calculate the factorial of a number?", "Sure! Here is a Python function to calculate the factorial of a number:"], ["Generate a web page with a heading that says 'Welcome to My Website!'", "action: GENERATE_HTML action_input=a heading that says 'Welcome to My Website!'"], ] def chat(purpose, message, agent_name, sys_prompt, temperature, max_new_tokens, top_p, repetition_penalty, history): if agent_name == "WEB_DEV": agent = WebDevAgent() elif agent_name == "AI_SYSTEM_PROMPT": agent = AiSystemPromptAgent() elif agent_name == "PYTHON_CODE_DEV": agent = PythonCodeDevAgent() else: agent = WebDevAgent() # Default to WEB_DEV if agent_name is invalid history = list(run_agent(purpose, history, agent)) return history, history submit_button.click(chat, inputs=[purpose, message, agent_name, sys_prompt, temperature, max_new_tokens, top_p, repetition_penalty, history], outputs=[chatbot, history]) demo.launch() if __name__ == "__main__": main()