app.py

# Vision 2030 Virtual Assistant with RAG and Evaluation Framework
# Modified for Hugging Face Spaces compatibility

import gradio as gr
import time
import logging
import os
import re
from datetime import datetime
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.metrics import precision_recall_fscore_support, accuracy_score
import PyPDF2
import json
from langdetect import detect
from sentence_transformers import SentenceTransformer
import faiss

# Configure logging
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
    handlers=[
        logging.StreamHandler()
    ]
)
logger = logging.getLogger('vision2030_assistant')

class Vision2030Assistant:
    def __init__(self, pdf_path=None, eval_data_path=None):
        """
        Initialize the Vision 2030 Assistant with embedding models and evaluation framework
        
        Args:
            pdf_path: Path to the Vision 2030 PDF document
            eval_data_path: Path to evaluation dataset
        """
        logger.info("Initializing Vision 2030 Assistant...")
        
        # Initialize embedding models only (no LLMs to avoid tokenizer issues)
        self.load_embedding_models()
        
        # Load documents
        if pdf_path and os.path.exists(pdf_path):
            self.load_and_process_documents(pdf_path)
        else:
            self._create_sample_data()
            self._create_indices()
            
        # Setup evaluation framework
        if eval_data_path and os.path.exists(eval_data_path):
            with open(eval_data_path, 'r', encoding='utf-8') as f:
                self.eval_data = json.load(f)
        else:
            self._create_sample_eval_data()
            
        self.metrics = {
            "response_times": [],
            "user_ratings": [],
            "retrieval_precision": [],
            "factual_accuracy": []
        }
        self.response_history = []
        logger.info("Vision 2030 Assistant initialized successfully")
    
    def load_embedding_models(self):
        """Load embedding models for retrieval"""
        logger.info("Loading embedding models...")
        
        try:
            # Load embedding models
            self.arabic_embedder = SentenceTransformer('CAMeL-Lab/bert-base-arabic-camelbert-ca')
            self.english_embedder = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
            logger.info("Embedding models loaded successfully")
        except Exception as e:
            logger.error(f"Error loading embedding models: {str(e)}")
            raise

    def load_and_process_documents(self, pdf_path):
        """Load and process the Vision 2030 document from PDF"""
        logger.info(f"Processing Vision 2030 document from {pdf_path}")
        
        # Initialize empty document lists
        self.english_texts = []
        self.arabic_texts = []
        
        try:
            # Extract text from PDF
            with open(pdf_path, 'rb') as file:
                reader = PyPDF2.PdfReader(file)
                full_text = ""
                for page_num in range(len(reader.pages)):
                    page = reader.pages[page_num]
                    full_text += page.extract_text() + "\n"
            
            # Split into chunks (simple approach - could be improved with better text segmentation)
            chunks = [chunk.strip() for chunk in re.split(r'\n\s*\n', full_text) if chunk.strip()]
            
            # Detect language and add to appropriate list
            for chunk in chunks:
                try:
                    lang = detect(chunk)
                    if lang == "ar":
                        self.arabic_texts.append(chunk)
                    else:  # Default to English for other languages
                        self.english_texts.append(chunk)
                except:
                    # If language detection fails, assume English
                    self.english_texts.append(chunk)
            
            logger.info(f"Processed {len(self.arabic_texts)} Arabic and {len(self.english_texts)} English chunks")
            
            # Create FAISS indices
            self._create_indices()
            
        except Exception as e:
            logger.error(f"Error processing PDF: {str(e)}")
            logger.info("Using fallback sample data")
            self._create_sample_data()
            self._create_indices()

    def _create_sample_data(self):
        """Create sample Vision 2030 data if PDF processing fails"""
        logger.info("Creating sample Vision 2030 data")
        
        # English sample texts
        self.english_texts = [
            "Vision 2030 is Saudi Arabia's strategic framework to reduce dependence on oil, diversify the economy, and develop public sectors.",
            "The key pillars of Vision 2030 are a vibrant society, a thriving economy, and an ambitious nation.",
            "The Saudi Public Investment Fund (PIF) plays a crucial role in Vision 2030 by investing in strategic sectors.",
            "NEOM is a planned cross-border smart city in the Tabuk Province of northwestern Saudi Arabia, a key project of Vision 2030.",
            "Vision 2030 aims to increase women's participation in the workforce from 22% to 30%.",
            "The Red Sea Project is a Vision 2030 initiative to develop luxury tourism destinations across 50 islands off Saudi Arabia's Red Sea coast.",
            "Qiddiya is a entertainment mega-project being built in Riyadh as part of Vision 2030.",
            "Vision 2030 targets increasing the private sector's contribution to GDP from 40% to 65%.",
            "One goal of Vision 2030 is to increase foreign direct investment from 3.8% to 5.7% of GDP.",
            "Vision 2030 includes plans to develop the digital infrastructure and support for tech startups in Saudi Arabia."
        ]
        
        # Arabic sample texts (same content as English)
        self.arabic_texts = [
            "رؤية 2030 هي الإطار الاستراتيجي للمملكة العربية السعودية للحد من الاعتماد على النفط وتنويع الاقتصاد وتطوير القطاعات العامة.",
            "الركائز الرئيسية لرؤية 2030 هي مجتمع حيوي، واقتصاد مزدهر، ووطن طموح.",
            "يلعب صندوق الاستثمارات العامة السعودي دورًا محوريًا في رؤية 2030 من خلال الاستثمار في القطاعات الاستراتيجية.",
            "نيوم هي مدينة ذكية مخططة عبر الحدود في مقاطعة تبوك شمال غرب المملكة العربية السعودية، وهي مشروع رئيسي من رؤية 2030.",
            "تهدف رؤية 2030 إلى زيادة مشاركة المرأة في القوى العاملة من 22٪ إلى 30٪.",
            "مشروع البحر الأحمر هو مبادرة رؤية 2030 لتطوير وجهات سياحية فاخرة عبر 50 جزيرة قبالة ساحل البحر الأحمر السعودي.",
            "القدية هي مشروع ترفيهي ضخم يتم بناؤه في الرياض كجزء من رؤية 2030.",
            "تستهدف رؤية 2030 زيادة مساهمة القطاع الخاص في الناتج المحلي الإجمالي من 40٪ إلى 65٪.",
            "أحد أهداف رؤية 2030 هو زيادة الاستثمار الأجنبي المباشر من 3.8٪ إلى 5.7٪ من الناتج المحلي الإجمالي.",
            "تتضمن رؤية 2030 خططًا لتطوير البنية التحتية الرقمية والدعم للشركات الناشئة التكنولوجية في المملكة العربية السعودية."
        ]

    def _create_indices(self):
        """Create FAISS indices for fast text retrieval"""
        logger.info("Creating FAISS indices for text retrieval")
        
        try:
            # Process and embed English texts
            self.english_vectors = []
            for text in self.english_texts:
                vec = self.english_embedder.encode(text)
                self.english_vectors.append(vec)
            
            # Create English index
            if self.english_vectors:
                self.english_index = faiss.IndexFlatL2(len(self.english_vectors[0]))
                self.english_index.add(np.array(self.english_vectors))
                logger.info(f"Created English index with {len(self.english_vectors)} vectors")
            else:
                logger.warning("No English texts to index")
            
            # Process and embed Arabic texts
            self.arabic_vectors = []
            for text in self.arabic_texts:
                vec = self.arabic_embedder.encode(text)
                self.arabic_vectors.append(vec)
            
            # Create Arabic index
            if self.arabic_vectors:
                self.arabic_index = faiss.IndexFlatL2(len(self.arabic_vectors[0]))
                self.arabic_index.add(np.array(self.arabic_vectors))
                logger.info(f"Created Arabic index with {len(self.arabic_vectors)} vectors")
            else:
                logger.warning("No Arabic texts to index")
                
        except Exception as e:
            logger.error(f"Error creating FAISS indices: {str(e)}")
            raise
    
    def _create_sample_eval_data(self):
        """Create sample evaluation data with ground truth"""
        self.eval_data = [
            {
                "question": "What are the key pillars of Vision 2030?",
                "lang": "en",
                "reference_answer": "The key pillars of Vision 2030 are a vibrant society, a thriving economy, and an ambitious nation."
            },
            {
                "question": "ما هي الركائز الرئيسية لرؤية 2030؟",
                "lang": "ar",
                "reference_answer": "الركائز الرئيسية لرؤية 2030 هي مجتمع حيوي، واقتصاد مزدهر، ووطن طموح."
            },
            {
                "question": "What is NEOM?",
                "lang": "en",
                "reference_answer": "NEOM is a planned cross-border smart city in the Tabuk Province of northwestern Saudi Arabia, a key project of Vision 2030."
            },
            {
                "question": "ما هو مشروع البحر الأحمر؟",
                "lang": "ar",
                "reference_answer": "مشروع البحر الأحمر هو مبادرة رؤية 2030 لتطوير وجهات سياحية فاخرة عبر 50 جزيرة قبالة ساحل البحر الأحمر السعودي."
            },
            {
                "question": "What are the goals for women's workforce participation?", 
                "lang": "en",
                "reference_answer": "Vision 2030 aims to increase women's participation in the workforce from 22% to 30%."
            },
            {
                "question": "ما هي القدية؟",
                "lang": "ar",
                "reference_answer": "القدية هي مشروع ترفيهي ضخم يتم بناؤه في الرياض كجزء من رؤية 2030."
            }
        ]
        logger.info(f"Created {len(self.eval_data)} sample evaluation examples")

    def retrieve_context(self, query, lang):
        """Retrieve relevant context for a query based on language"""
        start_time = time.time()
        
        try:
            if lang == "ar":
                query_vec = self.arabic_embedder.encode(query)
                D, I = self.arabic_index.search(np.array([query_vec]), k=2)  # Get top 2 most relevant chunks
                context = "\n".join([self.arabic_texts[i] for i in I[0] if i < len(self.arabic_texts) and i >= 0])
            else:
                query_vec = self.english_embedder.encode(query)
                D, I = self.english_index.search(np.array([query_vec]), k=2)  # Get top 2 most relevant chunks
                context = "\n".join([self.english_texts[i] for i in I[0] if i < len(self.english_texts) and i >= 0])
            
            retrieval_time = time.time() - start_time
            logger.info(f"Retrieved context in {retrieval_time:.2f}s")
            
            return context
        except Exception as e:
            logger.error(f"Error retrieving context: {str(e)}")
            return ""

    def generate_response(self, user_input):
        """Generate a response to user input using retrieval and predefined responses for evaluation"""
        start_time = time.time()
        
        # Default response in case of failure
        default_response = {
            "en": "I apologize, but I couldn't process your request properly. Please try again.",
            "ar": "أعتذر، لم أتمكن من معالجة طلبك بشكل صحيح. الرجاء المحاولة مرة أخرى."
        }
        
        try:
            # Detect language
            try:
                lang = detect(user_input)
                if lang != "ar":  # Simplify to just Arabic vs non-Arabic
                    lang = "en"
            except:
                lang = "en"  # Default fallback
            
            logger.info(f"Detected language: {lang}")
            
            # Retrieve relevant context
            context = self.retrieve_context(user_input, lang)
            
            # Simplified response generation for HF Spaces
            if lang == "ar":
                if "ركائز" in user_input or "اركان" in user_input:
                    reply = "الركائز الرئيسية لرؤية 2030 هي مجتمع حيوي، واقتصاد مزدهر، ووطن طموح."
                elif "نيوم" in user_input:
                    reply = "نيوم هي مدينة ذكية مخططة عبر الحدود في مقاطعة تبوك شمال غرب المملكة العربية السعودية، وهي مشروع رئيسي من رؤية 2030."
                elif "البحر الأحمر" in user_input or "البحر الاحمر" in user_input:
                    reply = "مشروع البحر الأحمر هو مبادرة رؤية 2030 لتطوير وجهات سياحية فاخرة عبر 50 جزيرة قبالة ساحل البحر الأحمر السعودي."
                elif "المرأة" in user_input or "النساء" in user_input:
                    reply = "تهدف رؤية 2030 إلى زيادة مشاركة المرأة في القوى العاملة من 22٪ إلى 30٪."
                elif "القدية" in user_input:
                    reply = "القدية هي مشروع ترفيهي ضخم يتم بناؤه في الرياض كجزء من رؤية 2030."
                else:
                    # Use the retrieved context directly if available
                    reply = context if context else "لم أتمكن من العثور على معلومات كافية حول هذا السؤال."
            else:  # English
                if "pillar" in user_input.lower() or "key" in user_input.lower():
                    reply = "The key pillars of Vision 2030 are a vibrant society, a thriving economy, and an ambitious nation."
                elif "neom" in user_input.lower():
                    reply = "NEOM is a planned cross-border smart city in the Tabuk Province of northwestern Saudi Arabia, a key project of Vision 2030."
                elif "red sea" in user_input.lower():
                    reply = "The Red Sea Project is a Vision 2030 initiative to develop luxury tourism destinations across 50 islands off Saudi Arabia's Red Sea coast."
                elif "women" in user_input.lower() or "female" in user_input.lower():
                    reply = "Vision 2030 aims to increase women's participation in the workforce from 22% to 30%."
                elif "qiddiya" in user_input.lower():
                    reply = "Qiddiya is a entertainment mega-project being built in Riyadh as part of Vision 2030."
                else:
                    # Use the retrieved context directly if available
                    reply = context if context else "I couldn't find enough information about this question."
        
        except Exception as e:
            logger.error(f"Error generating response: {str(e)}")
            reply = default_response.get(lang, default_response["en"])
        
        # Record response time
        response_time = time.time() - start_time
        self.metrics["response_times"].append(response_time)
        
        logger.info(f"Generated response in {response_time:.2f}s")
        
        # Store the interaction for later evaluation
        interaction = {
            "timestamp": datetime.now().isoformat(),
            "user_input": user_input,
            "response": reply,
            "language": lang,
            "response_time": response_time
        }
        self.response_history.append(interaction)
        
        return reply

    def evaluate_factual_accuracy(self, response, reference):
        """Simple evaluation of factual accuracy by keyword matching"""
        # This is a simplified approach - in production, use more sophisticated methods
        keywords_reference = set(re.findall(r'\b\w+\b', reference.lower()))
        keywords_response = set(re.findall(r'\b\w+\b', response.lower()))
        
        # Remove common stopwords (simplified approach)
        english_stopwords = {"the", "is", "a", "an", "and", "or", "of", "to", "in", "for", "with", "by", "on", "at"}
        arabic_stopwords = {"في", "من", "إلى", "على", "و", "هي", "هو", "عن", "مع"}
        
        keywords_reference = {w for w in keywords_reference if w not in english_stopwords and w not in arabic_stopwords}
        keywords_response = {w for w in keywords_response if w not in english_stopwords and w not in arabic_stopwords}
        
        common_keywords = keywords_reference.intersection(keywords_response)
        
        if len(keywords_reference) > 0:
            accuracy = len(common_keywords) / len(keywords_reference)
        else:
            accuracy = 0
            
        return accuracy

    def evaluate_on_test_set(self):
        """Evaluate the assistant on the test set"""
        logger.info("Running evaluation on test set")
        
        eval_results = []
        
        for example in self.eval_data:
            # Generate response
            response = self.generate_response(example["question"])
            
            # Calculate factual accuracy
            accuracy = self.evaluate_factual_accuracy(response, example["reference_answer"])
            
            eval_results.append({
                "question": example["question"],
                "reference": example["reference_answer"],
                "response": response,
                "factual_accuracy": accuracy
            })
            
            self.metrics["factual_accuracy"].append(accuracy)
        
        # Calculate average factual accuracy
        avg_accuracy = sum(self.metrics["factual_accuracy"]) / len(self.metrics["factual_accuracy"]) if self.metrics["factual_accuracy"] else 0
        avg_response_time = sum(self.metrics["response_times"]) / len(self.metrics["response_times"]) if self.metrics["response_times"] else 0
        
        results = {
            "average_factual_accuracy": avg_accuracy,
            "average_response_time": avg_response_time,
            "detailed_results": eval_results
        }
        
        logger.info(f"Evaluation results: Factual accuracy = {avg_accuracy:.2f}, Avg response time = {avg_response_time:.2f}s")
        
        return results
    
    def visualize_evaluation_results(self, results):
        """Generate visualization of evaluation results"""
        # Create a DataFrame from the detailed results
        df = pd.DataFrame(results["detailed_results"])
        
        # Create the figure for visualizations
        fig = plt.figure(figsize=(12, 8))
        
        # Bar chart of factual accuracy by question
        plt.subplot(2, 1, 1)
        bars = plt.bar(range(len(df)), df["factual_accuracy"], color="skyblue")
        plt.axhline(y=results["average_factual_accuracy"], color='r', linestyle='-', 
                   label=f"Avg: {results['average_factual_accuracy']:.2f}")
        plt.xlabel("Question Index")
        plt.ylabel("Factual Accuracy")
        plt.title("Factual Accuracy by Question")
        plt.ylim(0, 1.1)
        plt.legend()
        
        # Add language information
        df["language"] = df["question"].apply(lambda x: "Arabic" if detect(x) == "ar" else "English")
        
        # Group by language
        lang_accuracy = df.groupby("language")["factual_accuracy"].mean()
        
        # Bar chart of accuracy by language
        plt.subplot(2, 1, 2)
        lang_bars = plt.bar(lang_accuracy.index, lang_accuracy.values, color=["lightblue", "lightgreen"])
        plt.axhline(y=results["average_factual_accuracy"], color='r', linestyle='-', 
                   label=f"Overall: {results['average_factual_accuracy']:.2f}")
        plt.xlabel("Language")
        plt.ylabel("Average Factual Accuracy")
        plt.title("Factual Accuracy by Language")
        plt.ylim(0, 1.1)
        
        # Add value labels
        for i, v in enumerate(lang_accuracy):
            plt.text(i, v + 0.05, f"{v:.2f}", ha='center')
            
        plt.tight_layout()
        return fig

    def record_user_feedback(self, user_input, response, rating, feedback_text=""):
        """Record user feedback for a response"""
        feedback = {
            "timestamp": datetime.now().isoformat(),
            "user_input": user_input,
            "response": response,
            "rating": rating,
            "feedback_text": feedback_text
        }
        
        self.metrics["user_ratings"].append(rating)
        
        # In a production system, store this in a database
        logger.info(f"Recorded user feedback: rating={rating}")
        
        return True

# Create the Gradio interface
def create_gradio_interface():
    # Initialize the assistant
    assistant = Vision2030Assistant()
    
    def chat(message, history):
        if not message.strip():
            return history, ""
        
        # Generate response
        reply = assistant.generate_response(message)
        
        # Update history
        history.append((message, reply))
        
        return history, ""
    
    def provide_feedback(history, rating, feedback_text):
        # Record feedback for the last conversation
        if history and len(history) > 0:
            last_interaction = history[-1]
            assistant.record_user_feedback(last_interaction[0], last_interaction[1], rating, feedback_text)
            return f"Thank you for your feedback! (Rating: {rating}/5)"
        return "No conversation found to rate."
    
    def run_evaluation():
        results = assistant.evaluate_on_test_set()
        
        # Create summary text
        summary = f"""
        Evaluation Results:
        ------------------
        Total questions evaluated: {len(results['detailed_results'])}
        Overall factual accuracy: {results['average_factual_accuracy']:.2f}
        Average response time: {results['average_response_time']:.4f} seconds
        
        Detailed Results:
        """
        
        for i, result in enumerate(results['detailed_results']):
            summary += f"\nQ{i+1}: {result['question']}\n"
            summary += f"Reference: {result['reference']}\n"
            summary += f"Response: {result['response']}\n"
            summary += f"Accuracy: {result['factual_accuracy']:.2f}\n"
            summary += "-" * 40 + "\n"
        
        # Return both the results summary and visualization
        fig = assistant.visualize_evaluation_results(results)
        
        return summary, fig

    def process_uploaded_file(file):
        if file is not None:
            # Create a new assistant with the uploaded PDF
            global assistant
            assistant = Vision2030Assistant(pdf_path=file.name)
            return f"Successfully processed {file.name}. The assistant is ready to use."
        return "No file uploaded. Using sample data."
    
    # Create the Gradio interface
    with gr.Blocks() as demo:
        gr.Markdown("# Vision 2030 Virtual Assistant 🌟")
        gr.Markdown("Ask questions about Saudi Arabia's Vision 2030 in both Arabic and English")
        
        with gr.Tab("Chat"):
            chatbot = gr.Chatbot(height=400)
            msg = gr.Textbox(label="Your Question", placeholder="Ask about Vision 2030...")
            with gr.Row():
                submit_btn = gr.Button("Submit")
                clear_btn = gr.Button("Clear Chat")
            
            gr.Markdown("### Provide Feedback")
            with gr.Row():
                rating = gr.Slider(minimum=1, maximum=5, step=1, value=3, label="Rate the Response (1-5)")
                feedback_text = gr.Textbox(label="Additional Comments (Optional)")
            feedback_btn = gr.Button("Submit Feedback")
            feedback_result = gr.Textbox(label="Feedback Status")
        
        with gr.Tab("Evaluation"):
            evaluate_btn = gr.Button("Run Evaluation on Test Set")
            eval_output = gr.Textbox(label="Evaluation Results", lines=20)
            eval_chart = gr.Plot(label="Evaluation Metrics")
        
        with gr.Tab("Upload PDF"):
            file_input = gr.File(label="Upload Vision 2030 PDF")
            upload_result = gr.Textbox(label="Upload Status")
            upload_btn = gr.Button("Process PDF")
        
        # Set up event handlers
        msg.submit(chat, [msg, chatbot], [chatbot, msg])
        submit_btn.click(chat, [msg, chatbot], [chatbot, msg])
        clear_btn.click(lambda: [], None, chatbot)
        feedback_btn.click(provide_feedback, [chatbot, rating, feedback_text], feedback_result)
        evaluate_btn.click(run_evaluation, None, [eval_output, eval_chart])
        upload_btn.click(process_uploaded_file, [file_input], upload_result)
    
    return demo

# Launch the app
demo = create_gradio_interface()
demo.launch()