Update app.py

app.py

@@ -13,6 +13,8 @@ Primary goals for reconcile rate:
   4) clean common header artifacts (e.g. "DESCRIPTIONBeginning Balance", "BALANCE$3,447.10")
   5) recover fused first data row generically (two-signal guard: text fusion + money fusion must both fire)
   6) promote misplaced header row: when real column headers land in a data row, restructure the grid
+  7) fix fused key-value rows in summary sections (e.g. Interest Summary) appended to transaction tables
+- footer extraction enabled on all pages with deduplication guard (no double-print if body OCR already captured it)
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -68,7 +70,7 @@ ENABLE_CONTRAST = True
 DEFAULT_ZONE_FRAC = 0.12
 PDF_HEADER_BAND_FRAC = 0.10
 
-ENABLE_FOOTER_OCR = False
+ENABLE_FOOTER_OCR = True
 PDF_FOOTER_BAND_FRAC = 0.88
 
 MIN_CROP_HEIGHT = 112
@@ -572,6 +574,7 @@ _HEADER_KW_PATTERNS = [
     r"OPENING\s+BAL\.?",
 ]
 
+
 # Pre-compiled: each pattern anchored at start, case-insensitive
 _HEADER_KW_RES = [
     re.compile(r"^(" + p + r")(.*)", re.IGNORECASE | re.DOTALL)
@@ -823,6 +826,100 @@ def _promote_misplaced_header_row(grid):
     return new_grid
 
 
+
+# --------------------------
+# Fix 3: Fused key-value rows in summary sections
+# --------------------------
+
+# Matches a value suffix that looks like a number or percentage fused onto the
+# end of a label string.  Examples:
+#   "Beginning Interest Rate0.00%"   → label="Beginning Interest Rate"  value="0.00%"
+#   "Number of days in this Period31" → label="Number of days in this Period"  value="31"
+#   "Interest Earned this Period0.00" → label="Interest Earned this Period"    value="0.00"
+# Pattern: label text (non-digit) followed by a numeric/percentage token at the end.
+_FUSED_KV_RE = re.compile(
+    r"^(.+?)\s*(-?\d{1,3}(?:,\d{3})*(?:\.\d+)?%?)$",
+    re.DOTALL,
+)
+
+
+def _fix_fused_keyvalue_rows(grid):
+    """
+    Generic fix for rows where a label and its value are fused into the first
+    cell while all other cells in that row are empty.
+
+    This happens when a bank statement appends a summary section (e.g. Interest
+    Summary, Fee Summary) after the main transaction table without a clear
+    column separator.  The OCR wraps it inside the same table and each line
+    becomes a single-cell row like:
+
+        <td>Beginning Interest Rate0.00%</td> <td></td> <td></td> ...
+
+    The fix:
+      - For every data row where exactly the first cell is non-empty and all
+        other cells are empty, AND the first cell ends with a numeric/percentage
+        token fused directly onto text (no space separator), split it into
+        label (col 0) and value (last column of the table).
+      - If the cell already has a natural space before the value it is NOT
+        fused — leave it alone.  This prevents touching normal description rows
+        that happen to end with a number.
+
+    Guard conditions (no-op if not met — safe for all other tables):
+      - The row must have ALL cells empty except the first.
+      - The first cell must match the fused pattern (text immediately followed
+        by a digit with no whitespace gap before the digit run).
+      - The match must produce a non-empty label AND a non-empty value.
+    """
+    if not grid or len(grid) < 2:
+        return grid
+
+    ncols = len(grid[0])
+    new_grid = [grid[0]]  # keep header unchanged
+
+    for row in grid[1:]:
+        # Guard: only process rows where first cell is non-empty and rest are empty
+        cells = [str(c or "").strip() for c in row]
+        if not cells:
+            new_grid.append(row)
+            continue
+
+        first = cells[0]
+        rest_empty = all(c == "" for c in cells[1:])
+
+        if not rest_empty or not first:
+            new_grid.append(row)
+            continue
+
+        # Guard: fused pattern requires NO whitespace immediately before the
+        # numeric suffix — "Rate0.00%" is fused, "Rate 0.00%" is not.
+        # We check by requiring the character just before the digit run is NOT
+        # a space.
+        fused_match = re.search(r"\S\d", first)
+        if not fused_match:
+            new_grid.append(row)
+            continue
+
+        m = _FUSED_KV_RE.match(first)
+        if not m:
+            new_grid.append(row)
+            continue
+
+        label = m.group(1).strip()
+        value = m.group(2).strip()
+
+        if not label or not value:
+            new_grid.append(row)
+            continue
+
+        # Build the fixed row: label in col 0, value in last col, rest empty
+        new_row = [""] * ncols
+        new_row[0] = label
+        new_row[ncols - 1] = value
+        new_grid.append(new_row)
+
+    return new_grid
+
+
 # --------------------------
 # Normalizer entry point
 # --------------------------
@@ -863,6 +960,9 @@ def normalize_html_tables(text: str) -> str:
             # Fix 2: misplaced header row promotion (keyword-score guard — safe on clean PDFs)
             grid = _promote_misplaced_header_row(grid)
 
+                # Fix 3: fused key-value rows in summary sections (no-space guard — safe on all PDFs)
+            grid = _fix_fused_keyvalue_rows(grid)
+
             out.append(_grid_to_html(grid) if grid else table_html)
         except Exception:
             out.append(table_html)
@@ -1007,7 +1107,12 @@ def run_ocr(uploaded_file):
             if page_md and page_md.strip():
                 parts.append(stabilize_tables_and_text(page_md.strip()))
 
-            # Optional footer
+            # Footer extraction.
+            # Uses the same 3-step fallback as header: PDF text layer -> band words -> OCR crop.
+            # Deduplication guard: if the footer text is already present in the main body OCR
+            # output (accidental capture on some pages), we skip appending it to avoid
+            # printing it twice. Comparison is done on a normalised lowercase snippet of the
+            # first non-empty footer line so minor whitespace differences don't matter.
             if ENABLE_FOOTER_OCR and page_num < len(page_images):
                 ftr = ""
                 if is_pdf:
@@ -1017,7 +1122,18 @@ def run_ocr(uploaded_file):
                 if not (ftr and ftr.strip()):
                     ftr = ocr_zone(page_images[page_num], fs, 1.0)
                 if ftr and ftr.strip():
-                    parts.append(normalize_money_glyphs(ftr.strip()))
+                    ftr_clean = normalize_money_glyphs(ftr.strip())
+                    # Deduplication: check if the first meaningful line of the footer
+                    # already appears anywhere in the assembled page content so far.
+                    ftr_first_line = next(
+                        (ln.strip().lower() for ln in ftr_clean.splitlines() if ln.strip()),
+                        ""
+                    )
+                    already_present = ftr_first_line and any(
+                        ftr_first_line in part.lower() for part in parts
+                    )
+                    if not already_present:
+                        parts.append(ftr_clean)
 
             if parts:
                 all_pages.append("\n\n".join(parts))
@@ -1046,4 +1162,4 @@ with gr.Blocks(title="GLM-OCR") as demo:
     run_btn.click(fn=run_ocr, inputs=file_in, outputs=out)
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()