Update app.py

Add log to compare pymupdf result and parsed result

app.py

@@ -44,6 +44,9 @@ 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
@@ -132,7 +135,20 @@ def get_header_footer_zones(regions, norm_height=1000):
     return he_frac, fs_frac
 
 
+def _log_text(prefix, text, max_display=400):
+    """Log extracted text for comparison with OCR; truncate long content."""
+    if not text or not text.strip():
+        log.info("%s: (empty)", prefix)
+        return
+    s = text.strip()
+    if len(s) <= max_display:
+        log.info("%s: %d chars | %s", prefix, len(s), s)
+    else:
+        log.info("%s: %d chars | %s ...", prefix, len(s), s[:max_display].rstrip())
+
+
 def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
+    """Extract text from a horizontal band using a clip rect."""
     try:
         import pymupdf as fitz
         doc = fitz.open(pdf_path)
@@ -141,13 +157,52 @@ def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
         rect = fitz.Rect(0, h * y_start_frac, w, h * y_end_frac)
         text = page.get_text(clip=rect).strip()
         doc.close()
-        log.debug("extract_zone_text_pdf page=%s y=%.2f-%.2f -> %d chars", page_num, y_start_frac, y_end_frac, len(text))
+        _log_text("pymupdf clip page=%s y=%.2f-%.2f" % (page_num, y_start_frac, y_end_frac), text)
         return text
     except Exception as e:
         log.debug("extract_zone_text_pdf failed: %s", e)
         return ""
 
 
+def log_pymupdf_full_page(pdf_path, page_num):
+    """Log full-page PyMuPDF text for comparison (length + prefix)."""
+    try:
+        import pymupdf as fitz
+        doc = fitz.open(pdf_path)
+        page = doc[page_num]
+        text = page.get_text().strip()
+        doc.close()
+        _log_text("pymupdf full page=%s" % page_num, text, max_display=500)
+    except Exception as e:
+        log.debug("log_pymupdf_full_page failed: %s", e)
+
+
+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).
+    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)
+        page = doc[page_num]
+        h = page.rect.height
+        y_lo = h * y_start_frac
+        y_hi = h * y_end_frac
+        words = page.get_text("words")  # list of (x0, y0, x1, y1, word, ...)
+        doc.close()
+        parts = []
+        for w in words:
+            if len(w) >= 5:
+                y0, y1 = float(w[1]), float(w[3])
+                if y0 < y_hi and y1 > y_lo:
+                    parts.append(w[4])
+        text = " ".join(parts).strip()
+        _log_text("pymupdf band page=%s y=%.2f-%.2f" % (page_num, y_start_frac, y_end_frac), text)
+        return text
+    except Exception as e:
+        log.debug("extract_pdf_text_in_band failed: %s", e)
+        return ""
+
+
 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 instead of skipping."""
     zone_name = "header" if y_end_frac < 0.5 else "footer"
@@ -187,6 +242,7 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
             if out and getattr(out[0], "markdown_result", None):
                 text = (out[0].markdown_result or "").strip()
                 log.info("[%s] ocr_zone: crop %dx%d -> %d chars", zone_name, cw, ch, len(text))
+                _log_text("ocr_zone %s result" % zone_name, text)
                 return text
             log.info("[%s] ocr_zone: crop %dx%d -> empty result from API", zone_name, cw, ch)
         finally:
@@ -248,6 +304,8 @@ def run_ocr(uploaded_file):
 
         all_pages = []
         for page_num, page_result in enumerate(results):
+            if is_pdf:
+                log_pymupdf_full_page(path, page_num)
             page_md, regions = get_page_md_and_regions(page_result)
             header_end_frac, footer_start_frac = get_header_footer_zones(regions, 1000)
 
@@ -272,14 +330,17 @@ def run_ocr(uploaded_file):
                 hdr = ""
                 if is_pdf:
                     hdr = extract_zone_text_pdf(path, page_num, 0, he)
+                    if not (hdr and hdr.strip()):
+                        hdr = extract_pdf_text_in_band(path, page_num, 0, PDF_HEADER_BAND_FRAC)
                     if hdr and hdr.strip():
-                        log.info("run_ocr: page %d header from PDF extract: %d chars", page_num, len(hdr.strip()))
+                        log.info("run_ocr: page %d header from PDF: %d chars", page_num, len(hdr.strip()))
                 if not (hdr and hdr.strip()):
                     hdr = ocr_zone(img_path, 0, he)
                 if hdr and hdr.strip():
                     parts.append(hdr.strip())
+                    _log_text("run_ocr page %d final header (used in output)" % page_num, hdr.strip())
                 elif not (hdr and hdr.strip()) and he > 0:
-                    log.info("run_ocr: page %d header empty (PDF extract + ocr_zone)", page_num)
+                    log.info("run_ocr: page %d header empty (PDF + ocr_zone)", page_num)
 
             if page_md:
                 parts.append(page_md)
@@ -290,14 +351,17 @@ def run_ocr(uploaded_file):
                 ftr = ""
                 if is_pdf:
                     ftr = extract_zone_text_pdf(path, page_num, fs, 1.0)
+                    if not (ftr and ftr.strip()):
+                        ftr = extract_pdf_text_in_band(path, page_num, PDF_FOOTER_BAND_FRAC, 1.0)
                     if ftr and ftr.strip():
-                        log.info("run_ocr: page %d footer from PDF extract: %d chars", page_num, len(ftr.strip()))
+                        log.info("run_ocr: page %d footer from PDF: %d chars", page_num, len(ftr.strip()))
                 if not (ftr and ftr.strip()):
                     ftr = ocr_zone(img_path, fs, 1.0)
                 if ftr and ftr.strip():
                     parts.append(ftr.strip())
+                    _log_text("run_ocr page %d final footer (used in output)" % page_num, ftr.strip())
                 elif not (ftr and ftr.strip()) and fs < 1.0:
-                    log.info("run_ocr: page %d footer empty (PDF extract + ocr_zone)", page_num)
+                    log.info("run_ocr: page %d footer empty (PDF + ocr_zone)", page_num)
 
             if parts:
                 all_pages.append("\n\n".join(parts))