Update app.py

app.py

@@ -136,6 +136,9 @@ Primary goals for reconcile rate:
 - Fix FH-1: Drop a leading Date|Description|Amount (DA3) table when its (date, amount) multiset is a
             strict subset of a later DEPOSIT/WITHDRAWAL register (e.g. First Horizon OCR duplicate before
             ACCOUNT SUMMARY). Safe no-op when no such pair exists.
+- Fix FH-2: Normalize packed DAILY BALANCE SUMMARY and CHECKS PAID SUMMARY tables into one-row-per-item
+            layouts; also promote multiline register headers where row0 is placeholder and row1 has true
+            DATE/DESCRIPTION/DEPOSIT/WITHDRAWAL labels.
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -2037,6 +2040,51 @@ def _normalize_daily_balance_table(grid):
 
     return [new_header] + new_rows
 
+def _normalize_multiline_register_header(grid):
+    """
+    Normalize two-row register headers where row0 is mostly blank/Amount placeholders and
+    row1 carries the real labels (DATE, DESCRIPTION, DEPOSIT, WITHDRAWAL, CARD #).
+    """
+    if not grid or len(grid) < 3:
+        return grid
+    r0 = [str(c or "").strip() for c in grid[0]]
+    r1 = [str(c or "").strip() for c in grid[1]]
+    if not r1:
+        return grid
+
+    r1_join = " ".join(r1).upper()
+    if "DATE" not in r1_join:
+        return grid
+    if "WITHDRAWAL" not in r1_join and "DEPOSIT" not in r1_join:
+        return grid
+    if "DESCRIPTION" not in r1_join:
+        return grid
+
+    # Guard: only treat as multiline header if row0 is mostly empty / generic title-ish cells.
+    non_empty_r0 = [x for x in r0 if x]
+    if len(non_empty_r0) > max(3, len(r0) // 2):
+        if not ("ACCOUNT HISTORY" in " ".join(r0).upper()):
+            return grid
+    generic_r0 = all(
+        (not x) or (x.upper() in ("AMOUNT", "ACCOUNT HISTORY", "DATE", "DESCRIPTION", "DEPOSIT", "WITHDRAWAL", "CARD #"))
+        for x in r0
+    )
+    if not generic_r0:
+        return grid
+
+    # Use row1 as canonical header, keeping column count stable.
+    ncols = max(len(r0), len(r1))
+    new_header = []
+    for i in range(ncols):
+        v = (r1[i] if i < len(r1) else "").strip()
+        if not v and i < len(r0):
+            v = str(r0[i] or "").strip()
+        new_header.append(v)
+
+    out = [new_header]
+    out.extend(grid[2:])
+    return out
+
 def _normalize_checks_paid_table(grid):
     if not grid or len(grid) < 2:
         return grid
@@ -3249,6 +3297,7 @@ def normalize_html_tables(text: str) -> str:
             grid = _normalize_fused_date_posted_amount_table(grid)
 
             grid = _split_fused_multicolumn_header(grid)
+            grid = _normalize_multiline_register_header(grid)
 
             grid = _normalize_daily_balance_table(grid)
 
@@ -3275,6 +3324,97 @@ def normalize_html_tables(text: str) -> str:
     return "".join(out)
 
 
+def _normalize_packed_daily_balance_and_checks_summary_tables(text: str) -> str:
+    """
+    Normalize packed statement summary tables often seen in First Horizon OCR output:
+    - DAILY BALANCE SUMMARY with grouped date cells (e.g. "01/02 01/03 01/04 01/05")
+    - CHECKS PAID SUMMARY with grouped "DATE CHECK #" triplets + parallel amount columns
+    """
+    if not text or "<table" not in text.lower():
+        return text
+
+    pattern = re.compile(r"<table[^>]*>.*?</table>", re.DOTALL | re.IGNORECASE)
+    out = []
+    last = 0
+
+    _DATE_RE = re.compile(r"^\d{1,2}/\d{2}(?:/\d{2,4})?$")
+
+    for m in pattern.finditer(text):
+        out.append(text[last:m.start()])
+        tbl = m.group(0)
+        try:
+            p = TableGridParser()
+            p.feed(tbl)
+            g = _build_grid(p.rows)
+            if not g or len(g) < 3:
+                out.append(tbl)
+                last = m.end()
+                continue
+
+            h0 = " ".join(str(c or "").strip().upper() for c in g[0])
+
+            # DAILY BALANCE SUMMARY: unpack grouped dates into one row per date.
+            if "DAILY BALANCE SUMMARY" in h0:
+                rows = [["Date", "Balance"]]
+                for row in g[1:]:
+                    vals = [str(c or "").strip() for c in row]
+                    if not any(vals):
+                        continue
+                    if any("BALANCE" == v.upper() for v in vals):
+                        continue
+                    date_tokens = []
+                    for cell in vals:
+                        for tok in re.split(r"[\s\r\n]+", cell):
+                            t = tok.strip()
+                            if _DATE_RE.match(t):
+                                date_tokens.append(t)
+                    amt_cells = [v for v in vals if _DA3_STRICT_AMOUNT.match(v)]
+                    for d, a in zip(date_tokens, amt_cells):
+                        rows.append([d, a])
+                if len(rows) >= 3:
+                    out.append(_grid_to_html(rows))
+                    last = m.end()
+                    continue
+
+            # CHECKS PAID SUMMARY: unpack grouped DATE/CHECK# with parallel amount columns.
+            if "CHECKS PAID SUMMARY" in h0:
+                rows = [["Date", "Check #", "Amount"]]
+                for row in g[1:]:
+                    vals = [str(c or "").strip() for c in row]
+                    if not any(vals):
+                        continue
+                    if "DATE CHECK #" in " ".join(vals).upper():
+                        continue
+                    left = vals[0] if vals else ""
+                    amounts = [v for v in vals[1:] if _DA3_STRICT_AMOUNT.match(v)]
+                    # tokens like: 01/24 4701 01/18 4704 * 01/18 916831 *
+                    tokens = [t for t in re.split(r"\s+", left) if t]
+                    pairs = []
+                    i = 0
+                    while i < len(tokens):
+                        if _DATE_RE.match(tokens[i]) and i + 1 < len(tokens):
+                            chk = tokens[i + 1]
+                            if chk != "*":
+                                pairs.append((tokens[i], chk))
+                            i += 2
+                            continue
+                        i += 1
+                    for (d, chk), amt in zip(pairs, amounts):
+                        rows.append([d, chk, amt])
+                if len(rows) >= 2:
+                    out.append(_grid_to_html(rows))
+                    last = m.end()
+                    continue
+
+            out.append(tbl)
+        except Exception:
+            out.append(tbl)
+        last = m.end()
+
+    out.append(text[last:])
+    return "".join(out)
+
+
 def _transaction_row_dedupe_key(cells):
     parts = [str(c or "").strip() for c in cells[:3]]
     while len(parts) < 3:
@@ -4533,6 +4673,13 @@ def _fh_withdrawal_column_index(grid) -> Optional[int]:
                 return ci
     return None
 
+def _fh_deposit_column_index(grid) -> Optional[int]:
+    for row in grid[: min(6, len(grid))]:
+        for ci, c in enumerate(row):
+            if str(c or "").strip().upper() == "DEPOSIT":
+                return ci
+    return None
+
 
 def _fh_is_deposit_withdrawal_register(grid) -> bool:
     if not grid or len(grid) < 2:
@@ -4540,7 +4687,10 @@ def _fh_is_deposit_withdrawal_register(grid) -> bool:
     blob = " ".join(
         " ".join(str(c or "") for c in row) for row in grid[: min(5, len(grid))]
     ).upper()
-    return "DEPOSIT" in blob and "WITHDRAWAL" in blob and "DATE" in blob
+    has_date = "DATE" in blob
+    has_desc = "DESCRIPTION" in blob
+    has_amt_side = ("WITHDRAWAL" in blob) or ("DEPOSIT" in blob)
+    return has_date and has_desc and has_amt_side
 
 
 def _drop_da3_duplicate_of_later_deposit_withdrawal_register(text: str) -> str:
@@ -4588,25 +4738,29 @@ def _drop_da3_duplicate_of_later_deposit_withdrawal_register(text: str) -> str:
                 return None
             return c
 
-        def _register_wd_counter(g) -> Optional[Counter]:
+        def _register_amount_counter(g) -> Optional[Counter]:
             if not _fh_is_deposit_withdrawal_register(g):
                 return None
+            dci = _fh_deposit_column_index(g)
             wci = _fh_withdrawal_column_index(g)
-            if wci is None:
+            if dci is None and wci is None:
                 return None
             c = Counter()
             for row in g[1:]:
                 cells = [str(x or "").strip() for x in row]
-                while len(cells) <= wci:
+                need_cols = [x for x in (dci, wci) if x is not None]
+                max_idx = max(need_cols) if need_cols else 0
+                while len(cells) <= max_idx:
                     cells.append("")
                 if not _DA3_LEADING_DATE.match(cells[0]):
                     continue
-                amt = cells[wci]
-                if not amt.strip():
-                    continue
-                if not _DA3_STRICT_AMOUNT.match(amt):
-                    continue
-                c[_norm_da_key_date_amt(cells[0], amt)] += 1
+                for ci in need_cols:
+                    amt = cells[ci]
+                    if not amt.strip():
+                        continue
+                    if not _DA3_STRICT_AMOUNT.match(amt):
+                        continue
+                    c[_norm_da_key_date_amt(cells[0], amt)] += 1
             if sum(c.values()) < 3:
                 return None
             return c
@@ -4626,7 +4780,7 @@ def _drop_da3_duplicate_of_later_deposit_withdrawal_register(text: str) -> str:
                 continue
             for j in range(i + 1, len(grids)):
                 mj, gj = grids[j]
-                rk = _register_wd_counter(gj)
+                rk = _register_amount_counter(gj)
                 if rk is None:
                     continue
                 n_reg = sum(rk.values())
@@ -4637,7 +4791,7 @@ def _drop_da3_duplicate_of_later_deposit_withdrawal_register(text: str) -> str:
                 drop_starts.add(mi.start())
                 log.info(
                     "Fix FH-1: dropped DA3 (%d rows) duplicate of later "
-                    "DEPOSIT/WITHDRAWAL register (%d wd rows)",
+                    "DEPOSIT/WITHDRAWAL register (%d amount rows)",
                     n_frag,
                     n_reg,
                 )
@@ -5507,6 +5661,7 @@ def stabilize_tables_and_text(page_md: str) -> str:
     stabilized = _strip_orphan_continued_da3_flatline(stabilized)
     stabilized = _strip_standalone_continued_banner_line(stabilized)
     stabilized = normalize_html_tables(stabilized)
+    stabilized = _normalize_packed_daily_balance_and_checks_summary_tables(stabilized)
     stabilized = _dedupe_adjacent_duplicate_html_tables(stabilized)
     stabilized = _dedupe_adjacent_loose_register_duplicate_tables(stabilized)
     stabilized = _fix_markdown_fidelity_typos_and_escapes(stabilized)
@@ -5645,6 +5800,7 @@ def run_ocr(uploaded_file):
                     stabilized = _strip_orphan_continued_da3_flatline(stabilized)
                     stabilized = _strip_standalone_continued_banner_line(stabilized)
                     stabilized = normalize_html_tables(stabilized)
+                    stabilized = _normalize_packed_daily_balance_and_checks_summary_tables(stabilized)
                     stabilized = _dedupe_adjacent_duplicate_html_tables(stabilized)
                     stabilized = _dedupe_adjacent_loose_register_duplicate_tables(stabilized)
                     stabilized = _fix_markdown_fidelity_typos_and_escapes(stabilized)
@@ -5728,6 +5884,7 @@ def run_ocr(uploaded_file):
         if all_pages and merged and not merged.startswith("Error") and "<table" in merged.lower():
             merged = _strip_orphan_continued_da3_flatline(merged)
             merged = _strip_standalone_continued_banner_line(merged)
+            merged = _normalize_packed_daily_balance_and_checks_summary_tables(merged)
             merged = _dedupe_adjacent_duplicate_html_tables(merged)
             merged = _dedupe_adjacent_loose_register_duplicate_tables(merged)
             merged = _fix_markdown_fidelity_typos_and_escapes(merged)