app.py

import gradio as gr
import requests
import io
from pypdf import PdfReader
from concurrent.futures import ThreadPoolExecutor, as_completed
import numpy as np
import faiss
from sentence_transformers import SentenceTransformer, CrossEncoder
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, pipeline as hf_pipeline
from accelerate import Accelerator
from langchain.text_splitter import NLTKTextSplitter
from rank_bm25 import BM25Okapi
import os
import pickle
import nltk

nltk.download('punkt_tab')

# --- Global Variables for Caching ---
index = None
chunks = None
embedding_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
generator = None

# --- PDF Processing and Embedding ---
def download_pdf(url):
    response = requests.get(url, stream=True)
    response.raise_for_status()
    return response.content

def custom_chunking(text, delimiter="\n\n"):
    """Splits text based on a specified delimiter."""
    return text.split(delimiter)
    
def extract_text_from_pdf(pdf_bytes, document_id):
    """Extracts text from a PDF, page by page, and then chunks each page."""
    pdf_file = io.BytesIO(pdf_bytes)
    reader = PdfReader(pdf_file)
    nltk_splitter = NLTKTextSplitter(chunk_size=500)

    extracted_data = []
    for page_num, page in enumerate(reader.pages):
        page_text = page.extract_text() or ""
        clean_text = " ".join(page_text.split())

        if clean_text:
            words = clean_text.split()
            section_header = " ".join(words[:20]) if words else "No Section Name Found"
            custom_chunks = custom_chunking(clean_text)

            for custom_chunk in custom_chunks:
                clean_custom_chunk = " ".join(custom_chunk.split())
                if clean_custom_chunk:
                    nltk_chunks = nltk_splitter.split_text(clean_custom_chunk)

                    for nltk_chunk in nltk_chunks:
                        clean_nltk_chunk = " ".join(nltk_chunk.split())
                        if clean_nltk_chunk:
                            extracted_data.append({
                                "document_id": document_id,
                                "section_header": section_header,
                                "text": clean_nltk_chunk
                            })
    return extracted_data

def process_single_pdf(url, doc_id):
    """Processes a single PDF."""
    pdf_bytes = download_pdf(url)
    return extract_text_from_pdf(pdf_bytes, doc_id)

def process_pdfs_parallel(pdf_urls, document_ids):
    """Processes multiple PDFs in parallel."""
    all_data = []
    with ThreadPoolExecutor() as pdf_executor:
        pdf_futures = [pdf_executor.submit(process_single_pdf, url, doc_id) for url, doc_id in zip(pdf_urls, document_ids)]
        for future in as_completed(pdf_futures):
            all_data.extend(future.result())
    return all_data

def create_embeddings_and_index(data):
    """Create Embeddings"""
    texts = [item['text'] for item in data]
    embeddings = embedding_model.encode(texts)
    dimension = embeddings.shape[1]
    index = faiss.IndexFlatL2(dimension)
    index.add(embeddings)
    return index, data

# --- Retrieval Functions ---
def bm25_retrieval(query, documents, top_k=10):
    tokenized_docs = [doc['text'].split() for doc in documents]
    bm25 = BM25Okapi(tokenized_docs)
    doc_scores = bm25.get_scores(query.split())
    top_indices = np.argsort(doc_scores)[::-1][:top_k]
    return [documents[i] for i in top_indices]

def adaptive_retrieval(query, index, chunks, top_k=10):
    query_embedding = embedding_model.encode([query], convert_to_numpy=True)
    _, indices = index.search(query_embedding, top_k)
    vector_results = [chunks[i] for i in indices[0]]
    bm25_results = bm25_retrieval(query, chunks, top_k)
    combined_results = vector_results + bm25_results

    unique_results = []
    seen_texts = set()
    for result in combined_results:
        if result['text'] not in seen_texts:
            unique_results.append(result)
            seen_texts.add(result['text'])

    return unique_results

def rerank(query, results, keyword_weight=0.3, cross_encoder_weight=0.7):
    """Combines keyword-based and cross-encoder reranking."""

    # Keyword-based scoring
    keywords = query.lower().split()
    def score_chunk_keywords(chunk):
        text = chunk['text'].lower()
        return sum(1 for keyword in keywords if keyword in text)

    keyword_scores = [score_chunk_keywords(chunk) for chunk in results]

    # Cross-encoder scoring
    rerank_model = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-12-v2')
    query_results = [[query, f"Document: {result['document_id']}, Section: {result['section_header']}, Text: {result['text']}"] for result in results]
    cross_encoder_scores = rerank_model.predict(query_results)

    # Combine scores
    combined_scores = [(keyword_scores[i] * keyword_weight) + (cross_encoder_scores[i] * cross_encoder_weight) for i in range(len(results))]

    # Rank and select top 3
    ranked_results = [results[i] for i in np.argsort(combined_scores)[::-1]]
    return ranked_results[:3]

def merge_chunks(retrieved_chunks):
    """Merges chunks based on their original order, including metadata."""
    merged_text = " ".join([
        f"Document: {chunk['document_id']}, Section: {chunk['section_header']}, Text: {chunk['text']}"
        for chunk in retrieved_chunks
    ])
    return merged_text

# --- Confidence Calculation ---
def calculate_confidence(query, context, answer):
    """Calculates confidence score based on question-context and context-answer similarity."""
    query_embedding = embedding_model.encode([query], convert_to_numpy=True)
    context_embedding = embedding_model.encode([context], convert_to_numpy=True)
    answer_embedding = embedding_model.encode([answer], convert_to_numpy=True)
    
    query_context_similarity = np.dot(query_embedding, context_embedding.T).item()
    context_answer_similarity = np.dot(context_embedding, answer_embedding.T).item()
    confidence = (query_context_similarity + context_answer_similarity) / 2.0 # Equal weights

    return confidence

# --- Response Generation ---
def generate_response(query, context):
    prompt = f"""Your task is to analyze the given Context and take the answer for the Question and provide a clear relevant answer in plain English.
        **Guidelines:**
        - JUST PROVIDE ONLY THE ANSWER.
        - Provide a elaborate, factual answer based strictly on the Context.
        - Avoid generating Python code, solutions, or any irrelevant information.
    Context: {context}  
    Question: {query}
    Answer:"""
    response = generator(prompt, max_new_tokens=500, num_return_sequences=1)[0]['generated_text']
    return response

# --- Guardrail ---
def is_sensitive_query(query):
    sensitive_keywords = ["personal", "address", "phone", "ssn", "credit card", "bank account", "password", "social security", "private", "location"]
    query_lower = query.lower()
    if any(keyword in query_lower for keyword in sensitive_keywords):
        return True

    classifier = hf_pipeline("text-classification", model="unitary/toxic-bert")
    result = classifier(query)[0]

    if result["label"] == "toxic" and result["score"] > 0.7:
        return True

    return False

# --- Process Query ---
def process_query(query):
    if is_sensitive_query(query):
        return "I cannot answer questions that involve sensitive or personal information, or that are toxic in nature."

    retrieved_chunks = adaptive_retrieval(query, index, chunks)
    reranked_chunks = rerank(query, retrieved_chunks)
    final_chunks = reranked_chunks[:3]
    merged_result = merge_chunks(final_chunks)
    answer = generate_response(query, merged_result)
    if "</think>" in answer:
        answer = answer.split("</think>", 1)[-1].strip()
    confidence = calculate_confidence(query, merged_result, answer)
    full_response = f"{answer}\n\nConfidence: {confidence:.2f}"
    return full_response

# --- Initialization ---
def initialize_app():
    global index, chunks, generator

    pdf_urls = ["https://www.latentview.com/wp-content/uploads/2023/07/LatentView-Annual-Report-2022-23.pdf",
    "https://www.latentview.com/wp-content/uploads/2024/08/LatentView-Annual-Report-2023-24.pdf"]
    document_ids = ["LatentView-Annual-Report-2022-23", "LatentView-Annual-Report-2023-24"]

    if os.path.exists('vector_cache.pkl'):
        with open('vector_cache.pkl', 'rb') as f:
            index, chunks = pickle.load(f)
    else:
        extracted_data = process_pdfs_parallel(pdf_urls, document_ids)
        index, chunks = create_embeddings_and_index(extracted_data)
        with open('vector_cache.pkl', 'wb') as f:
            pickle.dump((index, chunks), f)

    accelerator = Accelerator()
    accelerator.free_memory()
    MODEL_NAME = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="auto")
    model = accelerator.prepare(model)
    generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
    return "Initialization Complete!"

# --- Gradio Interface ---
def gradio_interface(query):
    return process_query(query)

iface = gr.Interface(
    fn=gradio_interface,
    inputs=gr.Textbox(lines=2, placeholder="Enter your question here..."),
    outputs=gr.Textbox(lines=5, placeholder="Answer will appear here..."),
    title="Annual Report Q&A Chatbot (LatentView Analytics)",
    description="Ask questions about the company's annual reports. (2022-23 & 2023-24)",
    examples=[
        ["What is the total revenue from operations for 2023-24?"],
        ["Who is the CEO of Latentview Analytics? "],
        ["Summarize the key financial highlights in 2023-24"],
        ["What were the total expenses for 2022-23?"],
    ],
    cache_examples=False,
)

with gr.Blocks() as demo:
    gr.Markdown("# Annual Report Q&A Chatbot (LatentView Analytics)")
    init_button = gr.Button("Initialize")
    init_output = gr.Textbox(label="Initialization Status")

    init_button.click(
        fn=initialize_app,
        inputs=[],
        outputs=init_output,
    )

    iface.render()

demo.launch()