app.py

import spaces
import gradio as gr
import fitz  # PyMuPDF
import ocrmypdf
import tempfile
import os

def extract_text_markdown(doc):
    markdown_output = ""
    image_counter = 1

    for page in doc:
        blocks = page.get_text("dict")["blocks"]
        lines_group = []
        elements = []

        for b in blocks:
            y = b["bbox"][1]

            if b["type"] == 0:  # Texto
                for line in b["lines"]:
                    spans = line["spans"]
                    cells = []
                    for span in spans:
                        cells.append((span["bbox"][0], span["text"]))  # (x, texto)
                    if len(cells) > 1:
                        lines_group.append((line["bbox"][1], cells))  # (y, columnas)
                    else:
                        line_text = " ".join(span["text"] for span in spans).strip()
                        if line_text:
                            elements.append((line["bbox"][1], line_text))

            elif b["type"] == 1:  # Imagen
                elements.append((y, f"[imagen_{image_counter}]()"))
                image_counter += 1

        # Detectar tablas rudimentarias
        if len(lines_group) >= 2:
            # Ordenar por coordenada vertical
            lines_group.sort(key=lambda x: x[0])
            header_cells = [cell[1].strip() for cell in lines_group[0][1]]
            markdown_output += "| " + " | ".join(header_cells) + " |\n"
            markdown_output += "| " + " | ".join(["---"] * len(header_cells)) + " |\n"

            for y, row in lines_group[1:]:
                row_cells = [cell[1].strip() for cell in row]
                markdown_output += "| " + " | ".join(row_cells) + " |\n"

            markdown_output += "\n"

        # Agregar líneas de texto sueltas
        elements.sort(key=lambda x: x[0])
        previous_y = None
        for y, content in elements:
            if previous_y is not None and abs(y - previous_y) > 10:
                markdown_output += "\n"
            markdown_output += content + "\n"
            previous_y = y

        markdown_output += "\n---\n\n"

    return markdown_output.strip()

@spaces.GPU
def convert(pdf_file):
    original_doc = fitz.open(pdf_file)
    plain_text = "\n".join([page.get_text() for page in original_doc])

    # Aplicar OCR solo si el PDF no tiene texto
    if len(plain_text.strip()) < 100:
        ocr_temp_path = tempfile.NamedTemporaryFile(suffix=".pdf", delete=False).name
        ocrmypdf.ocr(pdf_file, ocr_temp_path, force_ocr=True)
        doc = fitz.open(ocr_temp_path)
    else:
        doc = original_doc

    markdown = extract_text_markdown(doc)
    metadata = {}  # Si necesitas metadatos, se pueden agregar aquí
    return markdown, metadata

gr.Interface(
    fn=convert,
    inputs=[gr.File(label="Sube tu PDF", type="filepath")],
    outputs=[gr.Text(label="Markdown estructurado"), gr.JSON(label="Metadata")],
).launch()