app.py

import os
from groq import Groq
import gradio as gr 
import pytesseract
from sentence_transformers import SentenceTransformer, util
from PIL import Image
from typing import List
import torch
from transformers import BertTokenizer, BertModel, T5ForConditionalGeneration, T5Tokenizer
import torch.nn.functional as F

# Load pre-trained models
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
bert_model = BertModel.from_pretrained('bert-base-uncased')
sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')

# Load the pre-trained T5 model and tokenizer for grammar error detection
grammar_model = T5ForConditionalGeneration.from_pretrained('t5-base')
grammar_tokenizer = T5Tokenizer.from_pretrained('t5-base')

# Initialize Groq client
client = Groq(api_key=os.environ.get("GROQ_API_KEY"))

# System prompt for Groq
system_prompt = {
    "role": "system",
    "content": "You are a useful assistant. You reply with efficient answers."
}

async def chat_groq(message, history):
    messages = [system_prompt]
    for msg in history:
        messages.append({"role": "user", "content": str(msg[0])})
        messages.append({"role": "assistant", "content": str(msg[1])})
    messages.append({"role": "user", "content": str(message)})

    response_content = ''
    
    stream = client.chat.completions.create(
        model="llama3-70b-8192",
        messages=messages,
        max_tokens=1024,
        temperature=1.3,
        stream=True
    )

    for chunk in stream:
        content = chunk.choices[0].delta.content
        if content:
            response_content += chunk.choices[0].delta.content
        yield response_content

def extract_text_from_image(filepath: str, languages: List[str]):
    image = Image.open(filepath)
    lang_str = '+'.join(languages)
    return pytesseract.image_to_string(image=image, lang=lang_str)

def assign_badge(grade):
    if grade == 5:
        return "Gold Badge 🌟"
    elif grade == 4:
        return "Silver Badge 🥈"
    elif grade == 3:
        return "Bronze Badge 🥉"
    else:
        return "Keep Improving Badge 💪"

def detailed_feedback(similarity_score):
    if similarity_score >= 0.9:
        return {"Clarity": "Excellent", "Completeness": "Complete", "Accuracy": "Accurate"}
    elif similarity_score >= 0.8:
        return {"Clarity": "Good", "Completeness": "Almost Complete", "Accuracy": "Mostly Accurate"}
    elif similarity_score >= 0.7:
        return {"Clarity": "Fair", "Completeness": "Partial", "Accuracy": "Some Errors"}
    else:
        return {"Clarity": "Needs Improvement", "Completeness": "Incomplete", "Accuracy": "Inaccurate"}

def get_grade(similarity_score):
    if similarity_score >= 0.9:
        return 5
    elif similarity_score >= 0.8:
        return 4
    elif similarity_score >= 0.7:
        return 3
    elif similarity_score >= 0.6:
        return 2
    else:
        return 1

def get_bert_embedding(text):
    inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True)
    with torch.no_grad():
        outputs = bert_model(**inputs)
    embeddings = outputs.last_hidden_state.mean(dim=1)
    return embeddings

def calculate_cosine_similarity(embedding1, embedding2):
    similarity = F.cosine_similarity(embedding1, embedding2)
    return similarity.item()

def calculate_sentence_similarity(text1, text2):
    embedding1 = sentence_model.encode(text1, convert_to_tensor=True)
    embedding2 = sentence_model.encode(text2, convert_to_tensor=True)
    return util.pytorch_cos_sim(embedding1, embedding2).item()

# Grammar detection and penalization using T5 model
def detect_grammar_errors(text):
    input_text = f"grammar: {text}"
    inputs = grammar_tokenizer.encode(input_text, return_tensors='pt', max_length=512, truncation=True)
    outputs = grammar_model.generate(inputs, max_length=512, num_beams=4, early_stopping=True)
    grammar_analysis = grammar_tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Count the number of errors based on specific indicators (customize based on analysis)
    error_count = grammar_analysis.count('error')  # Use your own criteria
    return error_count

def penalize_for_grammar(student_answer):
    grammar_errors = detect_grammar_errors(student_answer)
    
    # Apply a penalty based on the number of grammar errors (max 50% penalty)
    penalty = max(0.5, 1 - 0.05 * grammar_errors)
    return penalty

def compare_answers(student_answer, teacher_answer):
    bert_similarity = calculate_cosine_similarity(get_bert_embedding(student_answer), get_bert_embedding(teacher_answer))
    sentence_similarity = calculate_sentence_similarity(student_answer, teacher_answer)
    
    # Use a higher weight for BERT similarity
    semantic_similarity = (0.75 * bert_similarity + 0.25 * sentence_similarity)
    
    # Apply grammar penalty
    grammar_penalty = penalize_for_grammar(student_answer)
    final_similarity = semantic_similarity * grammar_penalty
    
    return final_similarity

def extract_keywords(text):
    return set(text.lower().split())

def check_keywords(student_answer, model_answer):
    student_keywords = extract_keywords(student_answer)
    teacher_keywords = extract_keywords(model_answer)
    keyword_overlap = len(student_keywords.intersection(teacher_keywords))
    return keyword_overlap / (len(teacher_keywords) if len(teacher_keywords) > 0 else 1)

def evaluate_answer(image, languages, model_answer):
    student_answer = extract_text_from_image(image, languages)
    
    # Calculate semantic similarity
    semantic_similarity = compare_answers(student_answer, model_answer)
    
    # Calculate keyword similarity
    keyword_similarity = check_keywords(student_answer, model_answer)
    
    # Adjust the weight of keyword similarity
    combined_similarity = (0.9 * semantic_similarity + 0.1 * keyword_similarity)
    grade = get_grade(combined_similarity)
    feedback = f"Student's answer: {student_answer}\nTeacher's answer: {model_answer}"
    badge = assign_badge(grade)
    detailed_feedback_msg = detailed_feedback(combined_similarity)
    prompt = f"The student got grade: {grade} when the student's answer is: {student_answer} and the teacher's answer is: {model_answer}. Justify the grade given to the student."
    return grade, combined_similarity * 100, feedback, badge, detailed_feedback_msg, prompt

async def gradio_interface(image, languages: List[str], model_answer="The process of photosynthesis helps plants produce glucose using sunlight.", prompt="", history=[]):
    grade, similarity_score, feedback, badge, detailed_feedback_msg, prompt = evaluate_answer(image, languages, model_answer)
    response = ""
    async for result in chat_groq(prompt, history):
        response = result
    return grade, similarity_score, feedback, badge, detailed_feedback_msg, response

language_choices = pytesseract.get_languages()

interface = gr.Interface(
    fn=gradio_interface,
    inputs=[
        gr.Image(type="filepath", label="Input"), 
        gr.CheckboxGroup(language_choices, type="value", value=['eng'], label='Language'),
        gr.Textbox(lines=2, placeholder="Enter your model answer here", label="Model Answer"),
        gr.Textbox(lines=2, placeholder="Enter your prompt here", label="Prompt")
    ],
    outputs=[
        gr.Text(label="Grade"), 
        gr.Number(label="Similarity Score (%)"), 
        gr.Text(label="Feedback"), 
        gr.Text(label="Badge"), 
        gr.JSON(label="Detailed Feedback"), 
        gr.Text(label="Generated Response")
    ],
    title="Enhanced Automated Grading System",
    description="Upload an image of your answer sheet to get a grade from 1 to 5, similarity score, visual feedback, badge, and detailed feedback based on the model answer.",
    live=True
)

if __name__ == "__main__":
    interface.queue()
    interface.launch()