Update app.py

app.py

@@ -733,6 +733,88 @@ def strip_degenerate_html_tables(md: str) -> str:
     return re.sub(r"\n{3,}", "\n\n", out)
 
 
+# OCR sometimes emits a malformed leading pseudo-header row like:
+# Date05/09/25 | TypeDeposit | Amount2,270.00 | ...
+_GLUED_HEADER_ROW_RE = re.compile(r"date\d{1,2}/\d{1,2}|typedeposit|amount\d", re.IGNORECASE)
+
+
+def strip_malformed_leading_table_rows(md: str) -> str:
+    """
+    Remove malformed leading rows in HTML tables when they are clearly bogus
+    header artifacts and the table already contains a proper header row below.
+    """
+    if not md or "<table" not in md.lower():
+        return md
+
+    def repl_table(m: re.Match) -> str:
+        full = m.group(0)
+        trs = list(re.finditer(r"<tr\b[^>]*>.*?</tr>", full, flags=re.IGNORECASE | re.DOTALL))
+        if len(trs) < 2:
+            return full
+
+        first = trs[0].group(0)
+        second = trs[1].group(0)
+        first_plain = _cell_plain_text(first).lower()
+        second_plain = _cell_plain_text(second).lower()
+
+        first_is_malformed = bool(_GLUED_HEADER_ROW_RE.search(first_plain))
+        second_looks_header = (
+            ("date" in second_plain and "description" in second_plain)
+            or ("posting date" in second_plain and "description" in second_plain)
+            or ("date" in second_plain and "balance" in second_plain)
+        )
+        if not (first_is_malformed and second_looks_header):
+            return full
+
+        return full.replace(first, "", 1)
+
+    return re.sub(
+        r"<table\b[^>]*>.*?</table>",
+        repl_table,
+        md,
+        flags=re.IGNORECASE | re.DOTALL,
+    )
+
+
+def strip_subtotal_rows_from_transaction_tables(md: str) -> str:
+    """
+    Remove subtotal/total line-item rows from transaction tables so aggregate
+    amounts are not mistaken as individual transactions.
+    """
+    if not md or "<table" not in md.lower():
+        return md
+
+    def repl_table(m: re.Match) -> str:
+        full = m.group(0)
+        rows = list(re.finditer(r"<tr\b[^>]*>.*?</tr>", full, flags=re.IGNORECASE | re.DOTALL))
+        if len(rows) < 2:
+            return full
+        # Some tables start with a title row, then the actual header appears in row 2/3.
+        sample_headers = [_cell_plain_text(r.group(0)).lower() for r in rows[:4]]
+        is_txn_like = any(
+            ("date" in hp or "posting date" in hp)
+            and ("description" in hp)
+            and ("amount" in hp or "debit" in hp or "credit" in hp)
+            for hp in sample_headers
+        )
+        if not is_txn_like:
+            return full
+
+        out = full
+        for r in rows[1:]:
+            rp = _cell_plain_text(r.group(0)).lower()
+            if "subtotal" in rp or rp.startswith("total "):
+                out = out.replace(r.group(0), "", 1)
+        return out
+
+    return re.sub(
+        r"<table\b[^>]*>.*?</table>",
+        repl_table,
+        md,
+        flags=re.IGNORECASE | re.DOTALL,
+    )
+
+
 # Match long digit runs even when OCR glues trailing letters (e.g. 9257403599CCD).
 _LONG_NUM_TOKEN = re.compile(r"(?<!\d)\d{7,}(?!\d)")
 
@@ -1042,9 +1124,9 @@ def light_stabilize_markdown(page_md: str) -> str:
         else:
             out_blocks.append(coalesce_loose_ledger_lines(b))
     merged = close_unclosed_html("\n\n".join(out_blocks))
+    merged = strip_malformed_leading_table_rows(merged)
+    merged = strip_subtotal_rows_from_transaction_tables(merged)
     merged = mask_non_monetary_long_numbers(merged)
-    merged = directionalize_amount_headers(merged)
-    merged = directionalize_amount_headers_by_nearby_context(merged)
     merged = repair_thead_cell_semantics(merged)
     merged = stabilize_table_markup(merged, rounds=4)
     merged = strip_degenerate_html_tables(merged)
@@ -1789,11 +1871,9 @@ def run_ocr(uploaded_file):
 
         merged = "\n\n---page-separator---\n\n".join(all_pages) if all_pages else "(No content)"
         if merged and merged != "(No content)" and not merged.lstrip().startswith("Error:"):
+            merged = strip_malformed_leading_table_rows(merged)
+            merged = strip_subtotal_rows_from_transaction_tables(merged)
             merged = mask_non_monetary_long_numbers(merged)
-            merged = directionalize_amount_headers(merged)
-            merged = directionalize_amount_headers_by_nearby_context(merged)
-            merged = normalize_ambiguous_amount_balance_tables(merged)
-            merged = normalize_amount_only_transaction_tables(merged)
             merged = stabilize_table_markup(merged, rounds=4)
             merged = enrich_transaction_tables_with_daily_balances(merged)
             # Keep transaction amounts untouched; add explicit snapshot for metadata extraction.