app.py

import os
from typing import List, Optional
from pydantic import BaseModel, Field
import gradio as gr
from datasets import load_dataset
from huggingface_hub import InferenceClient
import black

# Initialize the inference client
HF_TOKEN = os.getenv("HF_TOKEN")
HF_API_URL = os.getenv("HF_API_URL", "Qwen/Qwen2.5-Coder-32B-Instruct")
client = InferenceClient(model=HF_API_URL, token=HF_TOKEN)

# Load questions from Hugging Face dataset
EXAM_MAX_QUESTIONS = int(os.getenv("EXAM_MAX_QUESTIONS", 1))
EXAM_DATASET_ID = "agents-course/smolagents-quiz-data"

# prep the dataset for the quiz
ds = load_dataset(EXAM_DATASET_ID, split="train", download_mode="force_redownload")
quiz_data = list(ds)
if EXAM_MAX_QUESTIONS:
    quiz_data = quiz_data[:EXAM_MAX_QUESTIONS]

# Check if dataset has image feature
HAS_IMAGE_FEATURE = "image" in ds.features


class CriterionFeedback(BaseModel):
    """Feedback for a single assessment criterion"""

    criterion: str = Field(..., description="The assessment criterion being evaluated")
    met: bool = Field(..., description="Whether the criterion was met")
    explanation: str = Field(
        ..., description="Detailed explanation of how well the criterion was met"
    )
    improvement_tips: Optional[str] = Field(
        None, description="Specific tips for improvement if needed"
    )


class CodeFeedback(BaseModel):
    """Structured feedback for code submission"""

    overall_feedback: str = Field(
        ..., description="Overall assessment of the code solution"
    )
    criteria_feedback: List[CriterionFeedback] = Field(
        ..., description="Detailed feedback for each assessment criterion"
    )


def format_python_code(code: str) -> str:
    """Format Python code using black."""
    try:
        return black.format_str(code, mode=black.Mode())
    except Exception as e:
        gr.Warning(f"Code formatting failed: {str(e)}")
        return code


EVALUATION_TEMPLATE = """Evaluate this Python code solution:

Challenge:
{challenge}

Reference Solution:
```python

{solution}

```

Student's Solution:

```python

{student_code}

```

Assessment Criteria:
{criteria}

Approach:
Be highly tollerent of differences in approach, as long as they meet Assessment Criteria.

Provide detailed feedback on how well each criterion was met."""


def check_code(
    user_code: str, solution: str, challenge: str, assessment_criteria: List[str]
) -> dict:
    """
    Use LLM to evaluate the user's code solution and provide structured feedback.
    """
    # Format both user code and solution
    formatted_user_code = format_python_code(user_code)
    formatted_solution = format_python_code(solution)

    # Format criteria as bullet points
    criteria_text = "\n".join(f"- {c}" for c in assessment_criteria)

    # Fill the template
    prompt = EVALUATION_TEMPLATE.format(
        challenge=challenge,
        solution=formatted_solution,
        student_code=formatted_user_code,
        criteria=criteria_text,
    )

    try:
        # Get structured feedback using response_format with schema from Pydantic model
        response = client.text_generation(
            prompt=prompt,
            grammar={
                "type": "json_object",
                "value": CodeFeedback.model_json_schema(),
            },
        )

        # Parse response into Pydantic model
        feedback = CodeFeedback.model_validate_json(response)

        # Format the feedback for display
        formatted_feedback = [
            f"### Overall Assessment\n{feedback.overall_feedback}\n\n"
        ]

        for cf in feedback.criteria_feedback:
            tip = cf.improvement_tips or ""
            tip_text = f"\n💡 Tip: {tip}" if tip else ""

            formatted_feedback.append(
                f"### {cf.criterion}\n"
                f"{'✅' if cf.met else '❌'} {cf.explanation}"
                f"{tip_text}\n"
            )

        return {"feedback": "\n".join(formatted_feedback)}

    except Exception as e:
        gr.Warning(f"Error generating feedback: {str(e)}")
        return {"feedback": "Unable to generate detailed feedback due to an error."}


def on_user_logged_in(token: gr.OAuthToken | None):
    """
    Handle user login state.
    On a valid token, hide the login button and reveal the Start button while keeping Next hidden.
    Also, clear the question text, code input, status, and image.
    """
    if token is not None:
        return (
            gr.update(visible=False),  # login_btn hidden
            gr.update(visible=True),  # start_btn shown
            gr.update(visible=False),  # next_btn hidden
            "",  # Clear question_text
            gr.update(value="", visible=False),  # Clear code_input
            "",  # Clear status_text
            gr.update(value="", visible=False),  # Clear question_image
        )
    else:
        return (
            gr.update(visible=True),  # login_btn visible
            gr.update(visible=False),  # start_btn hidden
            gr.update(visible=False),  # next_btn hidden
            "",
            gr.update(value="", visible=False),
            "",
            gr.update(value="", visible=False),
        )


def handle_quiz(question_idx, user_answers, submitted_code, is_start):
    """Handle quiz state and progression"""
    if is_start:
        question_idx = 0
    else:
        # If not the first question and there's a submission, store it
        if question_idx < len(quiz_data) and submitted_code.strip():
            current_q = quiz_data[question_idx]
            # Format the submitted code before checking
            formatted_code = format_python_code(submitted_code)
            feedback_dict = check_code(
                formatted_code,
                current_q["solution"],
                current_q["challenge"],
                current_q["assessment_criteria"],
            )
            user_answers.append(
                {
                    "challenge": current_q["challenge"],
                    "submitted_code": formatted_code,
                    "correct_solution": current_q["solution"],
                    "assessment_criteria": current_q["assessment_criteria"],
                    "feedback": feedback_dict["feedback"],
                }
            )
        question_idx += 1

    # If we've reached the end, show final results
    if question_idx >= len(quiz_data):
        results_text = """## Code Review Complete! 📚
This feedback should help you improve your skills. 

⛔️ The feedback uses Qwen/Qwen2.5-Coder-32B-Instruct to compare your response to a gold
standard solution. As we know, LLMs are not perfect. You should compare your work against
the assessment criteria if you doubt the feedback. 

Here's your detailed feedback:"""

        for idx, answer in enumerate(user_answers):
            # Format assessment criteria as bullet points
            criteria_bullets = "\n".join(
                f"- {c}" for c in answer["assessment_criteria"]
            )

            # Build the results text piece by piece
            results_text += (
                f"### Question {idx + 1}: {answer['challenge']}\n\n"
                "#### Your Solution:\n```python\n"
                f"{answer['submitted_code']}\n```\n\n"
                "#### Reference Solution:\n```python\n"
                f"{answer['correct_solution']}\n```\n\n"
                "#### Assessment Criteria:\n"
                f"{criteria_bullets}\n\n"
                "#### Feedback:\n"
                f"{answer['feedback']}\n\n"
                "---\n\n"
            )

        return (
            "",  # question_text cleared
            gr.update(value="", visible=False),  # hide code_input
            "Review your feedback below to improve your coding skills!",
            question_idx,  # updated question index
            user_answers,  # accumulated answers
            gr.update(visible=False),  # start_btn hidden
            gr.update(visible=False),  # next_btn hidden
            gr.update(value=results_text, visible=True),  # final_markdown
            gr.update(visible=False),  # question_image hidden
        )
    else:
        # Show the next question
        q = quiz_data[question_idx]
        # Format assessment criteria as bullet points
        criteria_bullets = "\n".join(f"- {c}" for c in q["assessment_criteria"])
        challenge_text = (
            f"## Question {question_idx + 1}\n\n"
            f"### Challenge:\n{q['challenge']}\n\n"
            "### Assessment Criteria:\n"
            f"{criteria_bullets}"
        )

        # Only show image if the feature exists and question has an image
        show_image = HAS_IMAGE_FEATURE and q.get("image") is not None
        image_update = gr.update(
            value=q.get("image") if show_image else None, visible=show_image
        )

        return (
            challenge_text,  # question_text
            gr.update(value=q["placeholder"], visible=True),  # code_input
            "Submit your solution and click 'Next' to continue.",
            question_idx,  # updated question_idx
            user_answers,  # user_answers
            gr.update(visible=False),  # start_btn hidden
            gr.update(visible=True),  # next_btn visible
            gr.update(visible=False),  # final_markdown hidden
            image_update,  # question_image
        )


with gr.Blocks() as demo:
    demo.title = f"Coding Quiz: {EXAM_DATASET_ID}"
    # State variables
    question_idx = gr.State(value=0)
    user_answers = gr.State(value=[])

    with gr.Row(variant="compact"):
        intro_text = """
## Welcome to the smolagents code reviewer

This application will review your smolagents code, and provide feedback on your solutions. This exercise is not reviewed or certified! It's about trying out smolagents for the first time.

ℹ️ Log in first, then click 'Start' to begin. Complete each coding challenge and click 'Next' to proceed. You'll get feedback on your solutions at the end."""
        intro_text = gr.Markdown(intro_text)
    with gr.Row(variant="panel"):
        with gr.Column():
            question_text = gr.Markdown("")
            question_image = gr.Image(
                label="Question Image",
                visible=True if HAS_IMAGE_FEATURE else False,
                type="pil",
            )  # Add image component
        with gr.Column():
            code_input = gr.Code(
                language="python", label="Your Solution", visible=False
            )

    with gr.Row(variant="compact"):
        status_text = gr.Markdown("")

    with gr.Row(variant="compact"):
        login_btn = gr.LoginButton()
        start_btn = gr.Button("Start")
        next_btn = gr.Button("Next ⏭️", visible=False)

    with gr.Row(variant="compact"):
        final_markdown = gr.Markdown("", visible=False)

    login_btn.click(
        fn=on_user_logged_in,
        inputs=None,
        outputs=[
            login_btn,
            start_btn,
            next_btn,
            question_text,
            code_input,
            status_text,
            question_image,
        ],
    )

    start_btn.click(
        fn=handle_quiz,
        inputs=[question_idx, user_answers, code_input, gr.State(True)],
        outputs=[
            question_text,  # Markdown with question text
            code_input,  # Code input field
            status_text,  # Status text (instructions/status messages)
            question_idx,  # Updated question index (state)
            user_answers,  # Updated user answers (state)
            start_btn,  # Update for start button (will be hidden)
            next_btn,  # Update for next button (shown for in-progress quiz)
            final_markdown,  # Final results markdown (hidden until quiz ends)
            question_image,  # Image update for the quiz question
        ],
    )

    next_btn.click(
        fn=handle_quiz,
        inputs=[question_idx, user_answers, code_input, gr.State(False)],
        outputs=[
            question_text,
            code_input,
            status_text,
            question_idx,
            user_answers,
            start_btn,
            next_btn,
            final_markdown,
            question_image,
        ],
    )

if __name__ == "__main__":
    demo.launch()