import json
import re
from transformers import AutoTokenizer, AutoModelForCausalLM

def load_model():
    # Using a public model for code evaluation.
    model_name = "Salesforce/codegen-350M-mono"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(model_name)
    return tokenizer, model

def evaluate_code(question, code):
    # Refined prompt to enforce JSON-only output.
    prompt = f"""You are an expert code evaluator.
Evaluate the user's solution to the following problem.
Return ONLY a JSON object with two keys:
- "stars": an integer between 0 and 5 (0 means completely incorrect, 5 means excellent).
- "feedback": a concise message.
Do not include any additional text.
Problem: "{question}"
Solution: "{code}"
"""
    # Load model and tokenizer.
    tokenizer, model = load_model()
    # Generate a response with reduced max tokens and a lower temperature for determinism.
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,
        temperature=0.2,
        pad_token_id=tokenizer.eos_token_id
    )
    response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Attempt to extract the JSON object from the response.
    match = re.search(r'\{.*\}', response_text)
    if match:
        json_text = match.group(0)
        try:
            result = json.loads(json_text)
        except Exception as e:
            result = {"stars": 0, "feedback": "Evaluation failed. Unable to parse AI response."}
    else:
        result = {"stars": 0, "feedback": "Evaluation failed. Unable to extract JSON from AI response."}
    
    return result

# For direct testing from the command line.
if __name__ == "__main__":
    import sys
    if len(sys.argv) < 3:
        print(json.dumps({"error": "Please provide a question and code as arguments"}))
        sys.exit(1)
    question = sys.argv[1]
    code = sys.argv[2]
    result = evaluate_code(question, code)
    print(json.dumps(result))