app.py

import streamlit as st
import torch
import numpy as np
import faiss
from transformers import AutoModelForCausalLM, AutoTokenizer
from sentence_transformers import SentenceTransformer
import fitz  # PyMuPDF for better PDF extraction
from langchain_text_splitters import RecursiveCharacterTextSplitter

# Configuration
MODEL_NAME = "ibm-granite/granite-3.1-1b-a400m-instruct"
EMBED_MODEL = "sentence-transformers/all-mpnet-base-v2"
CHUNK_SIZE = 512
CHUNK_OVERLAP = 64
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# Initialize session state
if "docs" not in st.session_state:
    st.session_state.docs = []
if "index" not in st.session_state:
    st.session_state.index = None

# Model loading with better error handling
@st.cache_resource
def load_models():
    try:
        tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
        
        model = AutoModelForCausalLM.from_pretrained(
            MODEL_NAME,
            device_map="auto" if DEVICE == "cuda" else None,
            torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
            low_cpu_mem_usage=True
        ).eval()
        
        embedder = SentenceTransformer(EMBED_MODEL, device=DEVICE)
        return tokenizer, model, embedder
    
    except Exception as e:
        st.error(f"Model loading failed: {str(e)}")
        st.stop()

tokenizer, model, embedder = load_models()

# Improved text processing
def process_text(text):
    splitter = RecursiveCharacterTextSplitter(
        chunk_size=CHUNK_SIZE,
        chunk_overlap=CHUNK_OVERLAP,
        length_function=len
    )
    return splitter.split_text(text)

# Enhanced PDF extraction
def extract_pdf_text(uploaded_file):
    try:
        doc = fitz.open(stream=uploaded_file.read(), filetype="pdf")
        return "\n".join([page.get_text() for page in doc])
    except Exception as e:
        st.error(f"PDF extraction error: {str(e)}")
        return ""

# Multi-step summarization
def generate_summary(text):
    chunks = process_text(text)[:10]  # Use first 10 chunks for summary
    summaries = []
    
    for chunk in chunks:
        prompt = f"""<|user|>
        Summarize this text section focusing on key themes, characters, and plot points:
        {chunk[:2000]}
        <|assistant|>
        """
        
        inputs = tokenizer(prompt, return_tensors="pt").to(DEVICE)
        outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.3)
        summaries.append(tokenizer.decode(outputs[0], skip_special_tokens=True))
    
    # Combine summaries
    combined = "\n".join(summaries)
    final_prompt = f"""<|user|>
    Combine these section summaries into a coherent book summary:
    {combined}
    <|assistant|>
    The comprehensive summary is:"""
    
    inputs = tokenizer(final_prompt, return_tensors="pt").to(DEVICE)
    outputs = model.generate(**inputs, max_new_tokens=500, temperature=0.5)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split(":")[-1].strip()

# Enhanced retrieval system
def build_faiss_index(texts):
    embeddings = embedder.encode(texts, show_progress_bar=True)
    dimension = embeddings.shape[1]
    index = faiss.IndexFlatIP(dimension)
    faiss.normalize_L2(embeddings)
    index.add(embeddings)
    return index

# Context-aware generation
def generate_answer(query, context):
    prompt = f"""<|user|>
    Using this context: {context}
    Answer the question precisely and truthfully. If unsure, say "I don't know".
    Question: {query}
    <|assistant|>
    """
    
    inputs = tokenizer(prompt, return_tensors="pt", max_length=1024, truncation=True).to(DEVICE)
    outputs = model.generate(
        **inputs,
        max_new_tokens=300,
        temperature=0.4,
        top_p=0.9,
        repetition_penalty=1.2,
        do_sample=True
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split("<|assistant|>")[-1].strip()

# Streamlit UI
st.set_page_config(page_title="📚 Smart Book Analyst", layout="wide")
st.title("📚 AI-Powered Book Analysis System")

uploaded_file = st.file_uploader("Upload book (PDF or TXT)", type=["pdf", "txt"])

if uploaded_file:
    with st.spinner("📖 Analyzing book content..."):
        try:
            if uploaded_file.type == "application/pdf":
                text = extract_pdf_text(uploaded_file)
            else:
                text = uploaded_file.read().decode()
            
            chunks = process_text(text)
            st.session_state.docs = chunks
            st.session_state.index = build_faiss_index(chunks)
            
            with st.expander("📝 Book Summary", expanded=True):
                summary = generate_summary(text)
                st.write(summary)
                
        except Exception as e:
            st.error(f"Processing failed: {str(e)}")

if st.session_state.index:
    query = st.text_input("Ask about the book:")
    if query:
        with st.spinner("🔍 Searching for answers..."):
            try:
                # Retrieve top 3 relevant chunks
                query_embed = embedder.encode([query])
                faiss.normalize_L2(query_embed)
                distances, indices = st.session_state.index.search(query_embed, k=3)
                
                context = "\n".join([st.session_state.docs[i] for i in indices[0]])
                answer = generate_answer(query, context)
                
                st.subheader("Answer")
                st.markdown(f"```\n{answer}\n```")
                st.caption("Retrieved context confidence: {:.2f}".format(distances[0][0]))
                
            except Exception as e:
                st.error(f"Query failed: {str(e)}")