Update app.py

app.py

@@ -1,3 +1,45 @@
+Hugging Face's logo
+Hugging Face
+Models
+Datasets
+Spaces
+Community
+Docs
+Enterprise
+Pricing
+
+
+Spaces:
+SimpleCodeAI
+/
+glm-ocr-fixed
+
+
+like
+0
+
+Logs
+App
+Files
+Community
+Settings
+glm-ocr-fixed
+/
+app.py
+
+hoytshao's picture
+hoytshao
+Update app.py
+f9f6641
+verified
+5 days ago
+raw
+
+Copy download link
+history
+blame
+
+12.7 kB
 """
 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.
@@ -40,6 +82,19 @@ MIN_CROP_PIXELS = int(os.environ.get("GLMOCR_MIN_CROP_PIXELS", "12544"))  # 112*
 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
 # ---------------------------------------------------------------------------
@@ -177,8 +232,8 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
         os.close(fd)
         try:
             crop.save(path, "JPEG", quality=92)
-            from glmocr import parse
-            out = parse(path)
+            parser = get_parser()
+            out = parser.parse(path)
             if not isinstance(out, list):
                 out = [out]
             if out and getattr(out[0], "markdown_result", None):
@@ -217,10 +272,10 @@ def run_ocr(uploaded_file):
         return "Please upload a file."
     try:
         import pymupdf as fitz
-        from glmocr import parse
 
         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)
@@ -231,10 +286,10 @@ def run_ocr(uploaded_file):
                 pix.save(img_path)
                 page_images.append(img_path)
             doc.close()
-            results = parse(page_images)
+            results = parser.parse(page_images)
         else:
             page_images = [path]
-            results = parse(path)
+            results = parser.parse(path)
 
         if not isinstance(results, list):
             results = [results]
@@ -270,8 +325,8 @@ def run_ocr(uploaded_file):
             if page_md:
                 parts.append(page_md)
 
-            # Only run footer band if API actually detected a footer region via bbox
-            if footer_start_frac is not None and page_num < len(page_images):
+            # Always run footer band (bottom 8–12% of page)
+            if page_num < len(page_images):
                 img_path = page_images[page_num]
                 ftr = ""
                 if is_pdf: