Update app.py

app.py

@@ -1255,63 +1255,44 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                     insert_pos = result.find(new_table_html) + len(new_table_html)
                     result = result[:insert_pos] + extra_html + result[insert_pos:]
 
-        # Case C: NO transaction table exists on this page at all, but text
-        # layer has valid rows. Build a brand-new table and prepend it to
-        # the page content. This handles pages where OCR dropped the entire
-        # transaction section (e.g. East West Bank page 7 with only 2 rows).
-        #
-        # Guard: only fire if no table on this page already has DATE+DESCRIPTION
-        # columns (avoids duplicating non-transaction tables like DAILY BALANCES).
-        table_pattern2 = re.compile(r"<table[^>]*>.*?</table>", re.DOTALL | re.IGNORECASE)
-        has_txn_table = False
-        for tbl in table_pattern2.finditer(result):
-            p2 = TableGridParser()
-            p2.feed(tbl.group(0))
-            g2 = _build_grid(p2.rows)
-            if len(g2) < 2:
-                continue
-            hdr2 = [str(c or "").strip().upper() for c in g2[0]]
-            has_date  = any(re.search(r"DATE", h) for h in hdr2)
-            has_desc  = any(re.search(r"(DESCRIPTION|DETAILS?|NARRATION|PARTICULARS?)", h) for h in hdr2)
-            if has_date and has_desc:
-                has_txn_table = True
-                break
-
-        if not has_txn_table and tl_rows:
-            # Infer date format from text layer to pick column header label
-            _hy = re.compile(r"^\d{1,2}-\d{2}")
-            _sl = re.compile(r"^\d{1,2}/\d{2}")
-            _sample_dates = [_norm(r["date"]) for r in tl_rows[:3]]
-            if any(_hy.match(d) for d in _sample_dates):
-                date_label = "Date"
-            elif any(_sl.match(d) for d in _sample_dates):
-                date_label = "Date"
-            else:
-                date_label = "Trans Date"
+        # Case C: page has text-layer rows but NO transaction table in OCR output.
+        # Build a new table from the text layer and prepend it.
+        # Guard: only fires when the page has NO table with both DATE + DESCRIPTION cols.
+        # This is placed AFTER the for-loop so it only runs once per page, not per table.
+        if tl_rows:
+            _has_txn_table = False
+            for _tbl in table_pattern.finditer(result):
+                _p2 = TableGridParser()
+                _p2.feed(_tbl.group(0))
+                _g2 = _build_grid(_p2.rows)
+                if len(_g2) < 2:
+                    continue
+                _h2 = [str(c or "").strip().upper() for c in _g2[0]]
+                if (any(re.search(r"\bDATE\b", x) for x in _h2) and
+                        any(re.search(r"\b(DESCRIPTION|DETAILS?|NARRATION|PARTICULARS?)\b", x) for x in _h2)):
+                    _has_txn_table = True
+                    break
 
-            has_post = any(r.get("post_date") for r in tl_rows)
-            if has_post:
-                header_row = [date_label, "Post Date", "Transaction Description", "Amount"]
-            else:
-                header_row = [date_label, "Transaction Description", "Amount"]
-
-            date_ci = 0
-            desc_ci = 2 if has_post else 1
-            amt_ci  = 3 if has_post else 2
-
-            new_rows = []
-            for r in tl_rows:
-                row = [""] * len(header_row)
-                row[date_ci] = r["date"]
-                row[desc_ci] = r["desc"]
-                row[amt_ci]  = r["amount"]
-                if has_post:
-                    row[1] = r.get("post_date", "")
-                new_rows.append(row)
-
-            new_table_html = _grid_to_html([header_row] + new_rows)
-            result = new_table_html + "\n\n" + result
-            log.info("page %d: created new table from text layer (%d rows)", page_num, len(new_rows))
+            if not _has_txn_table:
+                _has_post = any(r.get("post_date") for r in tl_rows)
+                if _has_post:
+                    _chdr = ["Date", "Post Date", "Transaction Description", "Amount"]
+                    _di2, _pi2, _xi2, _ai2 = 0, 1, 2, 3
+                else:
+                    _chdr = ["Date", "Transaction Description", "Amount"]
+                    _di2, _xi2, _ai2 = 0, 1, 2
+                _cnew_rows = []
+                for r in tl_rows:
+                    _crow = [""] * len(_chdr)
+                    _crow[_di2] = r["date"]
+                    _crow[_xi2] = r["desc"]
+                    _crow[_ai2] = r["amount"]
+                    if _has_post:
+                        _crow[_pi2] = r.get("post_date", "")
+                    _cnew_rows.append(_crow)
+                result = _grid_to_html([_chdr] + _cnew_rows) + "\n\n" + result
+                log.info("page %d: Case C — new table (%d rows) from text layer",
+                         page_num, len(_cnew_rows))
 
         return result
 
@@ -1646,13 +1627,13 @@ def run_ocr(uploaded_file):
                     # Guard 2: suppress footer that is a raw text-layer dump of
                     # transaction rows (e.g. East West Bank two-column layout).
                     # Detected when the footer contains 3+ date tokens (MM/DD or MM-DD)
-                    # AND the page already has an OCR table — the table already
-                    # captured this data correctly, so the footer text is redundant.
+                    # AND the footer itself contains amounts — meaning it is transaction
+                    # data, not a legitimate page footer like an address or disclaimer.
                     _footer_date_re = re.compile(r"\b\d{1,2}[-/]\d{2}\b")
-                    is_txn_dump = (
-                        len(_footer_date_re.findall(ftr_clean)) >= 3
-                        and any("<table" in part.lower() for part in parts)
-                    )
+                    _footer_amt_re  = re.compile(r"\b\d{1,3}(?:,\d{3})*\.\d{2}\b")
+                    _date_hits = len(_footer_date_re.findall(ftr_clean))
+                    _amt_hits  = len(_footer_amt_re.findall(ftr_clean))
+                    is_txn_dump = _date_hits >= 3 and _amt_hits >= 3
 
                     if not already_present and not is_txn_dump:
                         parts.append(ftr_clean)