Update app.py

app.py

@@ -139,6 +139,10 @@ Primary goals for reconcile rate:
 - Fix FH-2: Normalize packed DAILY BALANCE SUMMARY and CHECKS PAID SUMMARY tables into one-row-per-item
             layouts; also promote multiline register headers where row0 is placeholder and row1 has true
             DATE/DESCRIPTION/DEPOSIT/WITHDRAWAL labels.
+- Fix SE-multi: Multi-account deposit summary tables (Beginning/Deposits/Withdrawals/Ending) sometimes get a
+            bogus OCR sub-row with tiny amounts (e.g. -$1 / $2 / $1) next to a real account row. Drop those
+            dust rows when another row has statement-scale balances, then recompute the TOTAL row from the
+            remaining account rows. Generic; no institution names.
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -164,16 +168,24 @@ from collections import Counter, defaultdict
 from html.parser import HTMLParser
 
 import yaml
-import gradio as gr
-import glmocr
+
+# Gradio and glmocr are optional at import: pro_processor only needs stabilize_tables_and_text
+# (glm_ocr_client). HF Space / run_ocr still use glmocr + gradio when installed.
+try:
+    import glmocr
+
+    GLMOCR_BASE = os.path.dirname(glmocr.__file__)
+    CONFIG_PATH = os.path.join(GLMOCR_BASE, "config.yaml")
+    FORMATTER_PATH = os.path.join(GLMOCR_BASE, "postprocess", "result_formatter.py")
+except ImportError:
+    glmocr = None  # type: ignore
+    GLMOCR_BASE = ""
+    CONFIG_PATH = ""
+    FORMATTER_PATH = ""
 
 log = logging.getLogger("glmocr_app")
 logging.basicConfig(level=logging.INFO)
 
-GLMOCR_BASE = os.path.dirname(glmocr.__file__)
-CONFIG_PATH = os.path.join(GLMOCR_BASE, "config.yaml")
-FORMATTER_PATH = os.path.join(GLMOCR_BASE, "postprocess", "result_formatter.py")
-
 # ============================================================
 # HARD-CODED SETTINGS (edit these numbers to tune quality/speed)
 # ============================================================
@@ -214,33 +226,35 @@ def get_parser():
 # ---------------------------------------------------------------------------
 # Best-effort config tweaks (safe to fail on read-only HF env)
 # ---------------------------------------------------------------------------
-try:
-    with open(CONFIG_PATH, "r") as f:
-        config = yaml.safe_load(f)
-    config["pipeline"]["maas"]["enabled"] = True
-    config["pipeline"]["maas"]["api_key"] = GLMOCR_API_KEY
-    with open(CONFIG_PATH, "w") as f:
-        yaml.dump(config, f, default_flow_style=False, sort_keys=False)
-except Exception:
-    pass
+if CONFIG_PATH:
+    try:
+        with open(CONFIG_PATH, "r") as f:
+            config = yaml.safe_load(f)
+        config["pipeline"]["maas"]["enabled"] = True
+        config["pipeline"]["maas"]["api_key"] = GLMOCR_API_KEY
+        with open(CONFIG_PATH, "w") as f:
+            yaml.dump(config, f, default_flow_style=False, sort_keys=False)
+    except Exception:
+        pass
 
 # Best-effort formatter tweak: avoid stripping header/footer labels
-try:
-    with open(FORMATTER_PATH, "r") as f:
-        source = f.read()
-    for label in (
-        '"header"', "'header'",
-        '"footer"', "'footer'",
-        '"doc_header"', "'doc_header'",
-        '"doc_footer"', "'doc_footer'",
-    ):
-        source = re.sub(r",\s*" + re.escape(label), "", source)
-        source = re.sub(re.escape(label) + r"\s*,", "", source)
-        source = re.sub(re.escape(label), "", source)
-    with open(FORMATTER_PATH, "w") as f:
-        f.write(source)
-except Exception:
-    pass
+if FORMATTER_PATH:
+    try:
+        with open(FORMATTER_PATH, "r") as f:
+            source = f.read()
+        for label in (
+            '"header"', "'header'",
+            '"footer"', "'footer'",
+            '"doc_header"', "'doc_header'",
+            '"doc_footer"', "'doc_footer'",
+        ):
+            source = re.sub(r",\s*" + re.escape(label), "", source)
+            source = re.sub(re.escape(label) + r"\s*,", "", source)
+            source = re.sub(re.escape(label), "", source)
+        with open(FORMATTER_PATH, "w") as f:
+            f.write(source)
+    except Exception:
+        pass
 
 # --------------------------
 # Header/footer helpers
@@ -3247,6 +3261,162 @@ def _expand_td_continuation_three_col_register_grid(grid):
     return new_grid
 
 
+def _parse_summary_table_money(cell: str) -> Optional[float]:
+    """Parse a currency cell from a statement summary row; None if not numeric."""
+    if cell is None:
+        return None
+    t = str(cell).strip().replace(",", "").replace("$", "").replace(" ", "")
+    if not t or t in ("-", "—", "–"):
+        return None
+    t = t.replace("(", "-").replace(")", "")
+    try:
+        return float(t)
+    except ValueError:
+        return None
+
+
+def _format_summary_table_money(value: float) -> str:
+    sign = "-" if value < 0 else ""
+    v = abs(value)
+    s = f"{v:,.2f}"
+    return sign + s
+
+
+def _drop_phantom_micro_account_rows_in_multi_account_summary(grid):
+    """
+    Remove OCR phantom sub-account rows in credit-union / multi-account deposit summary tables
+    (Acct | Beginning | Deposits | Withdrawals | Ending | …) when one row is dust and another
+    is statement-scale; recompute TOTAL from remaining account rows.
+    """
+    if not grid or len(grid) < 4:
+        return grid
+
+    header = [str(c or "").strip() for c in grid[0]]
+    ncols = len(header)
+    if ncols < 5:
+        return grid
+
+    header_join_u = " ".join(h.upper() for h in header)
+    if "DEPOSIT" not in header_join_u or "WITHDRAW" not in header_join_u:
+        return grid
+    if "BEGINNING" not in header_join_u and "PREVIOUS" not in header_join_u:
+        return grid
+    has_ending = "ENDING" in header_join_u or (
+        "NEW" in header_join_u and "BALANCE" in header_join_u
+    )
+    if not has_ending:
+        return grid
+
+    def _col_match(pred):
+        for i, h in enumerate(header):
+            hu = h.upper()
+            if pred(hu):
+                return i
+        return None
+
+    i_beg = _col_match(lambda u: "BEGINNING" in u or ("PREVIOUS" in u and "BALANCE" in u))
+    i_dep = _col_match(lambda u: "DEPOSIT" in u and "WITHDRAW" not in u)
+    i_wd = _col_match(lambda u: "WITHDRAW" in u)
+    i_end = _col_match(
+        lambda u: "ENDING" in u
+        or ("NEW" in u and "BALANCE" in u)
+        or u.startswith("CLOSING")
+    )
+    if None in (i_beg, i_dep, i_wd, i_end):
+        return grid
+    if len({i_beg, i_dep, i_wd, i_end}) != 4:
+        return grid
+
+    money_idxs = (i_beg, i_dep, i_wd, i_end)
+
+    # Dust vs statement scale (tuned for OCR rows like -1 / 2 / 0 / 1 next to 18k withdrawals)
+    _DUST_MAX = 10.0
+    _DUST_SUM = 15.0
+    _REAL_MIN = 3000.0
+
+    def _pad_row(row):
+        r = [str(c or "").strip() for c in row]
+        while len(r) < ncols:
+            r.append("")
+        return r[:ncols]
+
+    account_rows = []  # (grid_index, row_padded, max_abs, sum_abs, vals tuple)
+    total_idx: Optional[int] = None
+
+    for ri in range(1, len(grid)):
+        row = _pad_row(grid[ri])
+        fc = row[0].strip().upper()
+        if fc == "TOTAL" or fc.startswith("TOTAL "):
+            total_idx = ri
+            continue
+        vals = tuple(
+            _parse_summary_table_money(row[j]) if j < len(row) else None
+            for j in money_idxs
+        )
+        if all(v is None for v in vals):
+            continue
+        abs_vals = [abs(v) for v in vals if v is not None]
+        max_abs = max(abs_vals) if abs_vals else 0.0
+        sum_abs = sum(abs(v) for v in vals if v is not None)
+        account_rows.append((ri, row, max_abs, sum_abs, vals))
+
+    if len(account_rows) < 2:
+        return grid
+
+    global_max = max(t[2] for t in account_rows)
+    if global_max < _REAL_MIN:
+        return grid
+
+    phantom_ri = set()
+    for ri, row, max_abs, sum_abs, vals in account_rows:
+        if max_abs <= _DUST_MAX and sum_abs <= _DUST_SUM:
+            phantom_ri.add(ri)
+
+    if not phantom_ri:
+        return grid
+
+    kept_accounts = [
+        (ri, row, vals)
+        for ri, row, max_abs, sum_abs, vals in account_rows
+        if ri not in phantom_ri
+    ]
+    if not kept_accounts:
+        return grid
+
+    # Sum the four money columns across kept account rows
+    sums = [0.0, 0.0, 0.0, 0.0]
+    for _ri, _row, vals in kept_accounts:
+        for k in range(4):
+            if vals[k] is not None:
+                sums[k] += vals[k]
+
+    new_grid: List[List[str]] = [list(header)]
+    for ri in range(1, len(grid)):
+        if ri in phantom_ri:
+            continue
+        if ri == total_idx:
+            continue
+        new_grid.append(_pad_row(grid[ri]))
+
+    if total_idx is not None:
+        total_row = [""] * ncols
+        total_row[0] = "TOTAL"
+        for k, j in enumerate(money_idxs):
+            total_row[j] = _format_summary_table_money(sums[k])
+        old_total = _pad_row(grid[total_idx])
+        for j in range(ncols):
+            if j in (0,) + money_idxs:
+                continue
+            hj = header[j].upper() if j < len(header) else ""
+            if "DIVIDEND" in hj or "YTD" in hj:
+                total_row[j] = old_total[j] if old_total[j] else "0.00"
+            elif old_total[j]:
+                total_row[j] = old_total[j]
+        new_grid.append(total_row)
+
+    return new_grid
+
+
 def normalize_html_tables(text: str) -> str:
     if not text or "<table" not in text.lower():
         return text
@@ -3274,6 +3444,7 @@ def normalize_html_tables(text: str) -> str:
 
             grid = _drop_truly_empty_columns(grid)
             grid = _merge_blank_header_text_columns(grid)
+            grid = _drop_phantom_micro_account_rows_in_multi_account_summary(grid)
             _seq_linear = _split_wide_checks_sequenced_grid(grid)
             if _seq_linear is not None:
                 grid = _seq_linear
@@ -5916,12 +6087,17 @@ def run_ocr(uploaded_file):
             except Exception:
                 pass
 
-with gr.Blocks(title="GLM-OCR") as demo:
-    gr.Markdown("# GLM-OCR\nUpload a PDF or image. Headers included; tables stabilized.")
-    file_in = gr.File(label="Upload PDF or image", file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp"])
-    run_btn = gr.Button("Run OCR", variant="primary")
-    out = gr.Textbox(lines=40, label="Output (markdown)")
-    run_btn.click(fn=run_ocr, inputs=file_in, outputs=out)
+def _create_gradio_demo():
+    import gradio as gr
+
+    with gr.Blocks(title="GLM-OCR") as demo:
+        gr.Markdown("# GLM-OCR\nUpload a PDF or image. Headers included; tables stabilized.")
+        file_in = gr.File(label="Upload PDF or image", file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp"])
+        run_btn = gr.Button("Run OCR", variant="primary")
+        out = gr.Textbox(lines=40, label="Output (markdown)")
+        run_btn.click(fn=run_ocr, inputs=file_in, outputs=out)
+    return demo
+
 
 if __name__ == "__main__":
-    demo.launch()
+    _create_gradio_demo().launch()