app.py

import gradio as gr
import torch
import base64
import fitz  # PyMuPDF
import tempfile
from io import BytesIO
from PIL import Image
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration

from olmocr.data.renderpdf import render_pdf_to_base64png
from olmocr.prompts.anchor import get_anchor_text

import re
import markdown2
from latex2mathml.converter import convert as latex_to_mathml
import html

model = Qwen2VLForConditionalGeneration.from_pretrained(
    "allenai/olmOCR-7B-0225-preview", torch_dtype=torch.bfloat16
).eval()
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

def convert_latex_to_mathml(text):
    def replacer(match):
        try:
            return f"<math>{latex_to_mathml(match.group(1))}</math>"
        except Exception:
            return html.escape(match.group(0))
    text = re.sub(r'\\\((.*?)\\\)', replacer, text)
    text = re.sub(r'\\\[(.*?)\\\]', replacer, text)
    return text

def clean_page_headers(text):
    lines = text.split("\n")
    cleaned = []
    for line in lines:
        if not re.match(r'^(\s*Page \d+|\s*\d{1,2}\s*/\s*\d{1,2}|^[A-Z][A-Za-z\s]{0,20}$)', line.strip()):
            cleaned.append(line)
    return "\n".join(cleaned)

def process_pdf_to_html(pdf_file, title, author):
    pdf_path = pdf_file.name
    doc = fitz.open(pdf_path)
    num_pages = len(doc)

    toc_entries = doc.get_toc()
    toc_by_page = {}
    for level, header, page in toc_entries:
        toc_by_page.setdefault(page, []).append((level, header))

    all_text = ""
    cover_img_html = ""

    for i in range(num_pages):
        page_num = i + 1
        print(f"Processing page {page_num}...")

        try:
            image_base64 = render_pdf_to_base64png(pdf_path, page_num, target_longest_image_dim=1024)
            anchor_text = get_anchor_text(pdf_path, page_num, pdf_engine="pdfreport", target_length=4000)

            prompt = (
                "Below is the image of one page of a document, as well as some raw textual content that was previously "
                "extracted for it. Just return the plain text representation of this document as if you were reading it naturally.\n"
                "Do not hallucinate.\n"
                "RAW_TEXT_START\n"
                f"{anchor_text}\n"
                "RAW_TEXT_END"
            )

            messages = [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": prompt},
                        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_base64}"}},
                    ],
                }
            ]
            text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
            image = Image.open(BytesIO(base64.b64decode(image_base64)))

            inputs = processor(
                text=[text],
                images=[image],
                padding=True,
                return_tensors="pt",
            )
            inputs = {k: v.to(device) for k, v in inputs.items()}

            output = model.generate(
                **inputs,
                temperature=0.8,
                max_new_tokens=5096,
                num_return_sequences=1,
                do_sample=True,
            )

            prompt_len = inputs["input_ids"].shape[1]
            new_tokens = output[:, prompt_len:].detach().cpu()

            decoded = "[No output generated]"
            if new_tokens.shape[1] > 0:
                decoded_list = processor.tokenizer.batch_decode(new_tokens, skip_special_tokens=True)
                raw_output = decoded_list[0].strip() if decoded_list else "[No output generated]"
                try:
                    parsed = json.loads(raw_output)
                    decoded = parsed.get("natural_text", raw_output)
                except json.JSONDecodeError:
                    decoded = raw_output

        except Exception as e:
            decoded = f"[Error on page {page_num}: {e}]"

        print(f"Decoded content for page {page_num}: {decoded}")

        cleaned_text = clean_page_headers(decoded)
        mathml_converted = convert_latex_to_mathml(cleaned_text)
        markdown_converted = markdown2.markdown(mathml_converted)
        html_page = markdown_converted.replace("\n", "<br>")

        if page_num in toc_by_page:
            for level, header in toc_by_page[page_num]:
                tag = f"h{min(level, 6)}"
                html_page = f"<{tag}>{html.escape(header)}</{tag}>\n" + html_page

        all_text += f"<div>{html_page}</div>\n"

        if page_num == 1:
            cover_img_html = f'<img src="data:image/png;base64,{image_base64}" alt="cover" style="max-width:100%; height:auto;"><hr>'

    mathjax_script = """
    <script type="text/javascript" id="MathJax-script" async
      src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js">
    </script>
    """

    full_html = f"""<!DOCTYPE html>
    <html>
    <head>
        <meta charset="utf-8">
        <title>{html.escape(title)}</title>
        {mathjax_script}
    </head>
    <body>
        <h1>{html.escape(title)}</h1>
        <h3>{html.escape(author)}</h3>
        {cover_img_html}
        {all_text}
    </body>
    </html>
    """

    with tempfile.NamedTemporaryFile(delete=False, suffix=".html", dir="/tmp", mode="w", encoding="utf-8") as tmp:
        tmp.write(full_html)
        return tmp.name

iface = gr.Interface(
    fn=process_pdf_to_html,
    inputs=[
        gr.File(label="Upload PDF", file_types=[".pdf"]),
        gr.Textbox(label="HTML Title"),
        gr.Textbox(label="Author(s)")
    ],
    outputs=gr.File(label="Download HTML"),
    title="PDF to HTML Converter with Structure (olmOCR)",
    description="Extracts text with structure, math, and footnotes using olmOCR and renders to styled HTML.",
    allow_flagging="never"
)

if __name__ == "__main__":
    iface.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=True,
        debug=True,
        allowed_paths=["/tmp"]
    )