app.py

import os
import gradio as gr
import faiss
import pickle
from PyPDF2 import PdfReader
from sentence_transformers import SentenceTransformer
import pdfplumber
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

# Use FLAN-T5 instead of DeepSeek
model_id = "google/flan-t5-small"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)
model.eval()

"""def generate_answer(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=512)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)"""

def generate_answer(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.9,            # Higher = more creative
            repetition_penalty=1.1,     # Penalize repeating the same phrases
            do_sample=True,             # Needed for temperature to work
            top_k=50,                   # Sample from top 50 tokens
            top_p=0.95                  # Nucleus sampling
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)


# Sentence embeddings
embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

DATA_DIR = "data"
os.makedirs(DATA_DIR, exist_ok=True)

# Save uploaded PDF and index
def save_pdf(file, title):
    folder = os.path.join(DATA_DIR, title.strip())
    if os.path.exists(folder):
        return f"'{title}' already exists. Use a different title."

    os.makedirs(folder, exist_ok=True)

    chunks = []
    page_numbers = []
    with pdfplumber.open(file.name) as pdf:
        for i, page in enumerate(pdf.pages):
            text = page.extract_text()
            if text:
                for j in range(0, len(text), 500):
                    chunk = text[j:j+500]
                    chunks.append(chunk)
                    page_numbers.append(i + 1)

    embeddings = embedder.encode(chunks)

    if len(embeddings.shape) != 2:
        raise ValueError(f"Expected 2D embeddings, got shape {embeddings.shape}")

    index = faiss.IndexFlatL2(embeddings.shape[1])
    index.add(embeddings)

    index_path = os.path.join(folder, "index.faiss")
    chunks_path = os.path.join(folder, "chunks.pkl")
    faiss.write_index(index, index_path)
    with open(chunks_path, "wb") as f:
        pickle.dump({"chunks": chunks, "page_numbers": page_numbers}, f)

    return f"✅ Saved and indexed '{title}'. You can now ask questions."

def list_titles():
    return [d for d in os.listdir(DATA_DIR) if os.path.isdir(os.path.join(DATA_DIR, d))]

# Ask question using PDF context
def ask_question(message, history, selected_titles):
    if not selected_titles:
        return "❗ Please select at least one PDF."

    combined_answer = ""
    for title in selected_titles:
        folder = os.path.join(DATA_DIR, title)
        try:
            index = faiss.read_index(os.path.join(folder, "index.faiss"))
            with open(os.path.join(folder, "chunks.pkl"), "rb") as f:
                data = pickle.load(f)
                chunks = data["chunks"]
                page_numbers = data["page_numbers"]

            q_embed = embedder.encode([message])
            D, I = index.search(q_embed, k=3)
            context = "\n".join([
                f"(Page {page_numbers[i]}): {chunks[i]}" for i in I[0]
            ])

            #prompt = f"""Answer the question using only the context below.\n\nContext:\n{context}\n\nQuestion: {message}"""
            prompt = f"""You are a helpful assistant. Provide a thorough and detailed answer to the following question using only the context.

                Context:
                {context}
                
                Question: {message}
                
                Answer in detail:
                """

            response = generate_answer(prompt)

            combined_answer += f"**{title}**:\n{response.strip()}\n\n"
        except Exception as e:
            combined_answer += f"⚠️ Error with {title}: {str(e)}\n\n"

    return combined_answer.strip()

# Gradio UI
with gr.Blocks() as demo:
    with gr.Tab("📄 Upload PDF"):
        file = gr.File(label="PDF File", file_types=[".pdf"])
        title = gr.Textbox(label="Title for PDF")
        upload_btn = gr.Button("Upload and Index")
        upload_status = gr.Textbox(label="Status")
        upload_btn.click(fn=save_pdf, inputs=[file, title], outputs=upload_status)

    with gr.Tab("💭 Chat with PDFs"):
        pdf_selector = gr.CheckboxGroup(label="Select PDFs", choices=list_titles())
        refresh_btn = gr.Button("🔄 Refresh PDF List")
        refresh_btn.click(fn=list_titles, outputs=pdf_selector)
        chat = gr.ChatInterface(fn=ask_question, additional_inputs=[pdf_selector])

demo.launch()