Update app.py

app.py

@@ -47,36 +47,36 @@ with open(formatter_path, "w") as f:
 print("✅ result_formatter.py fixed")
 
 # ── STEP 3: Helper functions ──────────────────────────────────
-def get_missed_zones(regions, img_height):
-    y_tops    = []
-    y_bottoms = []
+def get_top_gap(regions, img_height):
+    """
+    Find the top gap the API missed using bbox_2d coordinates.
+    Only extracts header (top gap) — never bottom — to avoid
+    duplicating body content the API already captured correctly.
+    """
+    y_tops = []
     for r in regions:
         bbox = r.get("bbox_2d")
         if bbox and len(bbox) == 4:
             y_tops.append(bbox[1])
-            y_bottoms.append(bbox[3])
     if not y_tops:
-        return None, None
+        return None
     first_y = min(y_tops)
-    last_y  = max(y_bottoms)
-    top_gap    = (first_y / img_height) if first_y > img_height * 0.08 else None
-    bottom_gap = (last_y  / img_height) if last_y  < img_height * 0.92 else None
-    return top_gap, bottom_gap
+    return (first_y / img_height) if first_y > img_height * 0.08 else None
 
 
 def extract_zone_text(pdf_path, page_num, y_start_frac, y_end_frac):
-    import fitz
+    """Extract text from a vertical slice of a PDF page."""
+    import pymupdf as fitz
     doc  = fitz.open(pdf_path)
     page = doc[page_num]
-    h    = page.rect.height
-    w    = page.rect.width
-    rect = fitz.Rect(0, h * y_start_frac, w, h * y_end_frac)
-    text = page.get_text(clip=rect).strip()
+    h, w = page.rect.height, page.rect.width
+    text = page.get_text(clip=fitz.Rect(0, h * y_start_frac, w, h * y_end_frac)).strip()
     doc.close()
     return text
 
 
 def get_page_data(page_result):
+    """Extract markdown and raw regions from one PipelineResult."""
     regions = []
     if hasattr(page_result, "json_result"):
         jr = page_result.json_result
@@ -97,12 +97,12 @@ def run_ocr(uploaded_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")
 
         if is_pdf:
-            import fitz
             doc          = fitz.open(path)
             page_images  = []
             page_heights = []
@@ -111,6 +111,7 @@ def run_ocr(uploaded_file):
                 img_path = f"/tmp/maas_page_{i}.png"
                 pix.save(img_path)
                 page_images.append(img_path)
+                # Image height at 1.5x matches bbox_2d coordinate space
                 page_heights.append(doc[i].rect.height * 1.5)
             doc.close()
             results = parse(page_images)
@@ -129,23 +130,20 @@ def run_ocr(uploaded_file):
             parts = []
 
             if is_pdf and page_num < len(page_heights):
-                top_gap, bottom_gap = get_missed_zones(regions, page_heights[page_num])
+                top_gap = get_top_gap(regions, page_heights[page_num])
 
-                # HEADER at TOP
+                # HEADER at TOP — from PDF text layer (API misses this zone)
                 if top_gap is not None:
                     hdr = extract_zone_text(path, page_num, 0, top_gap)
                     if hdr:
                         parts.append(hdr)
 
-                # BODY in MIDDLE
+                # BODY — from API (includes all content + footer text correctly)
                 if page_md:
                     parts.append(page_md)
 
-                # FOOTER at BOTTOM
-                if bottom_gap is not None:
-                    ftr = extract_zone_text(path, page_num, bottom_gap, 1.0)
-                    if ftr:
-                        parts.append(ftr)
+                # No bottom zone extraction — API already captures footer content.
+                # Extracting bottom zone causes duplication of body content.
             else:
                 if page_md:
                     parts.append(page_md)
@@ -153,6 +151,7 @@ def run_ocr(uploaded_file):
             if parts:
                 all_pages_md.append("\n\n".join(parts))
 
+        # Exactly like original SDK output format
         final = "\n\n---\n\n".join(all_pages_md) if all_pages_md else "(No content)"
         return final
 
@@ -166,7 +165,7 @@ with gr.Blocks(title="GLM-OCR") as demo:
     gr.Markdown("""
     # 🔍 GLM-OCR
     Upload a PDF or image to extract text.
-    Headers and footers included in correct positions.
+    Headers included at top of each page.
     """)
 
     file_input = gr.File(