app.py

import os
import json
import re
import gradio as gr
from mistralai.client import Mistral
from difflib import SequenceMatcher

api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)
MODEL_SMALL = "mistral-small-latest"
MODEL_LARGE = "mistral-large-latest"

def clean_text(text, mode="normal"):
    if not text:
        return ""
    text = re.sub(r'[أإآ]', 'ا', text)
    if mode != "حرفي":
        text = re.sub(r'ة\b', 'ه', text)
        text = re.sub(r'ى\b', 'ي', text)
    fillers = ['آآآ', 'أاا', 'ممم', 'إممم', 'ااا', 'ءءء']
    for filler in fillers:
        text = text.replace(filler, ' ')
    text = text.replace("\n", " ")
    return " ".join(text.split())

def get_model(mode):
    if mode == "فهمي":
        return MODEL_LARGE
    elif mode == "هجين":
        return MODEL_SMALL
    elif mode == "حرفي":
        return MODEL_SMALL
    else:
        return None

# def literal_score(text1, text2):
#     return SequenceMatcher(None, text1, text2).ratio() * 100

def get_grade(score):
    if score >= 90:
        return "ممتاز"
    elif score >= 80:
        return "جدا جيد"
    elif score >= 70:
        return "جيد"
    elif score >= 60:
        return "مقبول"
    else:
        return "يرجى الاعادة"

def build_prompt(text1, text2, mode):
    # تعليمات مخصصة لكل نوع تقييم
    mode_instructions = {
        "حرفي": """
- Focus: Absolute Word-for-word accuracy (Memorization).
- Policy: Every word must match. Ignore minor diacritics (Tashkeel) or punctuation.
- Deductions: -10 points for any missing, added, or swapped word. 
- Strictness: If the student changes the vocabulary even if meaning is similar, they MUST lose points.
""",
        "فهمي": """
- Focus: Semantic meaning and core concepts. 
- Policy: Accept paraphrasing and synonyms.
- Strictness: If the topic is different (e.g. Technology vs Reading), the score MUST be 0.
- Deductions: -25 points for each major missing idea.
""",
        "هجين": """
- Focus: Balance between literal words and overall meaning.
- Policy: Small word changes are okay, but key terms must remain.
- Strictness: If the topic is unrelated, the score MUST be 0.
- Deductions: -15 points for missing key terms, -10 for weak paraphrasing.
"""
    }

    return f"""
You are a **ruthless and strict** Arabic language examiner. Your goal is to grade a student's recitation (Tasmée) against an original reference.

### Context:
- **Original Text (The Reference):** {text1}
- **Student's Text (The Attempt):** {text2}
- **Evaluation Mode:** {mode}

### Grading Rubric:
1. **Topic Check:** If the student talks about a COMPLETELY different topic, assign a score of 0 and grade "يرجى الاعادة".
2. **Rules for "{mode}":**
{mode_instructions[mode]}
3. **Hallucination:** If the student adds info not in the reference, deduct 20 points.
4. **Direct Feedback:** The "reason" MUST be in Arabic, addressed DIRECTLY to the student (e.g., "لقد نسيت ذكر...", "أحسنت ولكن...").

### Output:
Return ONLY a valid JSON. No markdown, no comments.
{{
  "score": <integer 0-100>,
  "grade": "ممتاز | جدا جيد | جيد | مقبول | يرجى الاعادة",
  "reason": "<Direct Arabic feedback to the student>"
}}
"""


def extract_json(text):
    # إزالة markdown
    text = re.sub(r"```json", "", text)
    text = re.sub(r"```", "", text)
    text = text.strip()

    # محاولة إصلاح الأسطر داخل strings
    def fix_multiline_strings(s):
        in_string = False
        result = []
        for char in s:
            if char == '"' and (len(result) == 0 or result[-1] != '\\'):
                in_string = not in_string
            if char == '\n' and in_string:
                result.append('\\n')  # تحويلها إلى escape
            else:
                result.append(char)
        return ''.join(result)

    text = fix_multiline_strings(text)
    return text


def call_mistral(prompt, model):
    try:
        response = client.chat.complete(
            model=model,
            messages=[{"role": "user", "content": prompt}],
            temperature=0
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        return f"ERROR: {e}"

def evaluate_similarity(text1, text2, mode):
    clean_ocr = clean_text(text1, mode)
    clean_stu = clean_text(text2, mode)

    if not clean_ocr or not clean_stu:
        return json.dumps({"error": "يرجى إدخال النصوص"}, ensure_ascii=False)

    # # --- الوضع الحرفي ---
    # if mode == "حرفي":
    #     score = literal_score(clean_ocr, clean_stu)
    #     grade = get_grade(score)
    #     return json.dumps({
    #         "evaluation_mode": mode,
    #         "model_used": "algorithmic",
    #         "details": {
    #             "score": round(score, 2),
    #             "grade": grade,
    #             "reason": "تم التقييم بناءً على التطابق الحرفي بدون AI"
    #         }
    #     }, ensure_ascii=False, indent=2)

    # --- وضع AI ---
    model = get_model(mode)
    prompt = build_prompt(clean_ocr, clean_stu, mode)
    content = call_mistral(prompt, model)

    clean_content = extract_json(content)

    try:
        data = json.loads(clean_content)
    except:
        # محاولة إنقاذ JSON
        match = re.search(r'\{.*\}', clean_content, re.DOTALL)
        if match:
            try:
                data = json.loads(match.group())
            except:
                return json.dumps({
                    "error": "فشل تحليل JSON",
                    "raw_output": content
                }, ensure_ascii=False, indent=2)
        else:
            return json.dumps({
                "error": "المودل لم يرجع JSON قابل للاستخراج",
                "raw_output": content
            }, ensure_ascii=False, indent=2)

    return json.dumps({
        "evaluation_mode": mode,
        "model_used": model,
        "details": data
    }, ensure_ascii=False, indent=2)

# --- واجهة Gradio ---
demo = gr.Interface(
    fn=evaluate_similarity,
    inputs=[
        gr.Textbox(lines=8, label="النص الأصلي"),
        gr.Textbox(lines=8, label="نص الطالب"),
        gr.Radio(["حرفي", "فهمي", "هجين"], value="فهمي", label="نوع التقييم")
    ],
    outputs=gr.Code(label="النتيجة", language="json"),
    title="🎓 مقيم التسميع الذكي باستخدام Mistral",
    description="قيم نص الطالب مقابل النص الأصلي بطريقة حرفية أو فهمية أو هجين."
)

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