Update app.py

app.py

@@ -34,6 +34,10 @@ Primary goals for reconcile rate:
             (same normalized Date|Description|Amount keys), strip that suffix from table N. Fixes
             formats (e.g. Truist) where deposit rows are glued onto the withdrawals table and repeated
             under the proper deposits heading (K>=2; first table must keep >=1 data row).
+- Fix 19: DA3 rows whose Amount cell is not strict currency (e.g. fused card digits + merchant tail)
+            — take the rightmost plausible money token from Description+Amount text for the amount.
+- Fix 20: DA3 in-table dedupe — cap consecutive identical normalized keys at 2 rows (keeps legitimate
+            back-to-back identical charges; still collapses 3+ OCR stutters).
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -3099,9 +3103,12 @@ def normalize_html_tables(text: str) -> str:
             # (Hardin County Bank: "CHASE CREDIT CRD EPAY 8319249882 500.00")
             grid = _extract_fused_desc_amount(grid)
 
+            # Fix 19: junk Amount cell but money token still present in row text (Truist MSBILL/PIN).
+            grid = _repair_da3_garbled_amount_cells(grid)
+
             # Fix 3: fused key-value rows in summary sections (no-space guard — safe on all PDFs)
             grid = _fix_fused_keyvalue_rows(grid)
-            # Fix 16: duplicate data rows inside same Date|Description|Amount table (OCR repeats).
+            # Fix 16/20: duplicate data rows inside same Date|Description|Amount table (OCR repeats).
             grid = _dedupe_duplicate_rows_in_da3_table(grid)
             # Fix 12: split fused summary+daily-balance combined table, if detected.
             split_tables = _split_combined_summary_daily_balance_grid(grid)
@@ -3142,24 +3149,113 @@ def _is_da3_header(grid):
     )
 
 
+_DA3_STRICT_AMOUNT = re.compile(r"^-?\$?\d{1,3}(?:,\d{3})*\.\d{2}-?$")
+_DA3_MONEY_TOKEN = re.compile(
+    r"(?<![\d,])(-?\$?\d{1,3}(?:,\d{3})*\.\d{2}|-?\$?\d{2,6}\.\d{2})(?!\d)"
+)
+
+
+def _repair_da3_garbled_amount_cells(grid):
+    """
+    Fix 19: Amount column shows fused junk (letters, card digits) while the real debit/credit
+    still appears as a money token in the row text — use the rightmost plausible amount.
+    Only for Date|Description|Amount tables and strict MM/DD dates in column 0.
+    """
+    if not grid or len(grid) < 2 or not _is_da3_header(grid):
+        return grid
+
+    date_re = re.compile(r"^\d{1,2}/\d{1,2}(?:/\d{2,4})?$")
+    out = [list(grid[0])]
+    changed = False
+
+    for row in grid[1:]:
+        cells = [str(c or "").strip() for c in row]
+        while len(cells) < 3:
+            cells.append("")
+        c0, c1, c2 = cells[0], cells[1], cells[2]
+        if not date_re.match(c0):
+            out.append(cells)
+            continue
+        c2s = c2.strip()
+        if _DA3_STRICT_AMOUNT.match(c2s):
+            out.append(cells)
+            continue
+        blob = f"{c1} {c2}".strip()
+        if not blob:
+            out.append(cells)
+            continue
+        found = _DA3_MONEY_TOKEN.findall(blob)
+        if not found:
+            out.append(cells)
+            continue
+        pick_raw = None
+        for m in reversed(found):
+            t = m.replace("$", "").replace(",", "")
+            try:
+                v = float(t)
+            except ValueError:
+                continue
+            if 0.01 <= v <= 9_999_999.99:
+                pick_raw = m.replace("$", "").strip()
+                break
+        if not pick_raw:
+            out.append(cells)
+            continue
+        new_desc = c1
+        for suf in (
+            pick_raw,
+            pick_raw.replace(",", ""),
+            f"${pick_raw}",
+            f"${pick_raw.replace(',', '')}",
+        ):
+            if suf and new_desc.rstrip().endswith(suf):
+                new_desc = new_desc[: -len(suf)].rstrip()
+                break
+        out.append([c0, new_desc, pick_raw])
+        changed = True
+
+    return out if changed else grid
+
+
 def _dedupe_duplicate_rows_in_da3_table(grid):
-    """Drop repeated data rows within one Date|Description|Amount table (same normalized key)."""
+    """
+    Fix 16 + 20: Drop long runs of OCR-identical rows; allow up to 2 consecutive rows with the
+    same normalized Date|Description|Amount key (legitimate duplicate charges). Non-consecutive
+    repeats are kept (each run capped separately).
+    """
     if not grid or len(grid) < 2:
         return grid
     if not _is_da3_header(grid):
         return grid
-    seen = set()
-    out = [grid[0]]
+    data_rows = []
     for row in grid[1:]:
         cells = [str(c or "").strip() for c in row]
         if not any(cells):
             continue
+        data_rows.append(row)
+
+    if not data_rows:
+        return grid
+
+    out = [grid[0]]
+    i = 0
+    while i < len(data_rows):
+        row = data_rows[i]
+        cells = [str(c or "").strip() for c in row]
         k = _transaction_row_dedupe_key(cells)
-        if k in seen:
-            continue
-        seen.add(k)
-        out.append(row)
-    return out if len(out) < len(grid) else grid
+        j = i + 1
+        while j < len(data_rows):
+            ncells = [str(c or "").strip() for c in data_rows[j]]
+            if _transaction_row_dedupe_key(ncells) != k:
+                break
+            j += 1
+        run_len = j - i
+        take = min(run_len, 2)
+        for t in range(take):
+            out.append(data_rows[i + t])
+        i = j
+
+    return out
 
 
 def _row_text_full_ucb(grid, r: int) -> str: