gradio_ui.py

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
from system_prompt import get_system_prompt

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"""<span style="color: #bbbbc2; font-size: 12px;">{step_footnote}</span> """
    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 <end_code> and extra backticks, handling multiple possible formats
            model_output = re.sub(r"```\s*<end_code>", "```", model_output)  # handles ```<end_code>
            model_output = re.sub(r"<end_code>\s*```", "```", model_output)  # handles <end_code>```
            model_output = re.sub(r"```\s*\n\s*<end_code>", "```", model_output)  # handles ```\n<end_code>
            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*<end_code>\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_gurus_site, car_site, adf_lead, dealership_phone):
        import gradio as gr
        self.agent.prompt_templates["system_prompt"] = self.agent.prompt_templates["system_prompt"] + get_system_prompt(car_gurus_site, car_site, adf_lead, dealership_phone)
        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_gurus_site = gr.Textbox(label="Car Gurus Dealership Site", lines=2, value="https://www.cargurus.com/Cars/m-Ohio-Cars-sp458596", interactive=True)
                        car_site = gr.Textbox(label="Dealership Site", lines=2, value="https://www.Ohiocars.com", interactive=True)
                        adf_lead = gr.Textbox(label="ADF Lead", lines=4, value="<?xml version=\"1.0\"?><?ADF version=\"1.0\"?><adf><prospect><requestdate>2025-05-12T13:59:30</requestdate><vehicle status=\"used\"><id source=\"CarsForSale.com\">16f3114e-825f-4eb0-8165-ce43fe5143b6</id><year>2016</year><make>Toyota</make><model>Corolla</model><vin>5YFBURHE4GP511115</vin><stock></stock><comments>DP</comments><colorcombination><exteriorcolor>Super White</exteriorcolor></colorcombination><miles>131024.0</miles><price type=\"asking\">9950</price></vehicle><customer><contact><name part=\"first\">Test</name><name part=\"last\">Lead</name><name part=\"full\">Test Lead</name><email>123@gmail.com</email><phone>2582584568</phone><address><city></city><state></state><postalcode></postalcode></address></contact><comments><![CDATA[I'm interested and want to know more about the 2016 Toyota Corolla S Plus you have listed for $9,950 on Cars For Sale.]]></comments><timeframe><description></description></timeframe></customer><provider><id>19971</id><name part=\"full\">Carsforsale.com</name><service>Carsforsale.com</service><phone>866-388-9778</phone></provider><vendor><id>114483</id><vendorname>Ohio Cars</vendorname></vendor></prospect></adf>", interactive=False)
                        dealership_phone = gr.Textbox(label="Dealership Phone", lines=2, value="(513) 800-0805", interactive=True)
                # 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_gurus_site, car_site, adf_lead, dealership_phone], [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_gurus_site, car_site, adf_lead, dealership_phone], [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"]