Update app.py

app.py

@@ -8,7 +8,8 @@ Primary goal for reconcile rate:
 - preserve right-most columns (often "Balance") by higher DPI render + right padding
 - keep tables as tables (convert markdown pipe tables -> HTML table)
 - return ---page-separator--- between pages
-- expand colspan/rowspan in HTML tables so header and data rows align (fixes reconciliation)
+- normalize HTML tables so header/data columns align (expand colspan/rowspan)
+- clean common header artifacts (e.g. "DESCRIPTIONBeginning Balance", "BALANCE$3,447.10")
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -29,9 +30,9 @@ import os
 import re
 import html
 import tempfile
+from typing import List, Tuple
 from collections import defaultdict
 from html.parser import HTMLParser
-from typing import List, Tuple
 
 import yaml
 import gradio as gr
@@ -262,9 +263,46 @@ def close_unclosed_html(md: str) -> str:
     return md
 
 
+def clean_table_header_artifacts(text: str) -> str:
+    """
+    Generic cleanup for common OCR artifacts in table headers:
+    - 'DESCRIPTIONBeginning Balance' -> 'DESCRIPTION'
+    - 'DESCRIPTIONEnding Balance' -> 'DESCRIPTION'
+    - 'BALANCE$3,447.10' -> 'BALANCE'
+    - 'BALANCE 3,447.10' -> 'BALANCE'
+    Works for any PDF; no bank-specific logic.
+    """
+    if not text:
+        return text
+
+    # DESCRIPTION + (Beginning/Ending Balance) glued
+    text = re.sub(
+        r"(>[^<]*\bDESCRIPTION)\s*(Beginning Balance|Ending Balance)\b([^<]*<)",
+        r"\1\3",
+        text,
+        flags=re.IGNORECASE,
+    )
+
+    # BALANCE with a number glued or appended (keep the word BALANCE only)
+    text = re.sub(
+        r"(>[^<]*\bBALANCE)\s*\$?\s*\d{1,3}(?:,\d{3})*(?:\.\d{2})?\s*([^<]*<)",
+        r"\1\2",
+        text,
+        flags=re.IGNORECASE,
+    )
+    text = re.sub(
+        r"(>[^<]*\bBALANCE)\$",
+        r"\1",
+        text,
+        flags=re.IGNORECASE,
+    )
+
+    return text
+
+
 # ---- Colspan/rowspan expansion: align header and data rows for any PDF ----
 class TableGridParser(HTMLParser):
-    """Parse <table> HTML into a list of rows; each row is a list of (text, colspan, rowspan)."""
+    """Parse <table> HTML into rows of (text, colspan, rowspan)."""
 
     def __init__(self):
         super().__init__()
@@ -295,10 +333,7 @@ class TableGridParser(HTMLParser):
 
 
 def _build_grid(rows_data):
-    """
-    Build a 2D grid from rows_data (list of list of (text, colspan, rowspan)).
-    Expands colspan and rowspan so every row has the same number of columns.
-    """
+    """Expand colspan/rowspan into a rectangular grid."""
     if not rows_data:
         return []
     blocked = defaultdict(set)
@@ -325,7 +360,7 @@ def _build_grid(rows_data):
 
 
 def _grid_to_html(grid):
-    """Turn a 2D grid of cell texts into a single <table> HTML string (no colspan/rowspan)."""
+    """Emit a normalized table without colspan/rowspan."""
     if not grid:
         return ""
     lines = ["<table>"]
@@ -341,11 +376,7 @@ def _grid_to_html(grid):
 
 
 def expand_colspan_rowspan_in_tables(text):
-    """
-    Find every <table>...</table> in text and replace with a version where
-    colspan and rowspan are expanded so all rows have the same number of cells.
-    Works for any PDF; prevents header/data misalignment that breaks reconciliation.
-    """
+    """Normalize every <table>...</table> in text."""
     if not text or "<table" not in text.lower():
         return text
 
@@ -425,7 +456,7 @@ def normalize_money_glyphs(text: str) -> str:
 
 
 def stabilize_tables_and_text(page_md: str) -> str:
-    """Convert markdown pipe tables to HTML, expand colspan/rowspan, and close tags."""
+    """Convert markdown pipe tables to HTML, clean header artifacts, normalize tables, and close tags."""
     if not page_md:
         return page_md
     page_md = normalize_money_glyphs(page_md)
@@ -439,7 +470,14 @@ def stabilize_tables_and_text(page_md: str) -> str:
             out_blocks.append(b)
 
     stabilized = "\n\n".join(out_blocks)
+
+    # 1) Fix common header text artifacts
+    stabilized = clean_table_header_artifacts(stabilized)
+
+    # 2) Normalize tables to a rectangular grid (expand colspan/rowspan)
     stabilized = expand_colspan_rowspan_in_tables(stabilized)
+
+    # 3) Close any unclosed tags to prevent bleed
     return close_unclosed_html(stabilized)
 
 
@@ -539,11 +577,13 @@ def run_ocr(uploaded_file):
             he = header_end_frac if header_end_frac is not None else DEFAULT_ZONE_FRAC
             fs = footer_start_frac if footer_start_frac is not None else (1.0 - DEFAULT_ZONE_FRAC)
 
+            # clamp
             he = max(0.02, min(0.25, he))
             fs = max(0.75, min(0.98, fs))
 
             parts = []
 
+            # Header inclusion: PDF text -> band words -> OCR band
             hdr = ""
             if is_pdf:
                 hdr = extract_zone_text_pdf(path, page_num, 0, he)
@@ -554,9 +594,11 @@ def run_ocr(uploaded_file):
             if hdr and hdr.strip():
                 parts.append(fix_account_number(normalize_money_glyphs(hdr.strip())))
 
+            # Main OCR markdown, stabilized
             if page_md and page_md.strip():
                 parts.append(stabilize_tables_and_text(page_md.strip()))
 
+            # Optional footer
             if ENABLE_FOOTER_OCR and page_num < len(page_images):
                 ftr = ""
                 if is_pdf:
@@ -579,6 +621,7 @@ def run_ocr(uploaded_file):
         return f"Error: {e}\n\n{traceback.format_exc()}"
 
     finally:
+        # cleanup rendered images
         for p in page_images:
             try:
                 if isinstance(p, str) and p.endswith(".png") and "glmocr_page_" in os.path.basename(p):