app.py

import argparse
import json
import os
import re
import threading
import mimetypes
import shutil
from concurrent.futures import ThreadPoolExecutor, as_completed
from datetime import datetime
from pathlib import Path
from typing import List, Optional

import datasets
import pandas as pd
from dotenv import load_dotenv
from huggingface_hub import login
import gradio as gr

from scripts.reformulator import prepare_response
from scripts.run_agents import (
    get_single_file_description,
    get_zip_description,
)
from scripts.text_inspector_tool import TextInspectorTool
from scripts.text_web_browser import (
    ArchiveSearchTool,
    FinderTool,
    FindNextTool,
    PageDownTool,
    PageUpTool,
    SearchInformationTool,
    SimpleTextBrowser,
    VisitTool,
)
from scripts.visual_qa import visualizer
from tqdm import tqdm

from smolagents import (
    # MANAGED_AGENT_PROMPT,
    CodeAgent,
    HfApiModel,
    LiteLLMModel,
    Model,
    ToolCallingAgent,
)
from smolagents.agent_types import AgentText, AgentImage, AgentAudio
from smolagents.gradio_ui import pull_messages_from_step, handle_agent_output_types

AUTHORIZED_IMPORTS = [
    "requests",
    "zipfile",
    "os",
    "pandas",
    "numpy",
    "sympy",
    "json",
    "bs4",
    "pubchempy",
    "xml",
    "yahoo_finance",
    "Bio",
    "sklearn",
    "scipy",
    "pydub",
    "io",
    "PIL",
    "chess",
    "PyPDF2",
    "pptx",
    "torch",
    "datetime",
    "fractions",
    "csv",
]
load_dotenv(override=True)
login(os.getenv("HF_TOKEN"))

append_answer_lock = threading.Lock()

SET = "validation"

custom_role_conversions = {"tool-call": "assistant", "tool-response": "user"}

# skip
_ = """
### LOAD EVALUATION DATASET

eval_ds = datasets.load_dataset("gaia-benchmark/GAIA", "2023_all")[SET]
eval_ds = eval_ds.rename_columns({"Question": "question", "Final answer": "true_answer", "Level": "task"})

def preprocess_file_paths(row):
    if len(row["file_name"]) > 0:
        row["file_name"] = f"data/gaia/{SET}/" + row["file_name"]
    return row

eval_ds = eval_ds.map(preprocess_file_paths)
eval_df = pd.DataFrame(eval_ds)
print("Loaded evaluation dataset:")
print(eval_df["task"].value_counts())
# """

user_agent = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/119.0.0.0 Safari/537.36 Edg/119.0.0.0"

BROWSER_CONFIG = {
    "viewport_size": 1024 * 5,
    "downloads_folder": "downloads_folder",
    "request_kwargs": {
        "headers": {"User-Agent": user_agent},
        "timeout": 300,
    },
    "serpapi_key": os.getenv("SERPAPI_API_KEY"),
}

os.makedirs(f"./{BROWSER_CONFIG['downloads_folder']}", exist_ok=True)

# --- 수정된 부분: OPENAI_API_BASE의 잘못된 엔드포인트 제거 ---
openai_api_base = os.getenv("OPENAI_API_BASE", "https://api.openai.com/v1")
# 만약 환경변수에 불필요한 엔드포인트가 포함되어 있다면 제거합니다.
if openai_api_base.endswith("/chat/completions"):
    openai_api_base = openai_api_base.rsplit("/chat/completions", 1)[0]

model = LiteLLMModel(
    os.getenv("MODEL_ID", "gpt-4o-mini"),
    custom_role_conversions=custom_role_conversions,
    api_base=openai_api_base,
    api_key=os.getenv("OPENAI_API_KEY"),
)
# --- 수정 끝 ---

text_limit = 20000
ti_tool = TextInspectorTool(model, text_limit)

browser = SimpleTextBrowser(**BROWSER_CONFIG)

WEB_TOOLS = [
    SearchInformationTool(browser),
    VisitTool(browser),
    PageUpTool(browser),
    PageDownTool(browser),
    FinderTool(browser),
    FindNextTool(browser),
    ArchiveSearchTool(browser),
    TextInspectorTool(model, text_limit),
]

agent = CodeAgent(
    model=model,
    tools=[visualizer] + WEB_TOOLS,
    max_steps=5,
    verbosity_level=2,
    additional_authorized_imports=AUTHORIZED_IMPORTS,
    planning_interval=4,
)

document_inspection_tool = TextInspectorTool(model, 20000)


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."""
    for step_log in agent.run(task, stream=True, reset=reset_agent_memory, additional_args=additional_args):
        for message in pull_messages_from_step(step_log):
            yield message

    final_answer = step_log  # Last log is the run's final_answer
    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, file_upload_folder: str | None = None):
        self.agent = agent
        self.file_upload_folder = file_upload_folder
        if self.file_upload_folder is not None:
            if not os.path.exists(file_upload_folder):
                os.mkdir(file_upload_folder)

    def interact_with_agent(self, prompt, messages):
        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 upload_file(
        self,
        file,
        file_uploads_log,
        allowed_file_types=[
            "application/pdf",
            "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
            "text/plain",
        ],
    ):
        """
        Handle file uploads, default allowed types are .pdf, .docx, and .txt
        """
        if file is None:
            return gr.Textbox("No file uploaded", visible=True), file_uploads_log

        try:
            mime_type, _ = mimetypes.guess_type(file.name)
        except Exception as e:
            return gr.Textbox(f"Error: {e}", visible=True), file_uploads_log

        if mime_type 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

        type_to_ext = {}
        for ext, t in mimetypes.types_map.items():
            if t not in type_to_ext:
                type_to_ext[t] = ext

        # Ensure the extension correlates to the mime type
        sanitized_name = sanitized_name.split(".")[:-1]
        sanitized_name.append("" + type_to_ext[mime_type])
        sanitized_name = "".join(sanitized_name)

        # 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):
        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 ""
            ),
            "",
        )

    def launch(self, **kwargs):
        with gr.Blocks(theme="ocean", fill_height=True) as demo:
            gr.Markdown("""# open Deep Research - free the AI agents!

OpenAI just published [Deep Research](https://openai.com/index/introducing-deep-research/), a very nice assistant that can perform deep searches on the web to answer user questions.

However, their agent has a huge downside: it's not open. So we've started a 24-hour rush to replicate and open-source it. Our resulting [open-Deep-Research agent](https://github.com/huggingface/smolagents/tree/main/examples/open_deep_research) took the #1 rank of any open submission on the GAIA leaderboard! ✨

You can try a simplified version below. 👇""")
            stored_messages = gr.State([])
            file_uploads_log = gr.State([])
            chatbot = gr.Chatbot(
                label="open-Deep-Research",
                type="messages",
                avatar_images=(
                    None,
                    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/smolagents/mascot_smol.png",
                ),
                resizeable=True,
                scale=1,
            )
            # 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],
                )
            text_input = gr.Textbox(lines=1, label="Your request")
            text_input.submit(
                self.log_user_message,
                [text_input, file_uploads_log],
                [stored_messages, text_input],
            ).then(self.interact_with_agent, [stored_messages, chatbot], [chatbot])

        demo.launch(debug=True, share=True, **kwargs)

GradioUI(agent).launch()