app.py

import streamlit as st
from transformers import pipeline
import torch
from pdf2image import convert_from_bytes
from PIL import Image

st.set_page_config(page_title="Donut PDF QA", layout="centered")

@st.cache_resource
def load_model():
    return pipeline(
        task="document-question-answering",
        model="naver-clova-ix/donut-base-finetuned-docvqa",
        device=0 if torch.cuda.is_available() else -1,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
    )

qa_pipeline = load_model()


st.title("📄 Donut: PDF Question Answering")

uploaded_file = st.file_uploader("Upload a PDF file", type=["pdf"])
question = st.text_input("Ask a question about the document")

if uploaded_file and question:
    st.write("Reading and converting PDF...")
    images = convert_from_bytes(uploaded_file.read(), dpi=200)

    page_number = st.number_input("Select page", min_value=1, max_value=len(images), value=1, step=1)
    page_image = images[page_number - 1]
    st.image(page_image, caption=f"Page {page_number}")

    with st.spinner("Finding answer..."):
        result = qa_pipeline(image=page_image, question=question)
        st.success("Answer:")
        st.write(result[0]['answer'])