#!/usr/bin/env python # coding=utf-8 # Copyright 2024 The HuggingFace Inc. team. All rights reserved. # import os import re from typing import Optional from smolagents.agent_types import AgentAudio, AgentImage, AgentText, handle_agent_output_types from smolagents.agents import ActionStep, MultiStepAgent from smolagents.memory import MemoryStep from smolagents.utils import _is_package_available def pull_messages_from_step( step_log: MemoryStep, ): """Extract ChatMessage objects from agent steps with proper nesting""" 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 "" yield gr.ChatMessage(role="assistant", content=f"**{step_number}**") # First yield the thought/reasoning from the LLM if hasattr(step_log, "model_output") and step_log.model_output is not None: # Clean up the LLM output 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) model_output = re.sub(r"\s*```", "```", model_output) model_output = re.sub(r"```\s*\n\s*", "```", model_output) model_output = model_output.strip() yield gr.ChatMessage(role="assistant", content=model_output) # 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" parent_id = f"call_{len(step_log.tool_calls)}" # Tool call becomes the parent message with timing info args = first_tool_call.arguments if isinstance(args, dict): content = str(args.get("answer", str(args))) else: content = str(args).strip() if used_code: content = re.sub(r"```.*?\n", "", content) content = re.sub(r"\s*\s*", "", content) 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}", "id": parent_id, "status": "pending", }, ) yield parent_message_tool # Nesting execution logs under the tool call if they exist if hasattr(step_log, "observations") and ( step_log.observations is not None and step_log.observations.strip() ): 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"{log_content}", metadata={"title": "📝 Execution Logs", "parent_id": parent_id, "status": "done"}, ) # Nesting any errors under the tool call 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", "parent_id": parent_id, "status": "done"}, ) # Update parent message metadata to done status without yielding a new message parent_message_tool.metadata["status"] = "done" # Handle standalone errors but not from tool calls elif hasattr(step_log, "error") and step_log.error is not None: yield gr.ChatMessage(role="assistant", content=str(step_log.error), metadata={"title": "💥 Error"}) # Calculate duration and token information step_footnote = f"{step_number}" 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 = f"""{step_footnote} """ yield gr.ChatMessage(role="assistant", content=f"{step_footnote}") yield gr.ChatMessage(role="assistant", content="-----") # Handle screenshot if present in observations if hasattr(step_log, "observations") and step_log.observations: for line in step_log.observations.split("\n"): if line.startswith("Screenshot saved at:"): screenshot_path = line.replace("Screenshot saved at: ", "").strip() yield gr.ChatMessage( role="assistant", content={"path": screenshot_path, "mime_type": "image/png"}, metadata={"title": "📸 Screenshot"} ) def stream_to_gradio( agent, task: str, reset_agent_memory: bool = False, additional_args: Optional[dict] = None, ): """Runs an agent with the given task and streams the messages from the agent as gradio ChatMessages.""" if not _is_package_available("gradio"): raise ModuleNotFoundError( "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`" ) import gradio as gr total_input_tokens = 0 total_output_tokens = 0 for step_log in agent.run(task, stream=True, reset=reset_agent_memory, additional_args=additional_args): if hasattr(agent.model, "last_input_token_count"): total_input_tokens += agent.model.last_input_token_count total_output_tokens += agent.model.last_output_token_count if isinstance(step_log, ActionStep): step_log.input_token_count = agent.model.last_input_token_count step_log.output_token_count = agent.model.last_output_token_count for message in pull_messages_from_step( step_log, ): yield message final_answer = step_log final_answer = handle_agent_output_types(final_answer) if isinstance(final_answer, AgentText): yield gr.ChatMessage( role="assistant", content=f"**Final answer:**\n{final_answer.to_string()}\n", ) elif isinstance(final_answer, AgentImage): yield gr.ChatMessage( role="assistant", content={"path": final_answer.to_string(), "mime_type": "image/png"}, ) elif isinstance(final_answer, AgentAudio): yield gr.ChatMessage( role="assistant", content={"path": final_answer.to_string(), "mime_type": "audio/wav"}, ) else: yield gr.ChatMessage(role="assistant", content=f"**Final answer:** {str(final_answer)}") class GradioUI: """A one-line interface to launch your agent in Gradio""" def __init__(self, agent: MultiStepAgent): if not _is_package_available("gradio"): raise ModuleNotFoundError( "Please install 'gradio' extra to use the GradioUI: `pip install 'smolagents[gradio]'`" ) self.agent = agent def interact_with_agent(self, prompt, messages): import gradio as gr messages.append(gr.ChatMessage(role="user", content=prompt)) yield messages for msg in stream_to_gradio(self.agent, task=prompt, reset_agent_memory=False): messages.append(msg) yield messages yield messages def launch(self, **kwargs): import gradio as gr with gr.Blocks(fill_height=True) as demo: stored_messages = gr.State([]) chatbot = gr.Chatbot( label="Web Navigation Agent", type="messages", avatar_images=( None, "./icon.png" ), scale=1, ) text_input = gr.Textbox(lines=1, label="Enter URL and request (e.g., navigate to https://www.google.com and search for gemini)") text_input.submit( self.interact_with_agent, [text_input, stored_messages], [chatbot] ) demo.launch(debug=True, **kwargs) __all__ = ["stream_to_gradio", "GradioUI"]