app.py

import os
import logging
import openai
from openai import OpenAI
import gradio as gr
import requests
import pandas as pd
from smolagents import CodeAgent, DuckDuckGoSearchTool, tool

# --- Setup logging ---
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
logger = logging.getLogger(__name__)

# --- Constants ---
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
MAX_PROMPT_LENGTH = 15000  # characters, naive cap to avoid token overflow

# --- Configure OpenAI SDK & Client ---
openai_api_key = os.getenv("OPENAI_API_KEY")
if not openai_api_key:
    raise RuntimeError("Environment variable OPENAI_API_KEY is required.")
openai.api_key = openai_api_key
client = OpenAI()
OPENAI_MODEL_ID = os.getenv("OPENAI_MODEL_ID", "gpt-4.1")

# --- Adapter so CodeAgent can call the OpenAI client correctly ---
class OpenAIModelWrapper:
    def __init__(self, model_id: str, client: OpenAI):
        self.model_id = model_id
        self.client = client

    def __call__(self, prompt: str, **kwargs) -> str:
        try:
            resp = self.client.responses.create(
                model=self.model_id,
                input=prompt
            )
            return getattr(resp, "output_text", str(resp))
        except Exception as e:
            logger.exception("OpenAI inference error")
            return f"AGENT ERROR (inference): {e}"

# --- Tool Definitions ---

@tool
def summarize_query(query: str) -> str:
    """
    Reframes an unclear search query to improve relevance.

    Args:
        query (str): The original search query needing refinement.

    Returns:
        str: A concise, improved version of the query.
    """
    return f"Summarize and reframe: {query}"

@tool
def wikipedia_search(page: str) -> str:
    """
    Fetches the summary extract of an English Wikipedia page via the REST API.

    Args:
        page (str): Title of the Wikipedia page (e.g. 'Mercedes_Sosa_discography').

    Returns:
        str: The page's summary (or an error message).
    """
    try:
        url = f"https://en.wikipedia.org/api/rest_v1/page/summary/{page}"
        resp = requests.get(url, timeout=10)
        resp.raise_for_status()
        extract = resp.json().get("extract", "")
        if not extract:
            logger.warning("Wikipedia returned empty extract for %s", page)
        return extract
    except Exception as e:
        logger.exception("Wikipedia lookup failed")
        return f"Wikipedia error: {e}"

# Instantiate tools
search_tool    = DuckDuckGoSearchTool()
wiki_tool      = wikipedia_search
summarize_tool = summarize_query

# --- ReACT Prompt ---

instruction_prompt = """
You are a ReACT agent with three tools: 
 • DuckDuckGoSearchTool(query: str)
 • wikipedia_search(page: str)
 • summarize_query(query: str)

Internally, for each question:
1. Thought: decide which tool to call.
2. Action: call the chosen tool.
3. Observation: record the result.
4. If empty/irrelevant:
   Thought: retry with summarize_query + DuckDuckGoSearchTool.
   Record new Observation.
5. Thought: integrate observations.

Finally, output your answer with the following template: FINAL ANSWER: [YOUR FINAL ANSWER]. YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list of numbers and/or strings. If you are asked for a number, don't use comma to write your number neither use units such as $ or percent sign unless specified otherwise. If you are asked for a string, don't use articles, neither abbreviations (e.g. for cities), and write the digits in plain text unless specified otherwise. If you are asked for a comma separated list, apply the above rules depending of whether the element to be put in the list is a number or a string.

"""

# --- Build CodeAgent ---

llm_wrapper = OpenAIModelWrapper(model_id=OPENAI_MODEL_ID, client=client)
smart_agent = CodeAgent(
    tools=[search_tool, wiki_tool, summarize_tool],
    model=llm_wrapper
)

# --- BasicAgent for Gradio ---

class BasicAgent:
    def __init__(self):
        logger.info("Initialized SmolAgent (GPT-4.1) with ReACT & tools.")

    def __call__(self, question: str) -> str:
        # Validate question
        if not question or len(question.strip()) == 0:
            return "AGENT ERROR: Empty question."
        # Build and truncate prompt
        prompt = instruction_prompt.strip() + "\n\nQUESTION: " + question.strip()
        if len(prompt) > MAX_PROMPT_LENGTH:
            prompt = prompt[:MAX_PROMPT_LENGTH]  # naive trim
            logger.warning("Prompt truncated to %d chars", MAX_PROMPT_LENGTH)
        try:
            return smart_agent.run(prompt)
        except Exception as e:
            logger.exception("Agent run error")
            return f"AGENT ERROR (run): {e}"

# --- Submission logic ---

def run_and_submit_all(profile: gr.OAuthProfile | None):
    if not profile:
        return "Please log in to Hugging Face.", None

    username = profile.username
    space_id = os.getenv("SPACE_ID", "")
    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
    agent = BasicAgent()

    # Fetch questions
    try:
        resp = requests.get(f"{DEFAULT_API_URL}/questions", timeout=15)
        resp.raise_for_status()
        questions = resp.json()
        if not isinstance(questions, list):
            raise ValueError("Invalid questions format")
        logger.info("Fetched %d questions", len(questions))
    except Exception as e:
        logger.exception("Failed to fetch questions")
        return f"Error fetching questions: {e}", None

    # Run agent
    logs, payload, skipped = [], [], 0
    for item in questions:
        tid = item.get("task_id")
        q   = item.get("question")
        if not tid or not q:
            skipped += 1
            continue
        ans = agent(q)
        logs.append({"Task ID": tid, "Question": q, "Submitted Answer": ans})
        payload.append({"task_id": tid, "submitted_answer": ans})
    if skipped:
        logger.warning("Skipped %d malformed items", skipped)

    if not payload:
        return "Agent did not produce any answers.", pd.DataFrame(logs)

    # Submit answers
    submission = {"username": username, "agent_code": agent_code, "answers": payload}
    try:
        post = requests.post(f"{DEFAULT_API_URL}/submit", json=submission, timeout=60)
        post.raise_for_status()
        res = post.json()
        status = (
            f"Submission Successful!\n"
            f"User: {res.get('username')}\n"
            f"Score: {res.get('score','N/A')}% "
            f"({res.get('correct_count','?')}/{res.get('total_attempted','?')})\n"
            f"Message: {res.get('message','')}"
        )
        return status, pd.DataFrame(logs)
    except Exception as e:
        logger.exception("Submission failed")
        return f"Submission Failed: {e}", pd.DataFrame(logs)

# --- Gradio UI ---

with gr.Blocks() as demo:
    gr.Markdown("# SmolAgent GAIA Runner 🚀")
    gr.Markdown("""
**Instructions:**
1. Clone this space.
2. Add `OPENAI_API_KEY` and optionally `OPENAI_MODEL_ID` in Settings → Secrets.
3. Log in to Hugging Face.
4. Click **Run Evaluation & Submit All Answers**.
""")
    gr.LoginButton()
    run_btn = gr.Button("Run Evaluation & Submit All Answers")
    status_out = gr.Textbox(label="Status", lines=5, interactive=False)
    table_out  = gr.DataFrame(label="Questions & Answers", wrap=True)
    run_btn.click(fn=run_and_submit_all, outputs=[status_out, table_out])

if __name__ == "__main__":
    demo.launch(debug=True, share=False)