Update app.py

app.py

@@ -1140,18 +1140,37 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                 ocr_freq[key] = ocr_freq.get(key, 0) + 1
                 ocr_rows_indexed.append((ri, key))
 
-            # Guard: overlap required
+            # Guard: overlap check for Case A (duplicate restore).
+            # For Case A we require at least one row in both OCR and text layer
+            # to confirm we are patching the right table.
+            # For Case B (entirely missing rows) we use a lighter structural check:
+            # if the OCR table has DATE+DESCRIPTION+AMOUNT columns (already verified)
+            # and the text layer date format matches the OCR date format,
+            # we can safely inject — even when zero rows overlap.
             overlap = set(ocr_freq.keys()) & set(tl_freq.keys())
-            if not overlap:
-                continue
+
+            # Detect date format used in OCR table (MM/DD vs MM-DD vs Mon DD)
+            _ocr_dates = [
+                _norm(str(row[date_col] or ""))
+                for row in grid[1:]
+                if len(row) > date_col and str(row[date_col] or "").strip()
+            ]
+            _tl_dates = [_norm(r["date"]) for r in tl_rows]
+            _slash_re  = re.compile(r"^\d{1,2}/\d{2}")
+            _hyphen_re = re.compile(r"^\d{1,2}-\d{2}")
+            def _date_fmt(dates):
+                if any(_slash_re.match(d) for d in dates):  return "slash"
+                if any(_hyphen_re.match(d) for d in dates): return "hyphen"
+                return "other"
+            ocr_fmt = _date_fmt(_ocr_dates)
+            tl_fmt  = _date_fmt(_tl_dates)
+            date_fmt_match = (ocr_fmt == tl_fmt) or "other" in (ocr_fmt, tl_fmt)
 
             # Determine missing rows.
             # Case A: row exists in OCR but count is too low (duplicate dropped)
-            # Case B: row exists in text layer but is completely absent from OCR
-            #         (entire section dropped by OCR model, e.g. Capital One purchases)
-            # For Case B we build the row content from the text layer directly.
-            to_inject = {}       # key -> count of missing copies
-            to_inject_new = {}   # key -> row content for brand-new rows
+            # Case B: row exists in text layer but completely absent from OCR
+            to_inject = {}
+            to_inject_new = {}
 
             for key in tl_freq:
                 ocr_count = ocr_freq.get(key, 0)
@@ -1159,18 +1178,20 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                 if tl_count > ocr_count:
                     missing = tl_count - ocr_count
                     if ocr_count > 0:
-                        # Case A: restore missing duplicates (existing row template)
-                        to_inject[key] = missing
+                        # Case A: restore duplicates — requires overlap confirmation
+                        if overlap:
+                            to_inject[key] = missing
                     else:
-                        # Case B: completely new row — build from text layer data
-                        # Find the original text-layer row for this key
-                        tl_row_data = next(
-                            (r for r in tl_rows
-                             if (_norm(r["date"]), _norm(r["desc"]), _norm(r["amount"])) == key),
-                            None
-                        )
-                        if tl_row_data:
-                            to_inject_new[key] = (missing, tl_row_data)
+                        # Case B: entirely new row — requires date format match
+                        # (lighter guard: no overlap needed, just structural match)
+                        if date_fmt_match:
+                            tl_row_data = next(
+                                (r for r in tl_rows
+                                 if (_norm(r["date"]), _norm(r["desc"]), _norm(r["amount"])) == key),
+                                None
+                            )
+                            if tl_row_data:
+                                to_inject_new[key] = (missing, tl_row_data)
 
             if not to_inject and not to_inject_new:
                 continue