functionbloom.py

from typing import Optional, Dict
import streamlit as st
import requests
import json
import fitz  # PyMuPDF
from fpdf import FPDF
import os
import tempfile
from dotenv import load_dotenv
import torch
from transformers import DistilBertForSequenceClassification, DistilBertTokenizer
from torch.nn.functional import softmax
from doctr.models import ocr_predictor
from doctr.io import DocumentFile
import tempfile

load_dotenv()

model = DistilBertForSequenceClassification.from_pretrained('./fine_tuned_distilbert')
tokenizer = DistilBertTokenizer.from_pretrained('./fine_tuned_distilbert')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
mapping = {"Remembering": 0, "Understanding": 1, "Applying": 2, "Analyzing": 3, "Evaluating": 4, "Creating": 5}
reverse_mapping = {v: k for k, v in mapping.items()}
modelocr = ocr_predictor(det_arch='db_resnet50', reco_arch='crnn_vgg16_bn', pretrained=True)

def save_uploaded_file(uploaded_file):
    if uploaded_file is not None:
        file_extension = uploaded_file.name.split('.')[-1].lower()
        temp_file = tempfile.NamedTemporaryFile(delete=False, suffix = f'.{file_extension}')
        temp_file.write(uploaded_file.getvalue())
        temp_file.close()
        return temp_file.name
    return None

# Previous functions from Question Generator
def get_pdf_path(pdf_source=None, uploaded_file=None):
    try:
        # If a file is uploaded locally
        if uploaded_file is not None:
            # Create a temporary file to save the uploaded PDF
            temp_dir = tempfile.mkdtemp()
            pdf_path = os.path.join(temp_dir, uploaded_file.name)
            
            # Save the uploaded file
            with open(pdf_path, "wb") as pdf_file:
                pdf_file.write(uploaded_file.getvalue())
            return pdf_path

        # If a URL is provided
        if pdf_source:
            response = requests.get(pdf_source, timeout=30)
            response.raise_for_status()
            
            # Create a temporary file
            temp_dir = tempfile.mkdtemp()
            pdf_path = os.path.join(temp_dir, "downloaded.pdf")
            
            with open(pdf_path, "wb") as pdf_file:
                pdf_file.write(response.content)
            return pdf_path

        # If no source is provided
        st.error("No PDF source provided.")
        return None
    except Exception as e:
        st.error(f"Error getting PDF: {e}")
        return None
    
    
def extract_text_pymupdf(pdf_path):
    try:
        doc = fitz.open(pdf_path)
        pages_content = []
        for page_num in range(len(doc)):
            page = doc[page_num]
            pages_content.append(page.get_text())
        doc.close()
        return " ".join(pages_content)  # Join all pages into one large context string
    except Exception as e:
        st.error(f"Error extracting text from PDF: {e}")
        return ""
    

def get_bloom_taxonomy_scores(question: str) -> Dict[str, float]:
    # Default scores in case of API failure
    default_scores = {
        "Remembering": 0.2,
        "Understanding": 0.2,
        "Applying": 0.15,
        "Analyzing": 0.15,
        "Evaluating": 0.15,
        "Creating": 0.15
    }
    
    try:
        scores = predict_with_loaded_model(question)
        for key, value in scores.items():
            if not (0 <= value <= 1):
                st.warning(f"Invalid score value for {key}. Using default scores.")
                return default_scores           
        return scores

    except Exception as e:
        st.warning(f"Unexpected error: {e}. Using default scores.")
        return default_scores
    
    
def generate_ai_response(api_key, assistant_context, user_query, role_description, response_instructions, bloom_taxonomy_weights, num_questions, question_length, include_numericals, user_input):
    try:
        url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash-latest:generateContent?key={api_key}"
        
        # Define length guidelines
        length_guidelines = {
            "Short": "Keep questions concise, around 10-15 words each.",
            "Medium": "Create moderately detailed questions, around 20-25 words each.",
            "Long": "Generate detailed, comprehensive questions, around 30-40 words each that may include multiple parts."
        }
        
        prompt = f"""
        You are a highly knowledgeable assistant. Your task is to assist the user with the following context from an academic paper.

        **Role**: {role_description}

        **Context**: {assistant_context}
        
        **User Query**: {user_input}

        **Instructions**: {response_instructions}
        Question Length Requirement: {length_guidelines[question_length]}

        **Bloom's Taxonomy Weights**:
        Knowledge: {bloom_taxonomy_weights['Knowledge']}%
        Comprehension: {bloom_taxonomy_weights['Comprehension']}%
        Application: {bloom_taxonomy_weights['Application']}%
        Analysis: {bloom_taxonomy_weights['Analysis']}%
        Synthesis: {bloom_taxonomy_weights['Synthesis']}%
        Evaluation: {bloom_taxonomy_weights['Evaluation']}%

        **Query**: {user_query}

        **Number of Questions**: {num_questions}

        **Include Numericals**: {include_numericals}
        """
        
        payload = {
            "contents": [
                {
                    "parts": [
                        {"text": prompt}
                    ]
                }
            ]
        }
        headers = {"Content-Type": "application/json"}

        response = requests.post(url, headers=headers, data=json.dumps(payload), timeout=60)
        response.raise_for_status()
        
        result = response.json()
        questions = result.get("candidates", [{}])[0].get("content", {}).get("parts", [{}])[0].get("text", "")
        questions_list = [question.strip() for question in questions.split("\n") if question.strip()]
        
        # Get Bloom's taxonomy scores for each question with progress bar
        questions_with_scores = []
        progress_bar = st.progress(0)
        for idx, question in enumerate(questions_list):
            scores = get_bloom_taxonomy_scores(question)
            if scores:  # Only add questions that got valid scores
                questions_with_scores.append((question, scores))
            progress_bar.progress((idx + 1) / len(questions_list))
        
        if not questions_with_scores:
            st.warning("Could not get Bloom's Taxonomy scores for any questions. Using default scores.")
            # Use default scores if no scores were obtained
            questions_with_scores = [(q, get_bloom_taxonomy_scores("")) for q in questions_list]
        
        # Update session state with scores
        st.session_state.question_scores = {q: s for q, s in questions_with_scores}
        
        # Return just the questions
        return [q for q, _ in questions_with_scores]
    except requests.RequestException as e:
        st.error(f"API request error: {e}")
        return []
    except Exception as e:
        st.error(f"Error generating questions: {e}")
        return []

def normalize_bloom_weights(bloom_weights):
    total = sum(bloom_weights.values())
    if total != 100:
        normalization_factor = 100 / total
        # Normalize each weight by multiplying it by the normalization factor
        bloom_weights = {key: round(value * normalization_factor, 2) for key, value in bloom_weights.items()}
    return bloom_weights

def generate_pdf(questions, filename="questions.pdf"):
    try:
        pdf = FPDF()
        pdf.set_auto_page_break(auto=True, margin=15)
        pdf.add_page()

        # Set font
        pdf.add_font("ArialUnicode", "", "ArialUnicodeMS.ttf", uni=True)
        pdf.set_font("ArialUnicode", size=12)
        
        # Add a title or heading
        pdf.cell(200, 10, txt="Generated Questions", ln=True, align="C")

        # Add space between title and questions
        pdf.ln(10)

        # Loop through questions and add them to the PDF
        for i, question in enumerate(questions, 1):
            # Using multi_cell for wrapping the text in case it's too long
            pdf.multi_cell(0, 10, f"Q{i}: {question}")

        # Save the generated PDF to the file
        pdf.output(filename)
        return filename
    except Exception as e:
        st.error(f"Error generating PDF: {e}")
        return None

def process_pdf_and_generate_questions(pdf_source, uploaded_file, api_key, role_description, response_instructions, bloom_taxonomy_weights, num_questions, question_length, include_numericals, user_input):
    try:

        pdf_path = get_pdf_path(pdf_source, uploaded_file)
        if not pdf_path:
            return []

        # Extract text
        pdf_text = extract_text_pymupdf(pdf_path)
        if not pdf_text:
            return []
        # Generate questions
        assistant_context = pdf_text
        user_query = "Generate questions based on the above context."
        normalized_bloom_weights = normalize_bloom_weights(bloom_taxonomy_weights)
        questions = generate_ai_response(
            api_key, 
            assistant_context, 
            user_query, 
            role_description, 
            response_instructions, 
            normalized_bloom_weights, 
            num_questions,
            question_length,
            include_numericals,
            user_input
        )

        # Clean up temporary PDF file
        try:
            os.remove(pdf_path)
            # Remove the temporary directory
            os.rmdir(os.path.dirname(pdf_path))
        except Exception as e:
            st.warning(f"Could not delete temporary PDF file: {e}")

        return questions
    except Exception as e:
        st.error(f"Error processing PDF and generating questions: {e}")
        return []
    
def get_bloom_taxonomy_details(question_scores: Optional[Dict[str, float]] = None) -> str:
    """
    Generate a detailed explanation of Bloom's Taxonomy scores.
    Handles missing or invalid scores gracefully.
    """
    try:
        if question_scores is None or not isinstance(question_scores, dict):
            return "Bloom's Taxonomy scores not available"
        
        # Validate scores
        valid_categories = {"Remembering", "Understanding", "Applying", 
                          "Analyzing", "Evaluating", "Creating"}
        
        if not all(isinstance(score, (int, float)) for score in question_scores.values()):
            return "Invalid score values detected"
            
        if not all(category in valid_categories for category in question_scores.keys()):
            return "Invalid score categories detected"
        
        details_text = "Bloom's Taxonomy Analysis:\n\n"
        
        try:
            # Sort scores by value in descending order
            sorted_scores = sorted(question_scores.items(), key=lambda x: x[1], reverse=True)
            
            # Format each score as a percentage
            for category, score in sorted_scores:
                percentage = min(max(score * 100, 0), 100)  # Ensure percentage is between 0 and 100
                details_text += f"{category}: {percentage:.1f}%\n"
            
            # Add the predicted level
            predicted_level = max(question_scores.items(), key=lambda x: x[1])[0]
            details_text += f"\nPredicted Level: {predicted_level}"
            
            return details_text.strip()
            
        except Exception as e:
            return f"Error processing scores: {str(e)}"
            
    except Exception as e:
        return f"Error generating taxonomy details: {str(e)}"
    
       
def predict_with_loaded_model(text):
    inputs = tokenizer(text, return_tensors='pt', padding=True, truncation=True, max_length=512)
    input_ids = inputs['input_ids'].to(device)
    model.eval()
    with torch.no_grad():
        outputs = model(input_ids)
        logits = outputs.logits
        probabilities = softmax(logits, dim=-1)
    probabilities = probabilities.squeeze().cpu().numpy()
    # Convert to float and format to 3 decimal places
    class_probabilities = {reverse_mapping[i]: float(f"{prob:.3f}") for i, prob in enumerate(probabilities)}
    return class_probabilities

def process_document(input_path):
    if input_path.lower().endswith(".pdf"):
        doc = DocumentFile.from_pdf(input_path)
        #print(f"Number of pages: {len(doc)}")
    elif input_path.lower().endswith((".jpg", ".jpeg", ".png", ".bmp", ".tiff")):
        doc = DocumentFile.from_images(input_path)
    else:
        raise ValueError("Unsupported file type. Please provide a PDF or an image file.")
    result = modelocr(doc)
    def calculate_average_confidence(result):
        total_confidence = 0
        word_count = 0
        for page in result.pages:
            for block in page.blocks:
                for line in block.lines:
                    for word in line.words:
                        total_confidence += word.confidence
                        word_count += 1
        average_confidence = total_confidence / word_count if word_count > 0 else 0
        return average_confidence
    average_confidence = calculate_average_confidence(result)
    string_result = result.render()
    return {'Avg_Confidence': average_confidence, 'String':string_result.split('\n')}

def sendtogemini(inputpath, question):
    if inputpath and inputpath.lower().endswith((".pdf", ".jpg", ".jpeg", ".png")):
        qw = process_document(inputpath)
    elif question:
        qw = {'String': [question]}
    else:
        raise ValueError("Unsupported file type. Please provide a PDF or an image file.")
    questionset = str(qw['String'])
    # send this prompt to gemini : 
    questionset += """You are given a list of text fragments containing questions fragments extracted by an ocr model. Your task is to:
    # only Merge the question fragments into complete and coherent questions.Don't answer then.
    # Separate each question , start a new question with @ to make them easily distinguishable for further processing."""
    url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash-latest:generateContent?key={os.getenv('GEMINI_API_KEY')}"
    
    payload = {
            "contents": [
                {
                    "parts": [
                        {"text": questionset}
                    ]
                }
            ]
        }
    headers = {"Content-Type": "application/json"}

    response = requests.post(url, headers=headers, data=json.dumps(payload), timeout=60)
    result = response.json()
    res1 = result.get("candidates", [{}])[0].get("content", {}).get("parts", [{}])[0].get("text", "")
    question = []
    for i in res1.split('\n'):
        i = i.strip()
        if len(i) > 0:
            if i[0] == '@':
                i = i[1:].strip().lower()
                if i[0] == 'q':
                    question.append(i[1:].strip())
                else:
                    question.append(i)
    data = []
    for i in question:
        d = {}
        d['question'] = i
        d['score'] = predict_with_loaded_model(i)
        data.append(d)
    return data