Update app.py

app.py

@@ -30,6 +30,10 @@ Primary goals for reconcile rate:
             Electronic Debits (PDF text order); drop orphan empty EC/ED headers that followed the merged table
 - Post-pass: dedupe 3-col Date|Description|Amount tables when one is a duplicate fragment
             (later subset of earlier, earlier subset of later, or identical row sets)
+- Fix 18: adjacent DA3 tables — if the last K data rows of table N match the first K rows of table N+1
+            (same normalized Date|Description|Amount keys), strip that suffix from table N. Fixes
+            formats (e.g. Truist) where deposit rows are glued onto the withdrawals table and repeated
+            under the proper deposits heading (K>=2; first table must keep >=1 data row).
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -3133,7 +3137,9 @@ def _is_da3_header(grid):
     h = [re.sub(r"\s+", " ", str(c or "").strip().lower()) for c in grid[0]]
     if len(h) != 3:
         return False
-    return h[0] == "date" and "description" in h[1] and h[2] == "amount"
+    return h[0] == "date" and "description" in h[1] and (
+        h[2] == "amount" or h[2].startswith("amount")
+    )
 
 
 def _dedupe_duplicate_rows_in_da3_table(grid):
@@ -3266,6 +3272,77 @@ def _split_ucb_deposits_electronic_sections_html(text: str) -> str:
     return _UCB_DEP_CONT_BLOCK.sub(_repl, text)
 
 
+def _parse_da3_grid_from_table_html(table_html: str):
+    """Return grid if table is Date|Description|Amount; else None."""
+    try:
+        p = TableGridParser()
+        p.feed(table_html)
+        grid = _build_grid(p.rows)
+        if grid and _is_da3_header(grid):
+            return grid
+    except Exception:
+        pass
+    return None
+
+
+def _trim_adjacent_da3_suffix_duplicating_next_table_prefix(text: str) -> str:
+    """
+    Remove trailing data rows from a DA3 table when they are repeated as the opening
+    rows of the immediately following DA3 table (OCR/LLM glued the next section onto
+    the prior table). Requires overlap length >= 2 and leaves at least one data row
+    in the first table. Safe no-op when patterns do not match.
+    """
+    if not text or "<table" not in text.lower():
+        return text
+
+    pattern = re.compile(r"<table[^>]*>.*?</table>", re.DOTALL | re.IGNORECASE)
+
+    for _ in range(48):
+        matches = list(pattern.finditer(text))
+        if len(matches) < 2:
+            break
+        changed = False
+        for i in range(len(matches) - 1):
+            g1 = _parse_da3_grid_from_table_html(matches[i].group(0))
+            g2 = _parse_da3_grid_from_table_html(matches[i + 1].group(0))
+            if not g1 or not g2 or len(g1) < 3 or len(g2) < 2:
+                continue
+            d1 = g1[1:]
+            d2 = g2[1:]
+            if len(d1) < 2 or len(d2) < 2:
+                continue
+            max_k = min(len(d1), len(d2))
+            best_k = 0
+            for k in range(max_k, 1, -1):
+                ok = True
+                for j in range(k):
+                    c1 = [str(x or "").strip() for x in d1[-k + j][:3]]
+                    c2 = [str(x or "").strip() for x in d2[j][:3]]
+                    while len(c1) < 3:
+                        c1.append("")
+                    while len(c2) < 3:
+                        c2.append("")
+                    if _transaction_row_dedupe_key(c1) != _transaction_row_dedupe_key(c2):
+                        ok = False
+                        break
+                if ok:
+                    best_k = k
+                    break
+            if best_k < 2:
+                continue
+            if len(d1) - best_k < 1:
+                continue
+            new_grid = [g1[0]] + d1[:-best_k]
+            new_html = _grid_to_html(new_grid)
+            m = matches[i]
+            text = text[: m.start()] + new_html + text[m.end() :]
+            changed = True
+            break
+        if not changed:
+            break
+    return text
+
+
 def _dedupe_subset_transaction_tables_html(text: str) -> str:
     """
     Remove duplicate 3-column Date|Description|Amount tables:
@@ -3275,7 +3352,6 @@ def _dedupe_subset_transaction_tables_html(text: str) -> str:
       - Identical row sets → drop the later table.
       - Fuzzy overlap (≥92% of later rows match earlier) → drop later (Electronic Debits vs main).
       - Runs multiple passes until stable.
-
     Row matching uses normalized keys so OCR variants (e.g. "0 27" vs "027", backticks)
     do not prevent duplicate tables from being detected.
     """
@@ -3397,6 +3473,7 @@ def stabilize_tables_and_text(page_md: str) -> str:
     stabilized = "\n\n".join(out_blocks)
     stabilized = convert_plaintext_bank_sections(stabilized)
     stabilized = normalize_html_tables(stabilized)
+    stabilized = _trim_adjacent_da3_suffix_duplicating_next_table_prefix(stabilized)
     stabilized = _split_ucb_deposits_electronic_sections_html(stabilized)
     stabilized = _dedupe_subset_transaction_tables_html(stabilized)
     return close_unclosed_html(stabilized)
@@ -3525,6 +3602,7 @@ def run_ocr(uploaded_file):
                     # _patch_ocr_with_textlayer may append Case B rows from the PDF text layer
                     # with fused UCB-style cells; run the same HTML table pass again (Fix 13/14).
                     stabilized = normalize_html_tables(stabilized)
+                    stabilized = _trim_adjacent_da3_suffix_duplicating_next_table_prefix(stabilized)
                     stabilized = _split_ucb_deposits_electronic_sections_html(stabilized)
                     stabilized = _dedupe_subset_transaction_tables_html(stabilized)
                 elif is_pdf and is_nfcu_doc:
@@ -3607,6 +3685,7 @@ def run_ocr(uploaded_file):
         merged = "\n\n---page-separator---\n\n".join(all_pages) if all_pages else "(No content)"
         # Cross-page: UCB section split + same transaction tables dedupe once on full doc.
         if all_pages and merged and not merged.startswith("Error") and "<table" in merged.lower():
+            merged = _trim_adjacent_da3_suffix_duplicating_next_table_prefix(merged)
             merged = _split_ucb_deposits_electronic_sections_html(merged)
             merged = _dedupe_subset_transaction_tables_html(merged)
         return merged
@@ -3633,4 +3712,3 @@ with gr.Blocks(title="GLM-OCR") as demo:
 
 if __name__ == "__main__":
     demo.launch()
-