refactor: split line_items_extractor into smaller modules with comprehensive tests

- Extract models.py (LineItem, LineItemsResult dataclasses)
- Extract html_table_parser.py (ColumnMapper, HtmlTableParser)
- Extract merged_cell_handler.py (MergedCellHandler for PP-StructureV3 merged cells)
- Reduce line_items_extractor.py from 971 to 396 lines
- Add constants for magic numbers (MIN_AMOUNT_THRESHOLD, ROW_GROUPING_THRESHOLD, etc.)
- Fix row grouping algorithm in text_line_items_extractor.py
- Demote INFO logs to DEBUG level in structure_detector.py
- Add 209 tests achieving 85%+ coverage on main modules

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

packages/backend/backend/table/html_table_parser.py

@@ -0,0 +1,204 @@
+"""
+HTML Table Parser
+
+Parses HTML tables into structured data and maps columns to field names.
+"""
+
+from html.parser import HTMLParser
+import logging
+
+logger = logging.getLogger(__name__)
+
+# Configuration constants
+# Minimum pattern length to avoid false positives from short substrings
+MIN_PATTERN_MATCH_LENGTH = 3
+# Exact match bonus for column mapping priority
+EXACT_MATCH_BONUS = 100
+
+# Swedish column name mappings
+# Extended to support multiple invoice types: product invoices, rental invoices, utility bills
+COLUMN_MAPPINGS = {
+    "article_number": [
+        "art nummer",
+        "artikelnummer",
+        "artikel",
+        "artnr",
+        "art.nr",
+        "art nr",
+        "objektnummer",  # Rental: property reference
+        "objekt",
+    ],
+    "description": [
+        "beskrivning",
+        "produktbeskrivning",
+        "produkt",
+        "tjänst",
+        "text",
+        "benämning",
+        "vara/tjänst",
+        "vara",
+        # Rental invoice specific
+        "specifikation",
+        "spec",
+        "hyresperiod",  # Rental period
+        "period",
+        "typ",  # Type of charge
+        # Utility bills
+        "förbrukning",  # Consumption
+        "avläsning",  # Meter reading
+    ],
+    "quantity": ["antal", "qty", "st", "pcs", "kvantitet", "m²", "kvm"],
+    "unit": ["enhet", "unit"],
+    "unit_price": ["á-pris", "a-pris", "pris", "styckpris", "enhetspris", "à pris"],
+    "amount": [
+        "belopp",
+        "summa",
+        "total",
+        "netto",
+        "rad summa",
+        # Rental specific
+        "hyra",  # Rent
+        "avgift",  # Fee
+        "kostnad",  # Cost
+        "debitering",  # Charge
+        "totalt",  # Total
+    ],
+    "vat_rate": ["moms", "moms%", "vat", "skatt", "moms %"],
+    # Additional field for rental: deductions/adjustments
+    "deduction": [
+        "avdrag",  # Deduction
+        "rabatt",  # Discount
+        "kredit",  # Credit
+    ],
+}
+
+# Keywords that indicate NOT a line items table
+SUMMARY_KEYWORDS = [
+    "frakt",
+    "faktura.avg",
+    "fakturavg",
+    "exkl.moms",
+    "att betala",
+    "öresavr",
+    "bankgiro",
+    "plusgiro",
+    "ocr",
+    "forfallodatum",
+    "förfallodatum",
+]
+
+
+class _TableHTMLParser(HTMLParser):
+    """Internal HTML parser for tables."""
+
+    def __init__(self):
+        super().__init__()
+        self.rows: list[list[str]] = []
+        self.current_row: list[str] = []
+        self.current_cell: str = ""
+        self.in_td = False
+        self.in_thead = False
+        self.header_row: list[str] = []
+
+    def handle_starttag(self, tag, attrs):
+        if tag == "tr":
+            self.current_row = []
+        elif tag in ("td", "th"):
+            self.in_td = True
+            self.current_cell = ""
+        elif tag == "thead":
+            self.in_thead = True
+
+    def handle_endtag(self, tag):
+        if tag in ("td", "th"):
+            self.in_td = False
+            self.current_row.append(self.current_cell.strip())
+        elif tag == "tr":
+            if self.current_row:
+                if self.in_thead:
+                    self.header_row = self.current_row
+                else:
+                    self.rows.append(self.current_row)
+        elif tag == "thead":
+            self.in_thead = False
+
+    def handle_data(self, data):
+        if self.in_td:
+            self.current_cell += data
+
+
+class HTMLTableParser:
+    """Parse HTML tables into structured data."""
+
+    def parse(self, html: str) -> tuple[list[str], list[list[str]]]:
+        """
+        Parse HTML table and return header and rows.
+
+        Args:
+            html: HTML string containing table.
+
+        Returns:
+            Tuple of (header_row, data_rows).
+        """
+        parser = _TableHTMLParser()
+        parser.feed(html)
+        return parser.header_row, parser.rows
+
+
+class ColumnMapper:
+    """Map column headers to field names."""
+
+    def __init__(self, mappings: dict[str, list[str]] | None = None):
+        """
+        Initialize column mapper.
+
+        Args:
+            mappings: Custom column mappings. Uses Swedish defaults if None.
+        """
+        self.mappings = mappings or COLUMN_MAPPINGS
+
+    def map(self, headers: list[str]) -> dict[int, str]:
+        """
+        Map column indices to field names.
+
+        Args:
+            headers: List of column header strings.
+
+        Returns:
+            Dictionary mapping column index to field name.
+        """
+        mapping = {}
+        for idx, header in enumerate(headers):
+            normalized = self._normalize(header)
+
+            if not normalized.strip():
+                continue
+
+            best_match = None
+            best_match_len = 0
+
+            for field_name, patterns in self.mappings.items():
+                for pattern in patterns:
+                    if pattern == normalized:
+                        # Exact match gets highest priority
+                        best_match = field_name
+                        best_match_len = len(pattern) + EXACT_MATCH_BONUS
+                        break
+                    elif pattern in normalized and len(pattern) > best_match_len:
+                        # Partial match requires minimum length to avoid false positives
+                        if len(pattern) >= MIN_PATTERN_MATCH_LENGTH:
+                            best_match = field_name
+                            best_match_len = len(pattern)
+
+                if best_match_len > EXACT_MATCH_BONUS:
+                    # Found exact match, no need to check other fields
+                    break
+
+            if best_match:
+                mapping[idx] = best_match
+
+        return mapping
+
+    def _normalize(self, header: str) -> str:
+        """Normalize header text for matching."""
+        return header.lower().strip().replace(".", "").replace("-", " ")

packages/backend/backend/table/merged_cell_handler.py

@@ -0,0 +1,423 @@
+"""
+Merged Cell Handler
+
+Handles detection and extraction of data from tables with merged cells,
+a common issue with PP-StructureV3 OCR output.
+"""
+
+import re
+import logging
+from typing import TYPE_CHECKING
+
+from .models import LineItem
+
+if TYPE_CHECKING:
+    from .html_table_parser import ColumnMapper
+
+logger = logging.getLogger(__name__)
+
+# Minimum positive amount to consider as line item (filters noise like row indices)
+MIN_AMOUNT_THRESHOLD = 100
+
+
+class MergedCellHandler:
+    """Handles tables with vertically merged cells from PP-StructureV3."""
+
+    def __init__(self, mapper: "ColumnMapper"):
+        """
+        Initialize handler.
+
+        Args:
+            mapper: ColumnMapper instance for header keyword detection.
+        """
+        self.mapper = mapper
+
+    def has_vertically_merged_cells(self, rows: list[list[str]]) -> bool:
+        """
+        Check if table rows contain vertically merged data in single cells.
+
+        PP-StructureV3 sometimes merges multiple table rows into single cells, e.g.:
+        ["Produktnr 1457280 1457280 1060381", "", "Antal 6ST 6ST 1ST", "Pris 127,20 127,20 159,20"]
+
+        Detection: cells contain repeating patterns of numbers or keywords suggesting multiple lines.
+        """
+        if not rows:
+            return False
+
+        for row in rows:
+            for cell in row:
+                if not cell or len(cell) < 20:
+                    continue
+
+                # Check for multiple product numbers (7+ digit patterns)
+                product_nums = re.findall(r"\b\d{7}\b", cell)
+                if len(product_nums) >= 2:
+                    logger.debug(f"has_vertically_merged_cells: found {len(product_nums)} product numbers in cell")
+                    return True
+
+                # Check for multiple prices (Swedish format: 123,45 or 1 234,56)
+                prices = re.findall(r"\b\d{1,3}(?:\s?\d{3})*[,\.]\d{2}\b", cell)
+                if len(prices) >= 3:
+                    logger.debug(f"has_vertically_merged_cells: found {len(prices)} prices in cell")
+                    return True
+
+                # Check for multiple quantity patterns (e.g., "6ST 6ST 1ST")
+                quantities = re.findall(r"\b\d+\s*(?:ST|st|PCS|pcs)\b", cell)
+                if len(quantities) >= 2:
+                    logger.debug(f"has_vertically_merged_cells: found {len(quantities)} quantities in cell")
+                    return True
+
+        return False
+
+    def split_merged_rows(
+        self, rows: list[list[str]]
+    ) -> tuple[list[str], list[list[str]]]:
+        """
+        Split vertically merged cells back into separate rows.
+
+        Handles complex cases where PP-StructureV3 merges content across
+        multiple HTML rows. For example, 5 line items might be spread across
+        3 HTML rows with content mixed together.
+
+        Strategy:
+        1. Merge all row content per column
+        2. Detect how many actual data rows exist (by counting product numbers)
+        3. Split each column's content into that many lines
+
+        Returns header and data rows.
+        """
+        if not rows:
+            return [], []
+
+        # Filter out completely empty rows
+        non_empty_rows = [r for r in rows if any(cell.strip() for cell in r)]
+        if not non_empty_rows:
+            return [], rows
+
+        # Determine column count
+        col_count = max(len(r) for r in non_empty_rows)
+
+        # Merge content from all rows for each column
+        merged_columns = []
+        for col_idx in range(col_count):
+            col_content = []
+            for row in non_empty_rows:
+                if col_idx < len(row) and row[col_idx].strip():
+                    col_content.append(row[col_idx].strip())
+            merged_columns.append(" ".join(col_content))
+
+        logger.debug(f"split_merged_rows: merged columns = {merged_columns}")
+
+        # Count how many actual data rows we should have
+        # Use the column with most product numbers as reference
+        expected_rows = self._count_expected_rows(merged_columns)
+        logger.debug(f"split_merged_rows: expecting {expected_rows} data rows")
+
+        if expected_rows <= 1:
+            # Not enough data for splitting
+            return [], rows
+
+        # Split each column based on expected row count
+        split_columns = []
+        for col_idx, col_text in enumerate(merged_columns):
+            if not col_text.strip():
+                split_columns.append([""] * (expected_rows + 1))  # +1 for header
+                continue
+            lines = self._split_cell_content_for_rows(col_text, expected_rows)
+            split_columns.append(lines)
+
+        # Ensure all columns have same number of lines (immutable approach)
+        max_lines = max(len(col) for col in split_columns)
+        split_columns = [
+            col + [""] * (max_lines - len(col))
+            for col in split_columns
+        ]
+
+        logger.debug(f"split_merged_rows: split into {max_lines} lines total")
+
+        # First line is header, rest are data rows
+        header = [col[0] for col in split_columns]
+        data_rows = []
+        for line_idx in range(1, max_lines):
+            row = [col[line_idx] if line_idx < len(col) else "" for col in split_columns]
+            if any(cell.strip() for cell in row):
+                data_rows.append(row)
+
+        logger.debug(f"split_merged_rows: header={header}, data_rows count={len(data_rows)}")
+        return header, data_rows
+
+    def _count_expected_rows(self, merged_columns: list[str]) -> int:
+        """
+        Count how many data rows should exist based on content patterns.
+
+        Returns the maximum count found from:
+        - Product numbers (7 digits)
+        - Quantity patterns (number + ST/PCS)
+        - Amount patterns (in columns likely to be totals)
+        """
+        max_count = 0
+
+        for col_text in merged_columns:
+            if not col_text:
+                continue
+
+            # Count product numbers (most reliable indicator)
+            product_nums = re.findall(r"\b\d{7}\b", col_text)
+            max_count = max(max_count, len(product_nums))
+
+            # Count quantities (e.g., "6ST 6ST 1ST 1ST 1ST")
+            quantities = re.findall(r"\b\d+\s*(?:ST|st|PCS|pcs)\b", col_text)
+            max_count = max(max_count, len(quantities))
+
+        return max_count
+
+    def _split_cell_content_for_rows(self, cell: str, expected_rows: int) -> list[str]:
+        """
+        Split cell content knowing how many data rows we expect.
+
+        This is smarter than split_cell_content because it knows the target count.
+        """
+        cell = cell.strip()
+
+        # Try product number split first
+        product_pattern = re.compile(r"(\b\d{7}\b)")
+        products = product_pattern.findall(cell)
+        if len(products) == expected_rows:
+            parts = product_pattern.split(cell)
+            header = parts[0].strip() if parts else ""
+            # Include description text after each product number
+            values = []
+            for i in range(1, len(parts), 2):  # Odd indices are product numbers
+                if i < len(parts):
+                    prod_num = parts[i].strip()
+                    # Check if there's description text after
+                    desc = parts[i + 1].strip() if i + 1 < len(parts) else ""
+                    # If description looks like text (not another pattern), include it
+                    if desc and not re.match(r"^\d{7}$", desc):
+                        # Truncate at next product number pattern if any
+                        desc_clean = re.split(r"\d{7}", desc)[0].strip()
+                        if desc_clean:
+                            values.append(f"{prod_num} {desc_clean}")
+                        else:
+                            values.append(prod_num)
+                    else:
+                        values.append(prod_num)
+            if len(values) == expected_rows:
+                return [header] + values
+
+        # Try quantity split
+        qty_pattern = re.compile(r"(\b\d+\s*(?:ST|st|PCS|pcs|M|m|KG|kg)\b)")
+        quantities = qty_pattern.findall(cell)
+        if len(quantities) == expected_rows:
+            parts = qty_pattern.split(cell)
+            header = parts[0].strip() if parts else ""
+            values = [p.strip() for p in parts[1:] if p.strip() and qty_pattern.match(p)]
+            if len(values) == expected_rows:
+                return [header] + values
+
+        # Try amount split for discount+totalsumma columns
+        cell_lower = cell.lower()
+        has_discount = any(kw in cell_lower for kw in ["rabatt", "discount"])
+        has_total = any(kw in cell_lower for kw in ["totalsumma", "total", "summa", "belopp"])
+
+        if has_discount and has_total:
+            # Extract only amounts (3+ digit numbers), skip discount percentages
+            amount_pattern = re.compile(r"\b(\d{3,}[,\.]\d{2})\b")
+            amounts = amount_pattern.findall(cell)
+            if len(amounts) >= expected_rows:
+                # Take the last expected_rows amounts (they are likely the totals)
+                return ["Totalsumma"] + amounts[:expected_rows]
+
+        # Try price split
+        price_pattern = re.compile(r"(\b\d{1,3}(?:\s?\d{3})*[,\.]\d{2}\b)")
+        prices = price_pattern.findall(cell)
+        if len(prices) >= expected_rows:
+            parts = price_pattern.split(cell)
+            header = parts[0].strip() if parts else ""
+            values = [p.strip() for p in parts[1:] if p.strip() and price_pattern.match(p)]
+            if len(values) >= expected_rows:
+                return [header] + values[:expected_rows]
+
+        # Fall back to original single-value behavior
+        return [cell]
+
+    def split_cell_content(self, cell: str) -> list[str]:
+        """
+        Split a cell containing merged multi-line content.
+
+        Strategies:
+        1. Look for product number patterns (7 digits)
+        2. Look for quantity patterns (number + ST/PCS)
+        3. Look for price patterns (with decimal)
+        4. Handle interleaved discount+amount patterns
+        """
+        cell = cell.strip()
+
+        # Strategy 1: Split by product numbers (common pattern: "Produktnr 1234567 1234568")
+        product_pattern = re.compile(r"(\b\d{7}\b)")
+        products = product_pattern.findall(cell)
+        if len(products) >= 2:
+            # Extract header (text before first product number) and values
+            parts = product_pattern.split(cell)
+            header = parts[0].strip() if parts else ""
+            values = [p for p in parts[1:] if p.strip() and re.match(r"\d{7}", p)]
+            return [header] + values
+
+        # Strategy 2: Split by quantities (e.g., "Antal 6ST 6ST 1ST")
+        qty_pattern = re.compile(r"(\b\d+\s*(?:ST|st|PCS|pcs|M|m|KG|kg)\b)")
+        quantities = qty_pattern.findall(cell)
+        if len(quantities) >= 2:
+            parts = qty_pattern.split(cell)
+            header = parts[0].strip() if parts else ""
+            values = [p.strip() for p in parts[1:] if p.strip() and qty_pattern.match(p)]
+            return [header] + values
+
+        # Strategy 3: Handle interleaved discount+amount (e.g., "Rabatt i% Totalsumma 10,0 686,88 10,0 686,88")
+        # Check if header contains two keywords indicating merged columns
+        cell_lower = cell.lower()
+        has_discount_header = any(kw in cell_lower for kw in ["rabatt", "discount"])
+        has_amount_header = any(kw in cell_lower for kw in ["totalsumma", "summa", "belopp", "total"])
+
+        if has_discount_header and has_amount_header:
+            # Extract all numbers and pair them (discount, amount, discount, amount, ...)
+            # Pattern for amounts: 3+ digit numbers with decimals (e.g., 686,88)
+            amount_pattern = re.compile(r"\b(\d{3,}[,\.]\d{2})\b")
+            amounts = amount_pattern.findall(cell)
+
+            if len(amounts) >= 2:
+                # Return header as "Totalsumma" (amount header) so it maps to amount field, not deduction
+                # This avoids the "Rabatt" keyword causing is_deduction=True
+                header = "Totalsumma"
+                return [header] + amounts
+
+        # Strategy 4: Split by prices (e.g., "Pris 127,20 127,20 159,20")
+        price_pattern = re.compile(r"(\b\d{1,3}(?:\s?\d{3})*[,\.]\d{2}\b)")
+        prices = price_pattern.findall(cell)
+        if len(prices) >= 2:
+            parts = price_pattern.split(cell)
+            header = parts[0].strip() if parts else ""
+            values = [p.strip() for p in parts[1:] if p.strip() and price_pattern.match(p)]
+            return [header] + values
+
+        # No pattern detected, return as single value
+        return [cell]
+
+    def has_merged_header(self, header: list[str] | None) -> bool:
+        """
+        Check if header appears to be a merged cell containing multiple column names.
+
+        This happens when OCR merges table headers into a single cell, e.g.:
+        "Specifikation 0218103-1201 2 rum och kök Hyra Avdrag" instead of separate columns.
+
+        Also handles cases where PP-StructureV3 produces headers like:
+        ["Specifikation ... Hyra Avdrag", "", "", ""] with empty trailing cells.
+        """
+        if header is None or not header:
+            return False
+
+        # Filter out empty cells to find the actual content
+        non_empty_cells = [h for h in header if h.strip()]
+
+        # Check if we have a single non-empty cell that contains multiple keywords
+        if len(non_empty_cells) == 1:
+            header_text = non_empty_cells[0].lower()
+            # Count how many column keywords are in this single cell
+            keyword_count = 0
+            for patterns in self.mapper.mappings.values():
+                for pattern in patterns:
+                    if pattern in header_text:
+                        keyword_count += 1
+                        break  # Only count once per field type
+
+            logger.debug(f"has_merged_header: header_text='{header_text}', keyword_count={keyword_count}")
+            return keyword_count >= 2
+
+        return False
+
+    def extract_from_merged_cells(
+        self, header: list[str], rows: list[list[str]]
+    ) -> list[LineItem]:
+        """
+        Extract line items from tables with merged cells.
+
+        For poorly OCR'd tables like:
+        Header: ["Specifikation 0218103-1201 2 rum och kök Hyra Avdrag"]
+        Row 1: ["", "", "", "8159"]      <- amount row
+        Row 2: ["", "", "", "-2 000"]    <- deduction row (separate line item)
+
+        Or:
+        Row: ["", "", "", "8159 -2 000"]  <- both in same row -> 2 line items
+
+        Each amount becomes its own line item. Negative amounts are marked as is_deduction=True.
+        """
+        items = []
+
+        # Amount pattern for Swedish format - match numbers like "8159" or "8 159" or "-2000" or "-2 000"
+        amount_pattern = re.compile(
+            r"(-?\d[\d\s]*(?:[,\.]\d+)?)"
+        )
+
+        # Try to parse header cell for description info
+        header_text = " ".join(h for h in header if h.strip()) if header else ""
+        logger.debug(f"extract_from_merged_cells: header_text='{header_text}'")
+        logger.debug(f"extract_from_merged_cells: rows={rows}")
+
+        # Extract description from header
+        description = None
+        article_number = None
+
+        # Look for object number pattern (e.g., "0218103-1201")
+        obj_match = re.search(r"(\d{7}-\d{4})", header_text)
+        if obj_match:
+            article_number = obj_match.group(1)
+
+        # Look for description after object number
+        desc_match = re.search(r"\d{7}-\d{4}\s+(.+?)(?:\s+(?:Hyra|Avdrag|Belopp))", header_text, re.IGNORECASE)
+        if desc_match:
+            description = desc_match.group(1).strip()
+
+        row_index = 0
+        for row in rows:
+            # Combine all non-empty cells in the row
+            row_text = " ".join(cell.strip() for cell in row if cell.strip())
+            logger.debug(f"extract_from_merged_cells: row text='{row_text}'")
+
+            if not row_text:
+                continue
+
+            # Find all amounts in the row
+            amounts = amount_pattern.findall(row_text)
+            logger.debug(f"extract_from_merged_cells: amounts={amounts}")
+
+            for amt_str in amounts:
+                # Clean the amount string
+                cleaned = amt_str.replace(" ", "").strip()
+                if not cleaned or cleaned == "-":
+                    continue
+
+                is_deduction = cleaned.startswith("-")
+
+                # Skip small positive numbers that are likely not amounts
+                # (e.g., row indices, small percentages)
+                if not is_deduction:
+                    try:
+                        val = float(cleaned.replace(",", "."))
+                        if val < MIN_AMOUNT_THRESHOLD:
+                            continue
+                    except ValueError:
+                        continue
+
+                # Create a line item for each amount
+                item = LineItem(
+                    row_index=row_index,
+                    description=description if row_index == 0 else "Avdrag" if is_deduction else None,
+                    article_number=article_number if row_index == 0 else None,
+                    amount=cleaned,
+                    is_deduction=is_deduction,
+                    confidence=0.7,
+                )
+                items.append(item)
+                row_index += 1
+                logger.debug(f"extract_from_merged_cells: created item amount={cleaned}, is_deduction={is_deduction}")
+
+        return items

packages/backend/backend/table/models.py

@@ -0,0 +1,61 @@
+"""
+Line Items Data Models
+
+Dataclasses for line item extraction results.
+"""
+
+from dataclasses import dataclass
+from decimal import Decimal, InvalidOperation
+
+
+@dataclass
+class LineItem:
+    """Single line item from invoice."""
+
+    row_index: int
+    description: str | None = None
+    quantity: str | None = None
+    unit: str | None = None
+    unit_price: str | None = None
+    amount: str | None = None
+    article_number: str | None = None
+    vat_rate: str | None = None
+    is_deduction: bool = False  # True if this row is a deduction/discount
+    confidence: float = 0.9
+
+
+@dataclass
+class LineItemsResult:
+    """Result of line items extraction."""
+
+    items: list[LineItem]
+    header_row: list[str]
+    raw_html: str
+    is_reversed: bool = False
+
+    @property
+    def total_amount(self) -> str | None:
+        """Calculate total amount from line items (deduction rows have negative amounts)."""
+        if not self.items:
+            return None
+
+        total = Decimal("0")
+        for item in self.items:
+            if item.amount:
+                try:
+                    # Parse Swedish number format (1 234,56)
+                    amount_str = item.amount.replace(" ", "").replace(",", ".")
+                    total += Decimal(amount_str)
+                except InvalidOperation:
+                    pass
+
+        if total == 0:
+            return None
+
+        # Format back to Swedish format
+        formatted = f"{total:,.2f}".replace(",", " ").replace(".", ",")
+        # Fix the space/comma swap
+        parts = formatted.rsplit(",", 1)
+        if len(parts) == 2:
+            return parts[0].replace(" ", " ") + "," + parts[1]
+        return formatted

packages/backend/backend/table/structure_detector.py

@@ -158,36 +158,36 @@ class TableDetector:
             return tables
 
         # Log raw result type for debugging
-        logger.info(f"PP-StructureV3 raw results type: {type(results).__name__}")
+        logger.debug(f"PP-StructureV3 raw results type: {type(results).__name__}")
 
         # Handle case where results is a single dict-like object (PaddleX 3.x)
         # rather than a list of results
         if hasattr(results, "get") and not isinstance(results, list):
             # Single result object - wrap in list for uniform processing
-            logger.info("Results is dict-like, wrapping in list")
+            logger.debug("Results is dict-like, wrapping in list")
             results = [results]
         elif hasattr(results, "__iter__") and not isinstance(results, (list, tuple)):
             # Iterator or generator - convert to list
             try:
                 results = list(results)
-                logger.info(f"Converted iterator to list with {len(results)} items")
+                logger.debug(f"Converted iterator to list with {len(results)} items")
             except Exception as e:
                 logger.warning(f"Failed to convert results to list: {e}")
                 return tables
 
-        logger.info(f"Processing {len(results)} result(s)")
+        logger.debug(f"Processing {len(results)} result(s)")
 
         for i, result in enumerate(results):
             try:
                 result_type = type(result).__name__
                 has_get = hasattr(result, "get")
                 has_layout = hasattr(result, "layout_elements")
-                logger.info(f"Result[{i}]: type={result_type}, has_get={has_get}, has_layout_elements={has_layout}")
+                logger.debug(f"Result[{i}]: type={result_type}, has_get={has_get}, has_layout_elements={has_layout}")
 
                 # Try PaddleX 3.x API first (dict-like with table_res_list)
                 if has_get:
                     parsed = self._parse_paddlex_result(result)
-                    logger.info(f"Result[{i}]: parsed {len(parsed)} tables via PaddleX path")
+                    logger.debug(f"Result[{i}]: parsed {len(parsed)} tables via PaddleX path")
                     tables.extend(parsed)
                     continue
 
@@ -201,14 +201,14 @@ class TableDetector:
                         if table_result and table_result.confidence >= self.config.min_confidence:
                             tables.append(table_result)
                             legacy_count += 1
-                    logger.info(f"Result[{i}]: parsed {legacy_count} tables via legacy path")
+                    logger.debug(f"Result[{i}]: parsed {legacy_count} tables via legacy path")
                 else:
                     logger.warning(f"Result[{i}]: no recognized API (not dict-like and no layout_elements)")
             except Exception as e:
                 logger.warning(f"Failed to parse result: {type(result).__name__}, error: {e}")
                 continue
 
-        logger.info(f"Total tables detected: {len(tables)}")
+        logger.debug(f"Total tables detected: {len(tables)}")
         return tables
 
     def _parse_paddlex_result(self, result: Any) -> list[TableDetectionResult]:
@@ -223,7 +223,7 @@ class TableDetector:
                 result_keys = list(result.keys())
             elif hasattr(result, "__dict__"):
                 result_keys = list(result.__dict__.keys())
-            logger.info(f"Parsing PaddleX result: type={result_type}, keys={result_keys}")
+            logger.debug(f"Parsing PaddleX result: type={result_type}, keys={result_keys}")
 
             # Get table results from PaddleX 3.x API
             # Handle both dict.get() and attribute access
@@ -234,8 +234,8 @@ class TableDetector:
                 table_res_list = getattr(result, "table_res_list", None)
                 parsing_res_list = getattr(result, "parsing_res_list", [])
 
-            logger.info(f"table_res_list: {type(table_res_list).__name__}, count={len(table_res_list) if table_res_list else 0}")
-            logger.info(f"parsing_res_list: {type(parsing_res_list).__name__}, count={len(parsing_res_list) if parsing_res_list else 0}")
+            logger.debug(f"table_res_list: {type(table_res_list).__name__}, count={len(table_res_list) if table_res_list else 0}")
+            logger.debug(f"parsing_res_list: {type(parsing_res_list).__name__}, count={len(parsing_res_list) if parsing_res_list else 0}")
 
             if not table_res_list:
                 # Log available keys/attributes for debugging
@@ -330,7 +330,7 @@ class TableDetector:
                     # Default confidence for PaddleX 3.x results
                     confidence = 0.9
 
-                    logger.info(f"Table {i}: html_len={len(html)}, cells={len(cells)}")
+                    logger.debug(f"Table {i}: html_len={len(html)}, cells={len(cells)}")
                     tables.append(TableDetectionResult(
                         bbox=(float(bbox[0]), float(bbox[1]), float(bbox[2]), float(bbox[3])),
                         html=html,
@@ -467,14 +467,14 @@ class TableDetector:
         if not pdf_path.exists():
             raise FileNotFoundError(f"PDF not found: {pdf_path}")
 
-        logger.info(f"detect_from_pdf: {pdf_path}, page={page_number}, dpi={dpi}")
+        logger.debug(f"detect_from_pdf: {pdf_path}, page={page_number}, dpi={dpi}")
 
         # Render specific page
         for page_no, image_bytes in render_pdf_to_images(str(pdf_path), dpi=dpi):
             if page_no == page_number:
                 image = Image.open(io.BytesIO(image_bytes))
                 image_array = np.array(image)
-                logger.info(f"detect_from_pdf: rendered page {page_no}, image shape={image_array.shape}")
+                logger.debug(f"detect_from_pdf: rendered page {page_no}, image shape={image_array.shape}")
                 return self.detect(image_array)
 
         raise ValueError(f"Page {page_number} not found in PDF")

packages/backend/backend/table/text_line_items_extractor.py

@@ -15,6 +15,11 @@ import logging
 
 logger = logging.getLogger(__name__)
 
+# Configuration constants
+DEFAULT_ROW_TOLERANCE = 15.0  # Max vertical distance (pixels) to consider same row
+MIN_ITEMS_FOR_VALID_EXTRACTION = 2  # Minimum items required for valid extraction
+MIN_TEXT_ELEMENTS_FOR_EXTRACTION = 5  # Minimum text elements needed to attempt extraction
+
 
 @dataclass
 class TextElement:
@@ -65,7 +70,10 @@ class TextLineItemsResult:
     extraction_method: str = "text_spatial"
 
 
-# Swedish amount pattern: 1 234,56 or 1234.56 or 1,234.56
+# Amount pattern matches Swedish, US, and simple numeric formats
+# Handles: "1 234,56", "1,234.56", "1234.56", "100 kr", "50:-", "-100,00"
+# Does NOT handle: amounts with more than 2 decimal places, scientific notation
+# See tests in test_text_line_items_extractor.py::TestAmountPattern
 AMOUNT_PATTERN = re.compile(
     r"(?<![0-9])(?:"
     r"-?\d{1,3}(?:\s\d{3})*(?:,\d{2})?"  # Swedish: 1 234,56
@@ -128,17 +136,17 @@ class TextLineItemsExtractor:
 
     def __init__(
         self,
-        row_tolerance: float = 15.0,  # Max vertical distance to consider same row
-        min_items_for_valid: int = 2,  # Minimum items to consider extraction valid
+        row_tolerance: float = DEFAULT_ROW_TOLERANCE,
+        min_items_for_valid: int = MIN_ITEMS_FOR_VALID_EXTRACTION,
     ):
         """
         Initialize extractor.
 
         Args:
             row_tolerance: Maximum vertical distance (pixels) between elements
-                          to consider them on the same row.
+                          to consider them on the same row. Default: 15.0
             min_items_for_valid: Minimum number of line items required for
-                                extraction to be considered successful.
+                                extraction to be considered successful. Default: 2
         """
         self.row_tolerance = row_tolerance
         self.min_items_for_valid = min_items_for_valid
@@ -161,10 +169,13 @@ class TextLineItemsExtractor:
 
         # Extract text elements from parsing results
         text_elements = self._extract_text_elements(parsing_res_list)
-        logger.info(f"TextLineItemsExtractor: found {len(text_elements)} text elements")
+        logger.debug(f"TextLineItemsExtractor: found {len(text_elements)} text elements")
 
-        if len(text_elements) < 5:  # Need at least a few elements
-            logger.debug("Too few text elements for line item extraction")
+        if len(text_elements) < MIN_TEXT_ELEMENTS_FOR_EXTRACTION:
+            logger.debug(
+                f"Too few text elements ({len(text_elements)}) for line item extraction, "
+                f"need at least {MIN_TEXT_ELEMENTS_FOR_EXTRACTION}"
+            )
             return None
 
         return self.extract_from_text_elements(text_elements)
@@ -183,11 +194,11 @@ class TextLineItemsExtractor:
         """
         # Group elements by row
         rows = self._group_by_row(text_elements)
-        logger.info(f"TextLineItemsExtractor: grouped into {len(rows)} rows")
+        logger.debug(f"TextLineItemsExtractor: grouped into {len(rows)} rows")
 
         # Find the line items section
         item_rows = self._identify_line_item_rows(rows)
-        logger.info(f"TextLineItemsExtractor: identified {len(item_rows)} potential item rows")
+        logger.debug(f"TextLineItemsExtractor: identified {len(item_rows)} potential item rows")
 
         if len(item_rows) < self.min_items_for_valid:
             logger.debug(f"Found only {len(item_rows)} item rows, need at least {self.min_items_for_valid}")
@@ -195,7 +206,7 @@ class TextLineItemsExtractor:
 
         # Extract structured items
         items = self._parse_line_items(item_rows)
-        logger.info(f"TextLineItemsExtractor: extracted {len(items)} line items")
+        logger.debug(f"TextLineItemsExtractor: extracted {len(items)} line items")
 
         if len(items) < self.min_items_for_valid:
             return None
@@ -209,7 +220,11 @@ class TextLineItemsExtractor:
     def _extract_text_elements(
         self, parsing_res_list: list[dict[str, Any]]
     ) -> list[TextElement]:
-        """Extract TextElement objects from parsing_res_list."""
+        """Extract TextElement objects from parsing_res_list.
+
+        Handles both dict and LayoutBlock object formats from PP-StructureV3.
+        Gracefully skips invalid elements with appropriate logging.
+        """
         elements = []
 
         for elem in parsing_res_list:
@@ -220,11 +235,15 @@ class TextLineItemsExtractor:
                     bbox = elem.get("bbox", [])
                     # Try both 'text' and 'content' keys
                     text = elem.get("text", "") or elem.get("content", "")
-                else:
+                elif hasattr(elem, "label"):
                     label = getattr(elem, "label", "")
                     bbox = getattr(elem, "bbox", [])
                     # LayoutBlock objects use 'content' attribute
                     text = getattr(elem, "content", "") or getattr(elem, "text", "")
+                else:
+                    # Element is neither dict nor has expected attributes
+                    logger.debug(f"Skipping element with unexpected type: {type(elem).__name__}")
+                    continue
 
                 # Only process text elements (skip images, tables, etc.)
                 if label not in ("text", "paragraph_title", "aside_text"):
@@ -232,6 +251,7 @@ class TextLineItemsExtractor:
 
                 # Validate bbox
                 if not self._valid_bbox(bbox):
+                    logger.debug(f"Skipping element with invalid bbox: {bbox}")
                     continue
 
                 # Clean text
@@ -250,8 +270,13 @@ class TextLineItemsExtractor:
                         ),
                     )
                 )
+            except (KeyError, TypeError, ValueError, AttributeError) as e:
+                # Expected format issues - log at debug level
+                logger.debug(f"Skipping element due to format issue: {e}")
+                continue
             except Exception as e:
-                logger.debug(f"Failed to parse element: {e}")
+                # Unexpected errors - log at warning level for visibility
+                logger.warning(f"Unexpected error parsing element: {type(e).__name__}: {e}")
                 continue
 
         return elements
@@ -270,6 +295,7 @@ class TextLineItemsExtractor:
         Group text elements into rows based on vertical position.
 
         Elements within row_tolerance of each other are considered same row.
+        Uses dynamic average center_y to handle varying element heights more accurately.
         """
         if not elements:
             return []
@@ -277,22 +303,22 @@ class TextLineItemsExtractor:
         # Sort by vertical position
         sorted_elements = sorted(elements, key=lambda e: e.center_y)
 
-        rows = []
-        current_row = [sorted_elements[0]]
-        current_y = sorted_elements[0].center_y
+        rows: list[list[TextElement]] = []
+        current_row: list[TextElement] = [sorted_elements[0]]
 
         for elem in sorted_elements[1:]:
-            if abs(elem.center_y - current_y) <= self.row_tolerance:
-                # Same row
+            # Calculate dynamic average center_y for current row
+            avg_center_y = sum(e.center_y for e in current_row) / len(current_row)
+
+            if abs(elem.center_y - avg_center_y) <= self.row_tolerance:
+                # Same row - add element and recalculate average on next iteration
                 current_row.append(elem)
             else:
-                # New row
-                if current_row:
-                    # Sort row by horizontal position
-                    current_row.sort(key=lambda e: e.center_x)
-                    rows.append(current_row)
+                # New row - finalize current row
+                # Sort row by horizontal position (left to right)
+                current_row.sort(key=lambda e: e.center_x)
+                rows.append(current_row)
                 current_row = [elem]
-                current_y = elem.center_y
 
         # Don't forget last row
         if current_row: