Update app.py

app.py

@@ -30,12 +30,17 @@ 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():
@@ -133,7 +138,8 @@ def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
 
 
 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)."""
+    """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)
@@ -155,7 +161,7 @@ def extract_pdf_text_in_band(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. Pads small crops to meet API minimum size."""
+    """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
@@ -167,6 +173,7 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
             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)
@@ -174,9 +181,9 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
                 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))
+                canvas.paste(crop, (0, 0))  # crop at top
             else:
-                canvas.paste(crop, (0, need_h - ch))
+                canvas.paste(crop, (0, need_h - ch))  # crop at bottom
             crop = canvas
             cw, ch = crop.size
         fd, path = tempfile.mkstemp(suffix=".jpg")
@@ -218,6 +225,14 @@ def get_page_md_and_regions(page_result):
     return md, regions
 
 
+def is_duplicate(text, existing_text):
+    """Check if text is already contained in existing_text using first 100 chars as fingerprint."""
+    if not text or not existing_text:
+        return False
+    fingerprint = text[:100].strip()
+    return fingerprint in existing_text
+
+
 def run_ocr(uploaded_file):
     if uploaded_file is None:
         return "Please upload a file."
@@ -250,16 +265,17 @@ def run_ocr(uploaded_file):
             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
+                he = DEFAULT_ZONE_FRAC  # ensure we always OCR a top band
             if fs >= 1.0:
-                fs = 1.0 - DEFAULT_ZONE_FRAC
+                fs = 1.0 - DEFAULT_ZONE_FRAC  # ensure we always OCR a bottom band
 
             parts = []
 
-            # Always run header band (top 8-12% of page)
+            # Always run header band (top 8–12% of page)
             if page_num < len(page_images):
                 img_path = page_images[page_num]
                 hdr = ""
@@ -275,7 +291,7 @@ def run_ocr(uploaded_file):
             if page_md:
                 parts.append(page_md)
 
-            # Always run footer band (bottom 8-12% of page)
+            # Always run footer band (bottom 8–12% of page)
             if page_num < len(page_images):
                 img_path = page_images[page_num]
                 ftr = ""
@@ -285,7 +301,8 @@ def run_ocr(uploaded_file):
                         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():
+                # Only append footer if NOT already present in page_md (avoids duplication)
+                if ftr and ftr.strip() and not is_duplicate(ftr, page_md):
                     parts.append(ftr.strip())
 
             if parts: