fix: downgrade paddlepaddle to 2.6.2 for older CPU AVX compatibility

extractor.py

@@ -110,7 +110,19 @@ def detect_regions(page_png_path):
     from surya.layout import batch_layout_detection
     img = Image.open(page_png_path).convert("RGB")
 
-    layout_predictions = batch_layout_detection([img], model, processor)
+    orig_w, orig_h = img.size
+    scale_factor = 1.0
+    # Surya has a numpy int32 overflow bug on Windows for very large images.
+    # Downscale for layout detection (Surya internally downscales anyway).
+    if max(orig_w, orig_h) > 2048:
+        scale_factor = 2048.0 / max(orig_w, orig_h)
+        new_w = int(orig_w * scale_factor)
+        new_h = int(orig_h * scale_factor)
+        img_for_surya = img.resize((new_w, new_h), Image.LANCZOS)
+    else:
+        img_for_surya = img
+
+    layout_predictions = batch_layout_detection([img_for_surya], model, processor)
 
     regions = []
 
@@ -131,7 +143,8 @@ def detect_regions(page_png_path):
         layout = layout_predictions[0]
         if hasattr(layout, 'bboxes'):
             for rid, box in enumerate(layout.bboxes, start=1):
-                bbox = [int(v) for v in box.bbox]
+                # Scale coordinates back up to original image dimensions
+                bbox = [int(v / scale_factor) for v in box.bbox]
                 raw_label = box.label
                 label = label_map.get(raw_label, "text").lower()
                 # Surya doesn't expose confidence in this API level easily, default to 1.0

requirements.txt

@@ -3,7 +3,7 @@ PyMuPDF>=1.23
 Pillow>=10.0
 numpy>=1.24
 paddleocr==3.5.0
-paddlepaddle==3.0.0
+paddlepaddle==2.6.2
 pytesseract>=0.3.10
 anthropic>=0.40
 surya-ocr==0.4.15