import json
import re
from transformers import AutoTokenizer, AutoModelForCausalLM

# Global variables for caching the model and tokenizer
tokenizer, model = None, None

def load_model():
    global tokenizer, model
    if tokenizer is None or model is None:
        # Use the DeepSeek instruct model for code evaluation.
        model_name = "deepseek-ai/deepseek-coder-1.3b-instruct"
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(model_name)
    return tokenizer, model

def extract_json(response_text):
    # Attempt to extract all JSON blocks (non-greedy, with DOTALL)
    matches = re.findall(r'\{.*?\}', response_text, re.DOTALL)
    for m in reversed(matches):
        try:
            temp = json.loads(m)
            if isinstance(temp, dict) and "stars" in temp and "feedback" in temp:
                return temp
        except Exception:
            continue
    return None

def evaluate_code(question, code):
    # Revised prompt that explicitly states the expected arithmetic operation for square.
    prompt = f"""You are an expert code evaluator.
Evaluate the following solution for the given problem.
The problem asks for a function that returns the square of a number.
A correct solution must multiply the number by itself (using x*x or x**2).
If the solution uses any other operation (such as addition), it is completely incorrect.
Rate the solution as follows:
  - 5 stars: Perfect solution; correct, efficient, and follows best practices.
  - 4 stars: Correct solution with minor issues.
  - 3 stars: Partially correct solution with noticeable issues.
  - 2 stars: Incorrect solution with some correct elements.
  - 1 star: Mostly incorrect solution.
  - 0 stars: Completely incorrect solution.
Respond with exactly one JSON object (with no extra text) that has exactly two keys:
  "stars": an integer between 0 and 5,
  "feedback": a concise string message explaining your rating.
The JSON must start with '{{' and end with '}}'.
Do not output any additional text.
Question: "{question}"
Solution: "{code}"
Your response:"""
    
    tokenizer, model = load_model()
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs,
        max_new_tokens=120,
        temperature=0.2,
        pad_token_id=tokenizer.eos_token_id,
        do_sample=True
    )
    response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    print("Raw model response:", response_text)  # Debug output

    result = extract_json(response_text)
    if result is None:
        result = {"stars": 0, "feedback": "Evaluation failed. Unable to extract valid JSON from AI response."}
    return result

# For direct command-line testing.
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))