Update app.py

1. MaaS 400 on header/footer crops
2. Single shared parser (fewer asyncio issues)
3. Asyncio BaseEventLoop.__del__ message

app.py

@@ -17,6 +17,22 @@ 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
+
+# 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
@@ -102,9 +118,9 @@ def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
 
 
 def ocr_zone(image_path, y_start_frac, y_end_frac):
+    """Run OCR on a horizontal band of the image. Skips MaaS if crop is too small (API returns 400)."""
     try:
         from PIL import Image
-        from glmocr import parse
         img = Image.open(image_path).convert("RGB")
         w, h = img.size
         y0 = max(0, int(h * y_start_frac))
@@ -112,11 +128,15 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
         if y1 <= y0:
             return ""
         crop = img.crop((0, y0, w, y1))
-        fd, path = tempfile.mkstemp(suffix=".png")
+        cw, ch = crop.size
+        if ch < MIN_CROP_HEIGHT or (cw * ch) < MIN_CROP_PIXELS:
+            return ""  # MaaS rejects very small images with 400
+        fd, path = tempfile.mkstemp(suffix=".jpg")
         os.close(fd)
-        crop.save(path)
         try:
-            out = parse(path)
+            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):
@@ -153,11 +173,11 @@ def run_ocr(uploaded_file):
     if uploaded_file is None:
         return "Please upload a file."
     try:
-        from glmocr import parse
         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)
@@ -168,10 +188,10 @@ 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]