app.py

import streamlit as st
st.set_page_config(page_title="RAG Book Analyzer", layout="wide")

import torch
import numpy as np
import faiss
from transformers import AutoModelForCausalLM, AutoTokenizer
from sentence_transformers import SentenceTransformer
import fitz  # PyMuPDF
import docx2txt
from langchain_text_splitters import RecursiveCharacterTextSplitter

# ------------------------
# Configuration (optimized for reliability)
# ------------------------
MODEL_NAME = "microsoft/phi-2"
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"  # Efficient embedding model
CHUNK_SIZE = 512
CHUNK_OVERLAP = 64
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MAX_TEXT_LENGTH = 3000  # To prevent OOM errors

# ------------------------
# Model Loading with Robust Error Handling
# ------------------------
@st.cache_resource(show_spinner="Loading AI models...")
def load_models():
    try:
        # Load tokenizer with special settings for Phi-2
        tokenizer = AutoTokenizer.from_pretrained(
            MODEL_NAME,
            trust_remote_code=True,
            padding_side="left"
        )
        tokenizer.pad_token = tokenizer.eos_token
        
        # Load model with safe defaults
        model = AutoModelForCausalLM.from_pretrained(
            MODEL_NAME,
            torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
            trust_remote_code=True,
            device_map="auto" if DEVICE == "cuda" else None,
            low_cpu_mem_usage=True
        )
        
        # Load efficient embedding model
        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()

# ------------------------
# Text Processing Functions
# ------------------------
def split_text(text):
    splitter = RecursiveCharacterTextSplitter(
        chunk_size=CHUNK_SIZE,
        chunk_overlap=CHUNK_OVERLAP,
        length_function=len
    )
    return splitter.split_text(text)

def extract_text(file):
    try:
        if file.type == "application/pdf":
            doc = fitz.open(stream=file.read(), filetype="pdf")
            return "\n".join([page.get_text() for page in doc])
        elif file.type == "text/plain":
            return file.read().decode("utf-8")
        elif file.type == "application/vnd.openxmlformats-officedocument.wordprocessingml.document":
            return docx2txt.process(file)
        else:
            st.error(f"Unsupported file type: {file.type}")
            return ""
    except Exception as e:
        st.error(f"Error processing file: {str(e)}")
        return ""

def build_index(chunks):
    embeddings = embedder.encode(chunks, show_progress_bar=False)
    dimension = embeddings.shape[1]
    index = faiss.IndexFlatL2(dimension)
    index.add(embeddings)
    return index

# ------------------------
# AI Generation Functions (with safeguards)
# ------------------------
def generate_summary(text):
    text = text[:MAX_TEXT_LENGTH]  # Prevent long inputs
    prompt = f"Instruction: Summarize this book in a concise paragraph\nText: {text}\nSummary:"
    
    inputs = tokenizer(
        prompt, 
        return_tensors="pt",
        max_length=1024,
        truncation=True
    ).to(DEVICE)
    
    outputs = model.generate(
        **inputs,
        max_new_tokens=200,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )
    
    summary = tokenizer.decode(
        outputs[0], 
        skip_special_tokens=True
    )
    
    # Extract just the summary part
    if "Summary:" in summary:
        return summary.split("Summary:")[-1].strip()
    return summary.replace(prompt, "").strip()

def generate_answer(query, context):
    context = context[:MAX_TEXT_LENGTH]  # Limit context size
    prompt = f"Instruction: Answer this question based on the context. If unsure, say 'I don't know'.\nQuestion: {query}\nContext: {context}\nAnswer:"
    
    inputs = tokenizer(
        prompt, 
        return_tensors="pt",
        max_length=1024,
        truncation=True
    ).to(DEVICE)
    
    outputs = model.generate(
        **inputs,
        max_new_tokens=150,
        temperature=0.4,
        top_p=0.85,
        repetition_penalty=1.1,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )
    
    answer = tokenizer.decode(
        outputs[0], 
        skip_special_tokens=True
    )
    
    # Extract just the answer part
    if "Answer:" in answer:
        return answer.split("Answer:")[-1].strip()
    return answer.replace(prompt, "").strip()

# ------------------------
# Streamlit UI
# ------------------------
st.title("📚 RAG-Based Book Analyzer")
st.write("Upload a book (PDF, TXT, DOCX) to get a summary and ask questions about its content.")
st.warning("Note: First run will download models (~1.5GB). Please be patient!")

uploaded_file = st.file_uploader("Upload File", type=["pdf", "txt", "docx"])

if uploaded_file:
    with st.spinner("Extracting text from file..."):
        text = extract_text(uploaded_file)
    
    if not text:
        st.error("Failed to extract text. Please try another file.")
        st.stop()
    
    st.success(f"✅ Extracted {len(text)} characters")
    
    with st.spinner("Generating summary (this may take a minute)..."):
        summary = generate_summary(text)
        st.markdown("### Book Summary")
        st.info(summary)
    
    with st.spinner("Preparing document for questions..."):
        chunks = split_text(text)
        index = build_index(chunks)
        st.session_state.chunks = chunks
        st.session_state.index = index
        st.success(f"✅ Document indexed with {len(chunks)} chunks")

st.divider()

if 'chunks' in st.session_state:
    st.markdown("### ❓ Ask a Question about the Book")
    query = st.text_input("Enter your question:", key="question")
    
    if query:
        with st.spinner("Searching for answers..."):
            # Retrieve top 3 relevant chunks
            query_embedding = embedder.encode([query])
            distances, indices = st.session_state.index.search(query_embedding, k=3)
            
            # Safely retrieve chunks
            retrieved_chunks = []
            for i in indices[0]:
                if i < len(st.session_state.chunks):
                    retrieved_chunks.append(st.session_state.chunks[i])
            
            context = "\n\n".join(retrieved_chunks)
            
            # Generate answer
            answer = generate_answer(query, context)
            
            # Display results
            st.markdown("### 💬 Answer")
            st.success(answer)
            
            with st.expander("View context used for answer"):
                st.text(context)