Update app.py

app.py

@@ -25,6 +25,12 @@ Included (data-driven, no institution names):
 
 Configure GLMOCR_API_KEY (environment variable). Optional: glmocr + gradio +
 pymupdf + pillow installed.
+
+When GLMOCR_PREFER_PDF_TEXT_BODY is 1 (default), searchable PDFs skip vision OCR
+for the body and return each page's embedded text so output matches the PDF text
+layer. Set GLMOCR_PREFER_PDF_TEXT_BODY=0 to force the image OCR path. Optional
+GLMOCR_MIN_PDF_BODY_CHARS (default 120) is the minimum characters per page required
+to use the text-layer path for the whole document.
 """
 
 # Patch asyncio first (before Gradio imports it) to reduce Python 3.13 loop noise
@@ -109,6 +115,8 @@ PAGE_PNG_COMPRESS_LEVEL = 3
 # JPEG quality for small header/footer crops sent to the API.
 ZONE_JPEG_QUALITY = 95
 
+MIN_PDF_TEXT_CHARS_NATIVE_LAYER = 1500
+
 _parser = None
 
 
@@ -186,6 +194,41 @@ def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
         return ""
 
 
+def _prefer_pdf_text_body() -> bool:
+    v = os.environ.get("GLMOCR_PREFER_PDF_TEXT_BODY", "1").strip().lower()
+    return v in ("1", "true", "yes", "")
+
+
+def extract_pdf_body_text_all_pages_if_suitable(pdf_path: str) -> Optional[List[str]]:
+    try:
+        import pymupdf as fitz
+    except ImportError:
+        return None
+    try:
+        min_c = int(os.environ.get("GLMOCR_MIN_PDF_BODY_CHARS", "120"))
+    except ValueError:
+        min_c = 120
+    try:
+        doc = fitz.open(pdf_path)
+    except Exception:
+        return None
+    try:
+        if len(doc) < 1:
+            return None
+        out: List[str] = []
+        for i in range(len(doc)):
+            t = (doc[i].get_text("text") or "").strip()
+            if len(t) < min_c:
+                return None
+            out.append(t)
+        return out
+    finally:
+        try:
+            doc.close()
+        except Exception:
+            pass
+
+
 def extract_pdf_text_in_band(pdf_path, page_num, y_start_frac, y_end_frac):
     try:
         import pymupdf as fitz
@@ -613,6 +656,12 @@ def run_ocr(uploaded_file):
     try:
         path = uploaded_file.name if hasattr(uploaded_file, "name") else str(uploaded_file)
         is_pdf = path.lower().endswith(".pdf")
+
+        if is_pdf and _prefer_pdf_text_body():
+            pdf_pages = extract_pdf_body_text_all_pages_if_suitable(path)
+            if pdf_pages is not None:
+                return "\n\n---page-separator---\n\n".join(pdf_pages)
+
         parser = get_parser()
 
         page_heights: List[int] = []
@@ -710,8 +759,8 @@ def _create_gradio_demo():
     with gr.Blocks(title="GLM-OCR (simple)") as demo:
         gr.Markdown(
             "# GLM-OCR (simple)\n"
-            "Upload a PDF or image. Header and footer bands use PDF text when available. "
-            "Body markdown is produced by the vision model and may disagree with the PDF text layer."
+            "Searchable PDFs use embedded page text by default (matches the PDF text layer). "
+            "Otherwise the vision model reads rasterized pages. Images always use vision OCR."
         )
         file_in = gr.File(
             label="Upload PDF or image",