Update app.py

app.py

@@ -2643,6 +2643,88 @@ def _split_fused_date_transaction_header(grid):
         new_grid.append(split_row)
     return new_grid
 
+def _normalize_fused_date_posted_amount_table(grid):
+    """
+    Normalize 2-column OCR tables with fused header like:
+      col1: "Date posted Transaction description [Reference number]"
+      col2: "Amount"
+    into stable 3-column transaction layout:
+      Date | Transaction Description | Amount
+    Strict guard: only activates on that specific fused-header signature.
+    """
+    if not grid or len(grid) < 2:
+        return grid
+
+    ncols = max(len(r) for r in grid if isinstance(r, list))
+    if ncols < 2:
+        return grid
+
+    def _row_cells(r):
+        return [str(c or "").strip() for c in (r + [""] * (ncols - len(r)))]
+
+    hdr_idx = None
+    for ri in range(min(6, len(grid))):
+        cells = _row_cells(grid[ri])
+        row_text = " ".join(cells).lower()
+        if (
+            "date posted transaction description" in row_text
+            and "amount" in row_text
+        ):
+            hdr_idx = ri
+            break
+    if hdr_idx is None:
+        return grid
+
+    date_re = re.compile(r"^\d{1,2}[/-]\d{1,2}(?:[/-]\d{2,4})?$")
+    out = []
+
+    for ri, row in enumerate(grid):
+        cells = _row_cells(row)
+        if ri == hdr_idx:
+            out.append(["Date", "Transaction Description", "Amount"])
+            continue
+        if ri < hdr_idx:
+            # Keep pre-header text as non-transaction row.
+            joined = " ".join(c for c in cells if c).strip()
+            out.append(["", joined, ""])
+            continue
+
+        c1 = cells[0].strip()
+        c2 = cells[1].strip() if len(cells) > 1 else ""
+
+        # Skip duplicated inline sub-headers that often appear in fused OCR tables.
+        lower_c1 = c1.lower()
+        if (
+            lower_c1.startswith("date posted transaction description")
+            or lower_c1 in {"ach additions", "ach deductions", "other additions", "other deductions", "service charges and fees"}
+        ):
+            continue
+
+        date = ""
+        desc = c1
+        if c1:
+            parts = c1.split()
+            if parts and date_re.match(parts[0]):
+                date = parts[0]
+                desc = " ".join(parts[1:]).strip()
+
+        # If amount is empty in column 2, try extracting trailing amount from description.
+        amount = c2
+        if not amount and desc:
+            m = re.search(r"(-?\$?\d{1,3}(?:,\d{3})*(?:\.\d{2})-?)\s*$", desc)
+            if m:
+                amount = m.group(1)
+                desc = desc[:m.start()].strip()
+
+        if not (date or desc or amount):
+            continue
+        out.append([date, desc, amount])
+
+    # Keep at least header + 1 row to avoid degrading clean tables on accidental match.
+    if len(out) < 2:
+        return grid
+    return out
+
 def normalize_html_tables(text: str) -> str:
     """
     For every <table>...</table>:
@@ -2694,6 +2776,10 @@ def normalize_html_tables(text: str) -> str:
             # Fix NF-2: split fused "Date Transaction Detail" header/cell.
             grid = _split_fused_date_transaction_header(grid)
 
+            # Fix 11: normalize fused "Date posted Transaction description ... / Amount"
+            # two-column tables into stable Date/Description/Amount structure.
+            grid = _normalize_fused_date_posted_amount_table(grid)
+
             # Fix 8: split fused multi-keyword header cell (First Horizon Bank)
             grid = _split_fused_multicolumn_header(grid)