app.py

import gradio as gr
import torch
from pypdf import PdfReader
from PIL import Image
import io

from transformers import (
    TrOCRProcessor,
    VisionEncoderDecoderModel,
    AutoTokenizer,
    AutoModelForCausalLM
)

# ============================================================
# Device
# ============================================================
device = "cuda" if torch.cuda.is_available() else "cpu"

# ============================================================
# Load Models (cached by HF Spaces)
# ============================================================
ocr_processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-printed")
ocr_model = VisionEncoderDecoderModel.from_pretrained(
    "microsoft/trocr-base-printed"
).to(device)

tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen2.5-1.5B-Instruct",
    trust_remote_code=True
)

qwen_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-1.5B-Instruct",
    device_map="auto",
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    trust_remote_code=True
)

# ============================================================
# Helpers
# ============================================================
def is_scanned_pdf(reader):
    for page in reader.pages:
        if page.extract_text():
            return False
    return True

def extract_text_from_pdf(file):
    reader = PdfReader(file)
    scanned = is_scanned_pdf(reader)

    extracted_text = []

    if not scanned:
        # Digital PDF
        for page in reader.pages:
            text = page.extract_text()
            if text:
                extracted_text.append(text)

    else:
        # OCR only embedded images (HF-safe)
        for page in reader.pages:
            if "/XObject" in page["/Resources"]:
                xobjects = page["/Resources"]["/XObject"].get_object()
                for obj in xobjects:
                    xobj = xobjects[obj]
                    if xobj["/Subtype"] == "/Image":
                        image = Image.open(io.BytesIO(xobj.get_data())).convert("RGB")
                        pixel_values = ocr_processor(
                            images=image,
                            return_tensors="pt"
                        ).pixel_values.to(device)

                        with torch.no_grad():
                            ids = ocr_model.generate(pixel_values)

                        text = ocr_processor.batch_decode(
                            ids,
                            skip_special_tokens=True
                        )[0]
                        extracted_text.append(text)

    return "\n\n".join(extracted_text)

def evaluate_text(text):
    prompt = f"""
You are a strict academic evaluator.

Evaluate the following document and assign marks out of 10.

Criteria:
- Clarity
- Structure
- Technical depth
- Language quality
- Completeness

DOCUMENT:
---------
{text[:6000]}
---------

Respond strictly in this format:

Score: X/10
Justification:
Strengths:
Weaknesses:
"""

    inputs = tokenizer(prompt, return_tensors="pt").to(device)

    with torch.no_grad():
        output = qwen_model.generate(
            **inputs,
            max_new_tokens=400,
            do_sample=False
        )

    return tokenizer.decode(output[0], skip_special_tokens=True)

# ============================================================
# Gradio Function
# ============================================================
def process_pdf(pdf_file):
    extracted_text = extract_text_from_pdf(pdf_file)
    evaluation = evaluate_text(extracted_text)

    return extracted_text, evaluation

# ============================================================
# Gradio UI
# ============================================================
with gr.Blocks(title="PDF Evaluator (OCR + Qwen)") as demo:
    gr.Markdown("""
    # 📄 PDF Evaluator
    Upload a PDF to:
    - Extract text (OCR if needed)
    - Evaluate content using Qwen
    - Get marks out of 10
    """)

    pdf_input = gr.File(label="Upload PDF", file_types=[".pdf"])
    extract_btn = gr.Button("Extract & Evaluate")

    extracted_output = gr.Textbox(
        label="Extracted Text",
        lines=20
    )

    evaluation_output = gr.Textbox(
        label="Evaluation",
        lines=10
    )

    extract_btn.click(
        process_pdf,
        inputs=pdf_input,
        outputs=[extracted_output, evaluation_output]
    )

demo.launch()