Update app.py

app.py

@@ -1024,12 +1024,17 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             # Line looks like: Jan 16 Jan 16 CAPITAL ONE MOBILE PYMT - $500.00
             # After consuming first date, check if next tokens are also a date
             remaining = line[desc_start:]
+            # Capture post_date if next tokens are also a date
+            post_date_str = ""
             if remaining and month_re.match(remaining[0]):
                 if len(remaining) > 1 and day_re.match(remaining[1]):
-                    # Skip the post date
+                    post_date_str = remaining[0] + " " + remaining[1]
                     remaining = remaining[2:]
                 elif len(remaining) > 0:
                     remaining = remaining[1:]
+            elif remaining and date_re.match(remaining[0]) and not month_re.match(remaining[0]):
+                post_date_str = remaining[0]
+                remaining = remaining[1:]
 
             if len(remaining) < 2:
                 continue
@@ -1051,7 +1056,7 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             if not desc:
                 continue
 
-            rows.append({"date": date_str, "desc": desc, "amount": amount_candidate})
+            rows.append({"date": date_str, "post_date": post_date_str, "desc": desc, "amount": amount_candidate})
 
         # Guard: require at least 2 rows to avoid false positives on non-transaction pages
         return rows if len(rows) >= 2 else []
@@ -1187,26 +1192,47 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                             new_grid.append(list(row))
                         to_inject[key] = 0
 
-            # Case B: append completely missing rows from the text layer.
-            # These are rows the OCR model dropped entirely (not just undercounted).
-            # We build new table rows using the text-layer data, placing values in
-            # the correct columns as identified from the OCR table header.
-            ncols = len(grid[0])
-            for key, (count, tl_row_data) in to_inject_new.items():
-                for _ in range(count):
-                    new_row = [""] * ncols
-                    new_row[date_col] = tl_row_data["date"]
-                    new_row[desc_col] = tl_row_data["desc"]
-                    new_row[amt_col]  = tl_row_data["amount"]
-                    new_grid.append(new_row)
-                    log.info(
-                        "page %d: injected missing row from text layer: %s %s",
-                        page_num, tl_row_data["date"], tl_row_data["desc"][:40]
-                    )
-
             new_table_html = _grid_to_html(new_grid)
             result = result.replace(table_html, new_table_html, 1)
 
+            # Case B: completely missing rows → build a SEPARATE new table
+            # appended after the patched OCR table. Using a separate table
+            # preserves the original PDF structure (e.g. Capital One has
+            # separate "Payments" and "Transactions" sections).
+            if to_inject_new:
+                ncols = len(grid[0])
+                header_row = grid[0]
+
+                # Find post_date column index if it exists in the header
+                post_date_col = None
+                for ci, h in enumerate(header_row):
+                    hh = str(h or "").strip().upper()
+                    if "POST" in hh and "DATE" in hh:
+                        post_date_col = ci
+                        break
+
+                new_rows = []
+                for key, (count, tl_row_data) in to_inject_new.items():
+                    for _ in range(count):
+                        new_row = [""] * ncols
+                        new_row[date_col] = tl_row_data["date"]
+                        new_row[desc_col] = tl_row_data["desc"]
+                        new_row[amt_col]  = tl_row_data["amount"]
+                        if post_date_col is not None:
+                            new_row[post_date_col] = tl_row_data.get("post_date", "")
+                        new_rows.append(new_row)
+                        log.info(
+                            "page %d: new table row from text layer: %s %s",
+                            page_num, tl_row_data["date"], tl_row_data["desc"][:40]
+                        )
+
+                if new_rows:
+                    extra_grid = [header_row] + new_rows
+                    extra_html = "\n\n" + _grid_to_html(extra_grid)
+                    # Insert immediately after the patched table
+                    insert_pos = result.find(new_table_html) + len(new_table_html)
+                    result = result[:insert_pos] + extra_html + result[insert_pos:]
+
         return result
 
     except Exception as e: