app.py

"""
GLM-OCR Hugging Face Space app with client-side header/footer fallback for MaaS.
Works for every PDF and every image: uses API bboxes when available, else minimal band.
"""
# Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
import asyncio
try:
    _orig_close = asyncio.BaseEventLoop.close
    def _safe_close(self):
        try:
            _orig_close(self)
        except (ValueError, OSError):
            pass
    asyncio.BaseEventLoop.close = _safe_close
except Exception:
    pass

import logging
import os
import re
import tempfile
import yaml
import gradio as gr

import glmocr

log = logging.getLogger("glmocr_app")

GLMOCR_BASE = os.path.dirname(glmocr.__file__)
CONFIG_PATH = os.path.join(GLMOCR_BASE, "config.yaml")
FORMATTER_PATH = os.path.join(GLMOCR_BASE, "postprocess", "result_formatter.py")

# When MaaS is enabled, the cloud API ignores local layout config and does not return
# header/footer regions. We always run client-side OCR on top/bottom bands as fallback.
DEFAULT_ZONE_FRAC = float(os.environ.get("GLMOCR_DEFAULT_ZONE_FRAC", "0.12"))
# MaaS rejects very small images (400). Only send crops that meet minimum dimensions.
MIN_CROP_HEIGHT = int(os.environ.get("GLMOCR_MIN_CROP_HEIGHT", "112"))
MIN_CROP_PIXELS = int(os.environ.get("GLMOCR_MIN_CROP_PIXELS", "12544"))  # 112*112
# Wider PDF bands for position-based extraction (capture account numbers etc. in top/bottom)
PDF_HEADER_BAND_FRAC = float(os.environ.get("GLMOCR_PDF_HEADER_BAND", "0.15"))   # top 15%
PDF_FOOTER_BAND_FRAC = float(os.environ.get("GLMOCR_PDF_FOOTER_BAND", "0.85"))   # bottom 15%

# Single shared parser to avoid "GLM-OCR initialized" per request and asyncio cleanup issues.
_parser = None

def get_parser():
    global _parser
    if _parser is None:
        from glmocr import GlmOcr
        _parser = GlmOcr(
            api_key=os.environ.get("GLMOCR_API_KEY", "4570c28bdea5493c9efae9dae68edc66.sGbA9DLlcX1GlvqV"),
            mode="maas",
        )
    return _parser

# ---------------------------------------------------------------------------
# 1. Config: set MaaS and optionally include headers/footers for non-MaaS
# ---------------------------------------------------------------------------
try:
    with open(CONFIG_PATH, "r") as f:
        config = yaml.safe_load(f)
    config["pipeline"]["maas"]["enabled"] = True
    config["pipeline"]["maas"]["api_key"] = os.environ.get("GLMOCR_API_KEY", "4570c28bdea5493c9efae9dae68edc66.sGbA9DLlcX1GlvqV")

    to_include_as_text = {"header", "footer"}
    layout_section = config.get("pipeline", {}).get("layout", {})
    if "label_task_mapping" in layout_section:
        mapping = layout_section["label_task_mapping"]
        abandon = mapping.get("abandon")
        if isinstance(abandon, list):
            mapping["abandon"] = [x for x in abandon if x not in to_include_as_text]
        text_labels = mapping.get("text")
        if isinstance(text_labels, list):
            for label in to_include_as_text:
                if label not in text_labels:
                    text_labels.append(label)

    formatter_section = config.get("pipeline", {}).get("result_formatter", {})
    if "abandon" in formatter_section:
        abandon = formatter_section["abandon"]
        if isinstance(abandon, list):
            formatter_section["abandon"] = [x for x in abandon if x not in to_include_as_text]
    if "label_visualization_mapping" in formatter_section:
        text_vis = formatter_section["label_visualization_mapping"].get("text")
        if isinstance(text_vis, list):
            for label in to_include_as_text:
                if label not in text_vis:
                    text_vis.append(label)

    with open(CONFIG_PATH, "w") as f:
        yaml.dump(config, f, default_flow_style=False, sort_keys=False)
except Exception:
    pass  # e.g. read-only package on Hugging Face; MaaS ignores layout anyway

# ---------------------------------------------------------------------------
# 2. result_formatter: stop stripping header/footer (best-effort; may be read-only)
# ---------------------------------------------------------------------------
try:
    with open(FORMATTER_PATH, "r") as f:
        source = f.read()
    for label in ('"header"', "'header'", '"footer"', "'footer'", '"doc_header"', "'doc_header'", '"doc_footer"', "'doc_footer'"):
        source = re.sub(r",\s*" + re.escape(label), "", source)
        source = re.sub(re.escape(label) + r"\s*,", "", source)
        source = re.sub(re.escape(label), "", source)
    with open(FORMATTER_PATH, "w") as f:
        f.write(source)
except Exception:
    pass


def get_header_footer_zones(regions, norm_height=1000):
    if not regions:
        return None, None
    y_tops, y_bottoms = [], []
    for r in regions:
        bbox = r.get("bbox_2d") if isinstance(r, dict) else getattr(r, "bbox_2d", None)
        if bbox and len(bbox) >= 4:
            y_tops.append(bbox[1])
            y_bottoms.append(bbox[3])
    if not y_tops:
        return None, None
    return min(y_tops) / norm_height, max(y_bottoms) / norm_height


def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
    """Extract text from a horizontal band using a clip rect."""
    try:
        import pymupdf as fitz
        doc = fitz.open(pdf_path)
        page = doc[page_num]
        h, w = page.rect.height, page.rect.width
        rect = fitz.Rect(0, h * y_start_frac, w, h * y_end_frac)
        text = page.get_text(clip=rect).strip()
        doc.close()
        return text
    except Exception:
        return ""


def extract_pdf_text_in_band(pdf_path, page_num, y_start_frac, y_end_frac):
    """Extract all text whose word bbox falls in the given vertical band (by position).
    Use a wider band (e.g. top 15% / bottom 15%) to capture header/footer that clip might miss."""
    try:
        import pymupdf as fitz
        doc = fitz.open(pdf_path)
        page = doc[page_num]
        h = page.rect.height
        y_lo = h * y_start_frac
        y_hi = h * y_end_frac
        words = page.get_text("words")  # list of (x0, y0, x1, y1, word, ...)
        doc.close()
        parts = []
        for w in words:
            if len(w) >= 5:
                y0, y1 = float(w[1]), float(w[3])
                if y0 < y_hi and y1 > y_lo:
                    parts.append(w[4])
        return " ".join(parts).strip()
    except Exception:
        return ""


def ocr_zone(image_path, y_start_frac, y_end_frac):
    """Run OCR on a horizontal band. Pads small crops to meet API minimum size instead of skipping."""
    zone_name = "header" if y_end_frac < 0.5 else "footer"
    try:
        from PIL import Image
        img = Image.open(image_path).convert("RGB")
        w, h = img.size
        y0 = max(0, int(h * y_start_frac))
        y1 = min(h, int(h * y_end_frac))
        if y1 <= y0:
            return ""
        crop = img.crop((0, y0, w, y1))
        cw, ch = crop.size
        # Pad to minimum size so MaaS accepts the image (avoids 400 on tiny strips)
        if ch < MIN_CROP_HEIGHT or (cw * ch) < MIN_CROP_PIXELS:
            need_h = max(ch, MIN_CROP_HEIGHT)
            need_w = max(cw, 1)
            if (need_w * need_h) < MIN_CROP_PIXELS:
                need_w = max(need_w, (MIN_CROP_PIXELS + need_h - 1) // need_h)
            canvas = Image.new("RGB", (need_w, need_h), (255, 255, 255))
            if zone_name == "header":
                canvas.paste(crop, (0, 0))  # crop at top
            else:
                canvas.paste(crop, (0, need_h - ch))  # crop at bottom
            crop = canvas
            cw, ch = crop.size
        fd, path = tempfile.mkstemp(suffix=".jpg")
        os.close(fd)
        try:
            crop.save(path, "JPEG", quality=92)
            parser = get_parser()
            out = parser.parse(path)
            if not isinstance(out, list):
                out = [out]
            if out and getattr(out[0], "markdown_result", None):
                text = (out[0].markdown_result or "").strip()
                return text
        finally:
            try:
                os.unlink(path)
            except Exception:
                pass
    except Exception as e:
        log.warning("[%s] ocr_zone failed: %s", zone_name, e, exc_info=True)
    return ""


def get_page_md_and_regions(page_result):
    md = ""
    if hasattr(page_result, "markdown_result") and page_result.markdown_result:
        md = (page_result.markdown_result or "").strip()
    regions = []
    if hasattr(page_result, "json_result"):
        jr = page_result.json_result
        if isinstance(jr, dict) and "regions" in jr:
            regions = jr.get("regions") or []
        elif isinstance(jr, list) and len(jr) > 0:
            r = jr[0] if isinstance(jr[0], list) else jr
            if isinstance(r, list):
                regions = r
            elif isinstance(r, dict) and "regions" in r:
                regions = r.get("regions") or []
    return md, regions


def run_ocr(uploaded_file):
    if uploaded_file is None:
        return "Please upload a file."
    try:
        import pymupdf as fitz

        path = uploaded_file.name if hasattr(uploaded_file, "name") else str(uploaded_file)
        is_pdf = path.lower().endswith(".pdf")
        parser = get_parser()

        if is_pdf:
            doc = fitz.open(path)
            page_images = []
            for i in range(len(doc)):
                pix = doc[i].get_pixmap(matrix=fitz.Matrix(1.5, 1.5), alpha=False)
                img_path = os.path.join(tempfile.gettempdir(), f"maas_page_{i}.png")
                pix.save(img_path)
                page_images.append(img_path)
            doc.close()
            results = parser.parse(page_images)
        else:
            page_images = [path]
            results = parser.parse(path)

        if not isinstance(results, list):
            results = [results]

        all_pages = []
        for page_num, page_result in enumerate(results):
            page_md, regions = get_page_md_and_regions(page_result)
            header_end_frac, footer_start_frac = get_header_footer_zones(regions, 1000)

            # MaaS does not return header/footer regions; always use at least default bands
            he = header_end_frac if header_end_frac is not None else DEFAULT_ZONE_FRAC
            fs = footer_start_frac if footer_start_frac is not None else (1.0 - DEFAULT_ZONE_FRAC)
            if he <= 0:
                he = DEFAULT_ZONE_FRAC  # ensure we always OCR a top band
            if fs >= 1.0:
                fs = 1.0 - DEFAULT_ZONE_FRAC  # ensure we always OCR a bottom band

            parts = []

            # Always run header band (top 8–12% of page)
            if page_num < len(page_images):
                img_path = page_images[page_num]
                hdr = ""
                if is_pdf:
                    hdr = extract_zone_text_pdf(path, page_num, 0, he)
                    if not (hdr and hdr.strip()):
                        hdr = extract_pdf_text_in_band(path, page_num, 0, PDF_HEADER_BAND_FRAC)
                if not (hdr and hdr.strip()):
                    hdr = ocr_zone(img_path, 0, he)
                if hdr and hdr.strip():
                    parts.append(hdr.strip())

            if page_md:
                parts.append(page_md)

            # Only run footer band if API actually detected a footer region via bbox
            # (footer_start_frac is None when MaaS finds no footer — avoids grabbing body content)
            if footer_start_frac is not None and page_num < len(page_images):
                img_path = page_images[page_num]
                ftr = ""
                if is_pdf:
                    ftr = extract_zone_text_pdf(path, page_num, fs, 1.0)
                    if not (ftr and ftr.strip()):
                        ftr = extract_pdf_text_in_band(path, page_num, PDF_FOOTER_BAND_FRAC, 1.0)
                if not (ftr and ftr.strip()):
                    ftr = ocr_zone(img_path, fs, 1.0)
                if ftr and ftr.strip():
                    parts.append(ftr.strip())

            if parts:
                all_pages.append("\n\n".join(parts))

        return "\n\n---page-separator---\n\n".join(all_pages) if all_pages else "(No content)"
    except Exception as e:
        import traceback
        log.exception("run_ocr failed: %s", e)
        return f"Error: {e}\n\n{traceback.format_exc()}"


with gr.Blocks(title="GLM-OCR") as demo:
    gr.Markdown("# GLM-OCR\nUpload a PDF or image. Headers and footers included.")
    file_in = gr.File(label="Upload PDF or image", file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp"])
    run_btn = gr.Button("Run OCR", variant="primary")
    out = gr.Textbox(lines=40, label="Output")
    run_btn.click(fn=run_ocr, inputs=file_in, outputs=out)

if __name__ == "__main__":
    demo.launch()