import os import re import shutil from pathlib import Path from smolagents.agent_types import AgentAudio, AgentImage, AgentText from smolagents.agents import MultiStepAgent, PlanningStep from smolagents.memory import ActionStep, FinalAnswerStep, MemoryStep from smolagents.models import ChatMessageStreamDelta from smolagents.utils import _is_package_available import xml.etree.ElementTree as ET def get_step_footnote_content(step_log: MemoryStep, step_name: str) -> str: """Get a footnote string for a step log with duration and token information""" step_footnote = f"**{step_name}**" if hasattr(step_log, "input_token_count") and hasattr(step_log, "output_token_count"): token_str = f" | Input tokens:{step_log.input_token_count:,} | Output tokens: {step_log.output_token_count:,}" step_footnote += token_str if hasattr(step_log, "duration"): step_duration = f" | Duration: {round(float(step_log.duration), 2)}" if step_log.duration else None step_footnote += step_duration step_footnote_content = f"""{step_footnote} """ return step_footnote_content def pull_messages_from_step(step_log: MemoryStep, skip_model_outputs: bool = False): """Extract ChatMessage objects from agent steps with proper nesting. Args: step_log: The step log to display as gr.ChatMessage objects. skip_model_outputs: If True, skip the model outputs when creating the gr.ChatMessage objects: This is used for instance when streaming model outputs have already been displayed. """ if not _is_package_available("gradio"): raise ModuleNotFoundError( "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`" ) import gradio as gr if isinstance(step_log, ActionStep): # Output the step number step_number = f"Step {step_log.step_number}" if step_log.step_number is not None else "Step" # First yield the thought/reasoning from the LLM if not skip_model_outputs: yield gr.ChatMessage(role="assistant", content=f"**{step_number}**", metadata={"status": "done"}) elif skip_model_outputs and hasattr(step_log, "model_output") and step_log.model_output is not None: model_output = step_log.model_output.strip() # Remove any trailing and extra backticks, handling multiple possible formats model_output = re.sub(r"```\s*", "```", model_output) # handles ``` model_output = re.sub(r"\s*```", "```", model_output) # handles ``` model_output = re.sub(r"```\s*\n\s*", "```", model_output) # handles ```\n model_output = model_output.strip() yield gr.ChatMessage(role="assistant", content=model_output, metadata={"status": "done"}) # For tool calls, create a parent message if hasattr(step_log, "tool_calls") and step_log.tool_calls is not None: first_tool_call = step_log.tool_calls[0] used_code = first_tool_call.name == "python_interpreter" # Tool call becomes the parent message with timing info # First we will handle arguments based on type args = first_tool_call.arguments if isinstance(args, dict): content = str(args.get("answer", str(args))) else: content = str(args).strip() if used_code: # Clean up the content by removing any end code tags content = re.sub(r"```.*?\n", "", content) # Remove existing code blocks content = re.sub(r"\s*\s*", "", content) # Remove end_code tags content = content.strip() if not content.startswith("```python"): content = f"```python\n{content}\n```" parent_message_tool = gr.ChatMessage( role="assistant", content=content, metadata={ "title": f"🛠️ Used tool {first_tool_call.name}", "status": "done", }, ) yield parent_message_tool # Display execution logs if they exist if hasattr(step_log, "observations") and ( step_log.observations is not None and step_log.observations.strip() ): # Only yield execution logs if there's actual content log_content = step_log.observations.strip() if log_content: log_content = re.sub(r"^Execution logs:\s*", "", log_content) yield gr.ChatMessage( role="assistant", content=f"```bash\n{log_content}\n", metadata={"title": "📝 Execution Logs", "status": "done"}, ) # Display any errors if hasattr(step_log, "error") and step_log.error is not None: yield gr.ChatMessage( role="assistant", content=str(step_log.error), metadata={"title": "💥 Error", "status": "done"}, ) # Update parent message metadata to done status without yielding a new message if getattr(step_log, "observations_images", []): for image in step_log.observations_images: path_image = AgentImage(image).to_string() yield gr.ChatMessage( role="assistant", content={"path": path_image, "mime_type": f"image/{path_image.split('.')[-1]}"}, metadata={"title": "🖼️ Output Image", "status": "done"}, ) # Handle standalone errors but not from tool calls if hasattr(step_log, "error") and step_log.error is not None: yield gr.ChatMessage( role="assistant", content=str(step_log.error), metadata={"title": "💥 Error", "status": "done"} ) yield gr.ChatMessage( role="assistant", content=get_step_footnote_content(step_log, step_number), metadata={"status": "done"} ) yield gr.ChatMessage(role="assistant", content="-----", metadata={"status": "done"}) elif isinstance(step_log, PlanningStep): yield gr.ChatMessage(role="assistant", content="**Planning step**", metadata={"status": "done"}) yield gr.ChatMessage(role="assistant", content=step_log.plan, metadata={"status": "done"}) yield gr.ChatMessage( role="assistant", content=get_step_footnote_content(step_log, "Planning step"), metadata={"status": "done"} ) yield gr.ChatMessage(role="assistant", content="-----", metadata={"status": "done"}) elif isinstance(step_log, FinalAnswerStep): final_answer = step_log.final_answer if isinstance(final_answer, AgentText): yield gr.ChatMessage( role="assistant", content=f"**Final answer:**\n{final_answer.to_string()}\n", metadata={"status": "done"}, ) elif isinstance(final_answer, AgentImage): yield gr.ChatMessage( role="assistant", content={"path": final_answer.to_string(), "mime_type": "image/png"}, metadata={"status": "done"}, ) elif isinstance(final_answer, AgentAudio): yield gr.ChatMessage( role="assistant", content={"path": final_answer.to_string(), "mime_type": "audio/wav"}, metadata={"status": "done"}, ) else: yield gr.ChatMessage( role="assistant", content=f"**Final answer:** {str(final_answer)}", metadata={"status": "done"} ) else: raise ValueError(f"Unsupported step type: {type(step_log)}") def stream_to_gradio( agent, task: str, task_images: list | None = None, reset_agent_memory: bool = False, additional_args: dict | None = None, ): """Runs an agent with the given task and streams the messages from the agent as gradio ChatMessages.""" total_input_tokens = 0 total_output_tokens = 0 if not _is_package_available("gradio"): raise ModuleNotFoundError( "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`" ) intermediate_text = "" for step_log in agent.run( task, images=task_images, stream=True, reset=reset_agent_memory, additional_args=additional_args ): # Track tokens if model provides them if getattr(agent.model, "last_input_token_count", None) is not None: total_input_tokens += agent.model.last_input_token_count total_output_tokens += agent.model.last_output_token_count if isinstance(step_log, (ActionStep, PlanningStep)): step_log.input_token_count = agent.model.last_input_token_count step_log.output_token_count = agent.model.last_output_token_count if isinstance(step_log, MemoryStep): intermediate_text = "" for message in pull_messages_from_step( step_log, # If we're streaming model outputs, no need to display them twice skip_model_outputs=getattr(agent, "stream_outputs", False), ): yield message elif isinstance(step_log, ChatMessageStreamDelta): intermediate_text += step_log.content or "" yield intermediate_text def extract_vehicle_info_as_string(adf_xml): root = ET.fromstring(adf_xml) # Find the vehicle element vehicle = root.find('.//vehicle') if vehicle is not None: year = vehicle.find('year').text if vehicle.find('year') is not None else "" make = vehicle.find('make').text if vehicle.find('make') is not None else "" model = vehicle.find('model').text if vehicle.find('model') is not None else "" vehicle_info = f"{year} {make} {model}".strip() # Extract first name first_name = "" name_element = root.find('.//name[@part="first"]') if name_element is not None: first_name = name_element.text.strip() if name_element.text else "" return first_name, vehicle_info class GradioUI: """A one-line interface to launch your agent in Gradio""" def __init__(self, agent: MultiStepAgent, file_upload_folder: str | None = None): if not _is_package_available("gradio"): raise ModuleNotFoundError( "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`" ) self.agent = agent self.file_upload_folder = Path(file_upload_folder) if file_upload_folder is not None else None self.name = getattr(agent, "name") or "OTTO: The Car Sales Agent" self.description = getattr(agent, "description", None) if self.file_upload_folder is not None: if not self.file_upload_folder.exists(): self.file_upload_folder.mkdir(parents=True, exist_ok=True) def interact_with_agent(self, prompt, messages, session_state, car_site, adf_lead): import gradio as gr self.agent.prompt_templates["system_prompt"] += f"\n\nWhen answering a customer's question about the dealership or other cars, use the following site to find the information:\n\nDealership Site: {car_site}\n\nWhen answering a customer's question about the specific car use the following ADF Lead:\n\nADF Lead: {adf_lead}" # Get the agent type from the template agent if "agent" not in session_state: session_state["agent"] = self.agent try: messages.append(gr.ChatMessage(role="user", content=prompt, metadata={"status": "done"})) yield messages for msg in stream_to_gradio(session_state["agent"], task=prompt, reset_agent_memory=False): if isinstance(msg, gr.ChatMessage): messages.append(msg) elif isinstance(msg, str): # Then it's only a completion delta try: if messages[-1].metadata["status"] == "pending": messages[-1].content = msg else: messages.append( gr.ChatMessage(role="assistant", content=msg, metadata={"status": "pending"}) ) except Exception as e: raise e yield messages yield messages except Exception as e: print(f"Error in interaction: {str(e)}") messages.append(gr.ChatMessage(role="assistant", content=f"Error: {str(e)}")) yield messages def upload_file(self, file, file_uploads_log, allowed_file_types=None): """ Handle file uploads, default allowed types are .pdf, .docx, and .txt """ import gradio as gr if file is None: return gr.Textbox(value="No file uploaded", visible=True), file_uploads_log if allowed_file_types is None: allowed_file_types = [".pdf", ".docx", ".txt"] file_ext = os.path.splitext(file.name)[1].lower() if file_ext not in allowed_file_types: return gr.Textbox("File type disallowed", visible=True), file_uploads_log # Sanitize file name original_name = os.path.basename(file.name) sanitized_name = re.sub( r"[^\w\-.]", "_", original_name ) # Replace any non-alphanumeric, non-dash, or non-dot characters with underscores # Save the uploaded file to the specified folder file_path = os.path.join(self.file_upload_folder, os.path.basename(sanitized_name)) shutil.copy(file.name, file_path) return gr.Textbox(f"File uploaded: {file_path}", visible=True), file_uploads_log + [file_path] def log_user_message(self, text_input, file_uploads_log): import gradio as gr return ( text_input + ( f"\nYou have been provided with these files, which might be helpful or not: {file_uploads_log}" if len(file_uploads_log) > 0 else "" ), "", gr.Button(interactive=False), ) def launch(self, share: bool = True, **kwargs): self.create_app().launch(debug=True, share=share, **kwargs) def create_app(self): import gradio as gr with gr.Blocks(theme="ocean", fill_height=True) as demo: # Add session state to store session-specific data session_state = gr.State({}) stored_messages = gr.State([]) file_uploads_log = gr.State([]) with gr.Sidebar(): gr.Markdown( f"# {self.name.replace('_', ' ')}" "\n> Test the OTTO Agent by asking it questions." + (f"\n\n**Agent description:**\n{self.description}" if self.description else "") ) with gr.Group(): gr.Markdown("**Your request**", container=True) text_input = gr.Textbox( lines=3, label="Chat Message", container=False, placeholder="Enter your prompt here and press Shift+Enter or press the button", ) submit_btn = gr.Button("Submit", variant="primary") with gr.Accordion("Dealership Info", open=False): car_site = gr.Textbox(label="Car Gurus Dealership Site", lines=2, value="https://www.cargurus.com/Cars/m-Ohio-Cars-sp458596", interactive=True) adf_lead = gr.Textbox(label="ADF Lead", lines=4, value="2025-05-12T13:59:3016f3114e-825f-4eb0-8165-ce43fe5143b62016ToyotaCorolla5YFBURHE4GP511115DPSuper White131024.09950TestLeadTest Lead123@gmail.com2582584568
19971Carsforsale.comCarsforsale.com866-388-9778114483Ohio Cars", interactive=False) # If an upload folder is provided, enable the upload feature if self.file_upload_folder is not None: upload_file = gr.File(label="Upload a file") upload_status = gr.Textbox(label="Upload Status", interactive=False, visible=False) upload_file.change( self.upload_file, [upload_file, file_uploads_log], [upload_status, file_uploads_log], ) first_name, vehicle_info = extract_vehicle_info_as_string(adf_lead.value) message = gr.ChatMessage(role="assistant", content=f"Hi {first_name}! The {vehicle_info} you're interested in is available at [OhioCars.com](https://www.ohiocars.com). Would you like to schedule a visit to check it out? We have appointment slots at 11 AM, 1 PM, or 3 PM. Which time works best for you?", metadata={"status": "done"}) # Main chat interface chatbot = gr.Chatbot( label="Agent", type="messages", value=[message], avatar_images=( None, "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/smolagents/mascot_smol.png", ), resizeable=True, scale=1, ) # Set up event handlers text_input.submit( self.log_user_message, [text_input, file_uploads_log], [stored_messages, text_input, submit_btn], ).then(self.interact_with_agent, [stored_messages, chatbot, session_state, car_site, adf_lead], [chatbot]).then( lambda: ( gr.Textbox( interactive=True, placeholder="Enter your prompt here and press Shift+Enter or the button" ), gr.Button(interactive=True), ), None, [text_input, submit_btn], ) submit_btn.click( self.log_user_message, [text_input, file_uploads_log], [stored_messages, text_input, submit_btn], ).then(self.interact_with_agent, [stored_messages, chatbot, session_state, car_site, adf_lead], [chatbot]).then( lambda: ( gr.Textbox( interactive=True, placeholder="Enter your prompt here and press Shift+Enter or the button" ), gr.Button(interactive=True), ), None, [text_input, submit_btn], ) return demo __all__ = ["stream_to_gradio", "GradioUI"]