Update app.py

app.py

@@ -108,6 +108,11 @@ Primary goals for reconcile rate:
             earlier larger DA3 table — catches OCR-truncated description duplicates such as
             the BoA service-fee page where a VIDDYOZE -1.11 fragment table appears after the
             full service-fee table with a longer description string.
+- Fix 32: Headless amount column — PDFs often leave the rightmost money column without a header;
+            OCR emits an empty <th>. Fill with \"Amount\" when the column is mostly currency.
+            Second pass: collapse Date|Transaction Description|Additions|Subtractions + empty 4th
+            when OCR splits \"Preauth Debit\" / \"Pre-Auth Credit\" into its own column before
+            the merchant text (East West–style) — merge back to Date|Description|Amount.
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -565,6 +570,156 @@ def _merge_blank_header_text_columns(grid):
         new_grid.append([cell for idx, cell in enumerate(row) if idx < len(keep) and keep[idx]])
     return new_grid
 
+
+def _fill_headless_trailing_amount_headers(grid):
+    """
+    Fix 32a: PDF layouts often omit a header above the rightmost amount column; OCR leaves a
+    trailing empty <th>. Downstream parsers then mis-associate columns. Fill only when the
+    column is predominantly strict currency (generic, bank-agnostic).
+    """
+    if not grid or len(grid) < 2:
+        return grid
+
+    header = [str(c or "").strip() for c in grid[0]]
+    ncols = len(header)
+    if ncols < 2:
+        return grid
+
+    i = ncols - 1
+    while i >= 0 and header[i] == "":
+        i -= 1
+    empty_tail_indices = list(range(i + 1, ncols))
+    if not empty_tail_indices:
+        return grid
+
+    body = grid[1:]
+    new_header = list(header)
+
+    for j in empty_tail_indices:
+        total = 0
+        amt_hits = 0
+        for row in body:
+            if j >= len(row):
+                continue
+            total += 1
+            v = str(row[j] or "").strip()
+            if v and _is_amount_like(v):
+                amt_hits += 1
+        if total == 0:
+            continue
+        if (amt_hits / total) < 0.45:
+            continue
+
+        # Right-align column is the numeric amount; PDF often omits a label for it.
+        new_header[j] = "Amount"
+
+    grid[0] = new_header
+    return grid
+
+
+_MMDD_DATE_RE = re.compile(r"^\d{1,2}[-/]\d{2}(?:[-/]\d{2,4})?$")
+
+# Leading transaction-type cell when OCR splits category vs merchant (East West / similar).
+_STACKED_TXN_TYPE_RE = re.compile(
+    r"^(?:Pre-Auth Credit|Pre-Auth Debit|Preauth Debit|"
+    r"Withdrawal Trns|Deposit(?:\s|$)|Outgoing Wire|Incoming Wire|"
+    r"Service Charge|ACH Debit|ACH Credit|ACH Payment|ACH PMT|"
+    r"Wire Transfer|POS Purchase|ATM Withdrawal|Transfer Trns)\b",
+    re.IGNORECASE | re.DOTALL,
+)
+
+
+def _collapse_four_col_stacked_transaction_type_to_three(grid):
+    """
+    Fix 32b: When the PDF is logically Date | one description | amount, OCR sometimes emits
+    four columns: Date | short type (e.g. Preauth Debit) | merchant tail | amount, with an
+    optional empty 4th header cell. Collapse to three columns so markdown matches the PDF row
+    semantics. Requires every non-empty data row to match the pattern (safe no-op otherwise).
+    """
+    if not grid or len(grid) < 3:
+        return grid
+
+    header = [str(c or "").strip() for c in grid[0]]
+    if len(header) != 4:
+        return grid
+
+    h0 = re.sub(r"\s+", " ", header[0]).strip().lower()
+    h1 = re.sub(r"\s+", " ", header[1]).strip().lower()
+    h2 = re.sub(r"\s+", " ", header[2]).strip().lower()
+    h3 = str(header[3] or "").strip().lower()
+
+    if h0 != "date":
+        return grid
+    if "transaction" not in h1 or "description" not in h1:
+        return grid
+    if not any(
+        k in h2
+        for k in (
+            "addition",
+            "subtraction",
+            "credit",
+            "debit",
+            "deposit",
+            "withdrawal",
+        )
+    ):
+        return grid
+    if h3 not in ("", "amount") and not any(
+        k in h3 for k in ("addition", "subtraction", "credit", "debit")
+    ):
+        return grid
+
+    def _row_matches_split_pattern(cells):
+        while len(cells) < 4:
+            cells = cells + [""]
+        c0, c1, c2, c3 = (
+            str(cells[0] or "").strip(),
+            str(cells[1] or "").strip(),
+            str(cells[2] or "").strip(),
+            str(cells[3] or "").strip(),
+        )
+        if not c0 or not _MMDD_DATE_RE.match(c0):
+            return False
+        if not c1 or not _STACKED_TXN_TYPE_RE.match(c1):
+            return False
+        if not c2:
+            return False
+        if _is_amount_like(c2):
+            return False
+        if not _is_amount_like(c3):
+            return False
+        return True
+
+    data_rows = [r for r in grid[1:] if any(str(c or "").strip() for c in r)]
+    if len(data_rows) < 2:
+        return grid
+
+    for row in data_rows:
+        cells = [str(c or "").strip() for c in row]
+        if not _row_matches_split_pattern(cells):
+            return grid
+
+    out = [
+        [
+            header[0],
+            header[1],
+            header[2],
+        ]
+    ]
+    for row in grid[1:]:
+        cells = [str(c or "").strip() for c in row]
+        while len(cells) < 4:
+            cells.append("")
+        if not any(cells):
+            out.append(["", "", ""])
+            continue
+        d0, t1, tail, amt = cells[0], cells[1], cells[2], cells[3]
+        desc = (t1 + " " + tail).strip()
+        out.append([d0, desc, amt])
+
+    return out
+
+
 # --------------------------
 # Shared keyword definitions
 # --------------------------
@@ -2652,6 +2807,9 @@ def normalize_html_tables(text: str) -> str:
             grid = _drop_truly_empty_columns(grid)
             grid = _merge_blank_header_text_columns(grid)
 
+            grid = _fill_headless_trailing_amount_headers(grid)
+            grid = _collapse_four_col_stacked_transaction_type_to_three(grid)
+
             grid = _reconstruct_separator_rows(grid)
 
             grid, recovered_row = _clean_header_artifacts(grid)