Update app.py

app.py

@@ -594,15 +594,12 @@ def _extract_fused_header_artifacts(header_row):
     """
     Generic detector for the OCR artifact where the first data row of a table
     gets fused into the header cells during OCR.
-
     The artifact pattern (BOTH signals must fire simultaneously):
       - Signal 1 — text-fused: a header cell contains KEYWORD + free descriptive text
         e.g.  "DESCRIPTIONBeginning Balance"  "NARRATIONOpening Balance"
       - Signal 2 — money-fused: a DIFFERENT header cell contains KEYWORD + money amount
         e.g.  "BALANCE$3,447.10"  "AMOUNT 5,000.00"
-
     Two-signal guard prevents false positives on clean PDFs from any bank.
-
     Returns:
         (cleaned_header : list[str], recovered_row : list[str] | None)
     """
@@ -707,15 +704,12 @@ def _promote_misplaced_header_row(grid):
     """
     Detects and fixes the OCR artifact where real column headers land in a
     <td> data row instead of the <th> header row.
-
     Two scenarios handled:
-
     Scenario A — Simple misplaced header (e.g. TD Bank):
       <th>ACCOUNT ACTIVITY</th>        ← section title, not a real header
       <td>DATE DESCRIPTION</td> ...    ← real column headers in a data row
       <td>01/05 ...</td>               ← data
     → Promote the keyword row to header, discard the section title rows above.
-
     Scenario B — Metadata header + misplaced column headers (e.g. Citi page 1):
       <th>208479667</th> <th>Beginning Balance:$20.48 Ending Balance:$3.46</th>
       <td>Date Description</td> <td>Debits</td> <td>Credits</td> <td>Balance</td>
@@ -725,7 +719,6 @@ def _promote_misplaced_header_row(grid):
     → Preserve metadata row cells as a plain-text prefix OUTSIDE the table
       by embedding them as a leading data row with colspan (kept for reference).
     → Split any "Date Description" fused cells in data rows.
-
     Guard conditions (no-op if not met — safe for all other PDFs):
       - Current header keyword score < threshold.
       - A data row within the first 5 rows has keyword score >= threshold.
@@ -821,10 +814,8 @@ def _fix_fused_keyvalue_rows(grid):
     Fix rows where label+value are fused into the first cell with all other
     cells empty — common in Interest Summary / Fee Summary sections appended
     to a transaction table by OCR.
-
     e.g. "Beginning Interest Rate0.00%"  → label="Beginning Interest Rate"  value="0.00%"
          "Number of days in this Period31" → label="..."  value="31"
-
     Guard conditions (no-op if not met — safe for all other tables):
       - All cells except first must be empty.
       - A non-space character must immediately precede the digit run (fused).
@@ -912,11 +903,9 @@ def _fix_fused_keyvalue_rows(grid):
 def _extract_textlayer_rows(pdf_path: str, page_num: int):
     """
     Extract structured transaction rows from the PDF text layer using pymupdf.
-
     Returns list of {"date": str, "desc": str, "amount": str} dicts, or []
     if the PDF has no text layer, pymupdf is unavailable, or fewer than 2
     transaction rows are found (guards against non-transaction pages).
-
     Supports multiple date formats:
       - Numeric:    MM/DD, MM/DD/YY, MM/DD/YYYY  (Chase, TD, BoA, Citi)
       - Month-name: Jan 16, Feb 7, Mar 05         (Capital One, Amex)
@@ -1047,7 +1036,6 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
         log.debug("_extract_textlayer_rows failed (page %d): %s", page_num, e)
         return []
 
-
 def _jb_extract_checks_fallback(text_layer: str) -> List[Tuple[str, str, str]]:
     """
     Johnson Bank: when pdfminer column extraction fails (e.g. pymupdf line text or
@@ -1081,18 +1069,15 @@ def _jb_extract_checks_fallback(text_layer: str) -> List[Tuple[str, str, str]]:
             out.append(t)
     return out
 
-
 def _extract_jb_summary_sections(text_layer: str, needs_checks: bool = True, needs_dab: bool = True) -> str:
     """
     Parse Johnson Bank Checks and Daily Account Balance from pdfminer column-separated
     page text and return as HTML tables.
-
     pdfminer outputs multi-column sections as separate column lists:
         Date      Number    Amount
         04-14     3259      45.72
     becomes:
         "Date\n04-14\n04-09\n\nNumber\n3259\n3260\n\nAmount\n45.72\n1,628.00"
-
     Uses pdfminer (always available) not pymupdf.
     """
     if not text_layer:
@@ -1211,7 +1196,6 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
     """
     Compare OCR table rows against PDF text-layer rows and inject any rows
     the OCR model silently dropped (typically identical consecutive rows).
-
     Guard conditions (all must pass for any injection):
       1. PDF text layer must exist and yield >= 2 transaction rows.
       2. The OCR table must have DATE + DESCRIPTION + AMOUNT columns.
@@ -1456,16 +1440,13 @@ def _reconstruct_separator_rows(grid):
     """
     Detect and reconstruct mid-table separator / label rows that OCR has
     incorrectly split across columns.
-
     The artifact (Bank of America and similar):
       A section label like "Card account # XXXX XXXX XXXX 2889" sits between
       transaction rows as a full-width label. OCR splits the trailing digits
       across columns because they are right-aligned, producing variations like:
-
         2-cell split:  ["Card account # XXXX XXXX XXXX 2", "",    "889"]
         3-cell split:  ["Card account # XXXX XXXX XXXX",   "2",   "889"]
         clean 1-cell:  ["Card account # XXXX XXXX XXXX 2889", "", ""]
-
     Detection criteria (ALL must hold — guards safe for all other PDFs):
       1. First cell is NOT a date token (transaction rows always start with date).
       2. First cell contains letters (it is a label, not a bare number).
@@ -1475,7 +1456,6 @@ def _reconstruct_separator_rows(grid):
               — these are the split-off tails of the account/card number.
       4. There must be at least one non-empty cell after the first (to detect
          the split; single-cell rows are also handled as clean label rows).
-
     The fix:
       Concatenate all non-empty cells in order, place the result in col 0,
       blank all other cells.
@@ -1551,19 +1531,15 @@ def _merge_split_rows(grid):
     """
     Fix OCR artifact where a transaction row is split across two <tr> rows
     because the description wrapped to a second line in the source PDF.
-
     Two patterns detected (Hardin County Bank and similar monospace statements):
-
     Pattern A — description + empty continuation:
       Row N:   [desc, '', '', '', '']          ← description, no date/money
       Row N+1: ['', debit, credit, date, bal]  ← money/date, no description
       → Merge into: [desc, debit, credit, date, bal]
-
     Pattern B — description + text continuation + money:
       Row N:   [desc, '', '', '', '']               ← first line of description
       Row N+1: [desc_cont, debit, credit, date, bal] ← continuation + money
       → Merge into: [desc + ' ' + desc_cont, debit, credit, date, bal]
-
     Guard conditions (no-op unless both rows match the pattern):
       - Row N must have non-empty col 0 (description).
       - Row N must have empty date column (col 3 or wherever DATE is).
@@ -1655,11 +1631,9 @@ def _extract_fused_desc_amount(grid):
     """
     Fix OCR artifact where a trailing money amount gets fused into the
     description cell instead of its own debit/credit column.
-
     Example (Hardin County Bank):
       OCR: ['CHASE CREDIT CRD EPAY 8319249882 500.00', '', '', '04/11/25', '16,699.04']
       Fix: ['CHASE CREDIT CRD EPAY 8319249882', '500.00', '', '04/11/25', '16,699.04']
-
     Guard conditions (all must hold):
       - Description ends with a space + money amount (NNN.NN or N,NNN.NN).
       - ALL debit AND credit columns are empty for this row.
@@ -1718,7 +1692,6 @@ def _is_junk_table(grid):
     """
     Detect and suppress known junk tables that are OCR artifacts from
     page headers/footers, not real transaction data.
-
     Current patterns:
     1. CUSTOMER INFORMATION fragment tables (First Horizon Bank):
        Single-column table with header "CUSTOMER INFORMATION" and
@@ -1741,15 +1714,12 @@ def _is_junk_table(grid):
 def _split_fused_multicolumn_header(grid):
     """
     Fix OCR artifact where multiple column headers are fused into one <th> cell.
-
     Example (First Horizon Bank alternating pages):
       Fused:   ['DATE DESCRIPTION CARD #', 'DEPOSIT', 'WITHDRAWAL']
       Correct: ['DATE', 'DESCRIPTION', 'DEPOSIT', 'WITHDRAWAL', 'CARD #']
-
     The data rows also have date + description + card# all in col 0:
       Fused:   ['01/16 PURCHASE - UBER TRIP... 4919', '', '$21.02']
       Correct: ['01/16', 'PURCHASE - UBER TRIP...', '', '$21.02', '4919']
-
     Detection: first header cell contains 2+ keyword terms separated by spaces,
     AND remaining header cells are valid money column keywords.
     """
@@ -1862,7 +1832,6 @@ def _normalize_daily_balance_table(grid):
     - OCR produces N DATE columns followed by N BALANCE columns
       instead of interleaved DATE|BALANCE|DATE|BALANCE pairs
     - Multiple dates are stacked in one cell "01/02\n01/03\n01/04\n01/05"
-
     Reorders columns and expands stacked date cells into proper rows.
     Only fires when ALL of these hold:
     - Header has exactly 2N columns where N >= 2
@@ -1940,9 +1909,7 @@ def _normalize_checks_paid_table(grid):
     - Remaining headers = "AMOUNT AMOUNT AMOUNT"
     - Data col0 = "01/24 4701 01/18 4704 * 01/18 916831 *" (all groups fused)
     - Remaining data cells = amount values
-
     Expands to proper: DATE | CHECK # | AMOUNT | DATE | CHECK # | AMOUNT | ...
-
     Detection: first header cell matches pattern (DATE CHECK #)+ 
     and remaining headers are all AMOUNT.
     """
@@ -2194,64 +2161,78 @@ def convert_plaintext_bank_sections(text: str) -> str:
 # Navy Federal full-page text-layer parser (Fix NF-1)
 # --------------------------
 
+def _is_nfcu_document(pdf_path: str) -> bool:
+    """
+    Strict detector for Navy Federal statements.
+    Used to gate NF-specific parsing so other banks are unaffected.
+    """
+    try:
+        import pymupdf as fitz
+        doc = fitz.open(pdf_path)
+        pages_to_check = min(2, len(doc))
+        text = ""
+        for i in range(pages_to_check):
+            text += "\n" + (doc[i].get_text() or "")
+        doc.close()
+        t = text.lower()
+        return (
+            ("statement of account" in t)
+            and ("access no." in t or "routing number" in t)
+            and ("navy federal" in t or "navyfederal.org" in t)
+        )
+    except Exception:
+        return False
+
+
 def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
     """
     Full text-layer parser for Navy Federal Credit Union statements.
-
     GLM-OCR fails on NFCU pages because:
       - The summary table has 5 wide columns with no visible borders
       - Transaction tables have Amount($) and Balance($) spatially far right
       - Multi-line descriptions (wrapped lines) confuse OCR table detection
-
     This function reads the PDF text layer directly via pdfplumber and produces
     correct, complete HTML for every section: summary table, transaction tables,
     items paid, savings, disclosures.
-
     Detection guards (all must pass — safe no-op for all other banks):
       1. Page text must contain 'Access No.' AND 'Statement of Account'
          (unique to Navy Federal statement format)
       2. pdfplumber must return extractable text (not a scanned page)
-
     Returns complete HTML string for the page, or "" if not NFCU / any error.
     """
     try:
-        # Extract full page text using pymupdf (installed on HuggingFace) or
-        # pdfplumber (installed locally) — try both so the function works in
-        # either environment.
         text = ""
         try:
-            import pymupdf as _fitz_nf
-            _doc_nf = _fitz_nf.open(pdf_path)
-            if page_num >= len(_doc_nf):
-                _doc_nf.close()
-                return ""
-            text = _doc_nf[page_num].get_text() or ""
-            _doc_nf.close()
+            import pdfplumber
+            with pdfplumber.open(pdf_path) as pdf:
+                if page_num >= len(pdf.pages):
+                    return ""
+                page = pdf.pages[page_num]
+                text = page.extract_text() or ""
         except Exception:
-            pass
-
-        if not text:
-            try:
-                import pdfplumber as _plumber_nf
-                with _plumber_nf.open(pdf_path) as _pdf_nf:
-                    if page_num >= len(_pdf_nf.pages):
-                        return ""
-                    text = _pdf_nf.pages[page_num].extract_text() or ""
-            except Exception:
-                pass
-
-        if not text:
-            return ""
+            # Fallback to pymupdf text extraction when pdfplumber isn't available.
+            import pymupdf as fitz
+            doc = fitz.open(pdf_path)
+            if page_num >= len(doc):
+                doc.close()
+                return ""
+            text = doc[page_num].get_text() or ""
+            doc.close()
 
         if not text:
             return ""
 
         # Guard: must match NFCU format on this page OR on page 0
         # (pages 2+ still have "Access No." and "Statement of Account")
+        t = text.lower()
         is_nfcu = (
-            ('Access No.' in text and 'Statement of Account' in text) or
-            ('navyfederal' in text.lower()) or
-            ('Navy Federal' in text)
+            ("statement of account" in t)
+            and (
+                ("access no." in t)
+                or ("routing number" in t)
+                or ("navy federal" in t)
+                or ("navyfederal.org" in t)
+            )
         )
         if not is_nfcu:
             return ""
@@ -2673,6 +2654,7 @@ def run_ocr(uploaded_file):
     try:
         path = uploaded_file.name if hasattr(uploaded_file, "name") else str(uploaded_file)
         is_pdf = path.lower().endswith(".pdf")
+        is_nfcu_doc = _is_nfcu_document(path) if is_pdf else False
         parser = get_parser()
 
         page_heights = []
@@ -2707,14 +2689,12 @@ def run_ocr(uploaded_file):
             # detected, the entire page output is replaced with clean text-layer
             # data and none of the garbled OCR pipeline runs at all.
             # Safe no-op for all other banks (_parse_nfcu_page returns "").
-            if is_pdf:
-                try:
-                    nfcu_html = _parse_nfcu_page(path, page_num)
-                    if nfcu_html:
-                        all_pages.append(nfcu_html)
-                        continue  # skip ALL OCR processing for this page
-                except Exception:
-                    pass  # fall through to normal OCR pipeline on any error
+            if is_pdf and is_nfcu_doc:
+                nfcu_html = _parse_nfcu_page(path, page_num)
+                if nfcu_html:
+                    all_pages.append(nfcu_html)
+                    continue  # skip ALL OCR processing for this page
+                log.warning("NFCU parser returned empty output for page %d", page_num + 1)
 
             # Header inclusion: PDF text -> band words -> OCR band
             hdr = ""