feat: simplify layout model selection and archive proposals

Changes:
- Replace PP-Structure 7-slider parameter UI with simple 3-option layout model selector
- Add layout model mapping: chinese (PP-DocLayout-S), default (PubLayNet), cdla
- Add LayoutModelSelector component and zh-TW translations
- Fix "default" model behavior with sentinel value for PubLayNet
- Add gap filling service for OCR track coverage improvement
- Add PP-Structure debug utilities
- Archive completed/incomplete proposals:
  - add-ocr-track-gap-filling (complete)
  - fix-ocr-track-table-rendering (incomplete)
  - simplify-ppstructure-model-selection (22/25 tasks)
- Add new layout model tests, archive old PP-Structure param tests
- Update OpenSpec ocr-processing spec with layout model requirements

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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

backend/app/core/config.py

@@ -3,7 +3,7 @@ Tool_OCR - Configuration Management
 Loads environment variables and provides centralized configuration
 """
 
-from typing import List
+from typing import List, Optional
 from pydantic_settings import BaseSettings
 from pydantic import Field
 from pathlib import Path
@@ -99,6 +99,33 @@ class Settings(BaseSettings):
     text_det_box_thresh: float = Field(default=0.3)  # Lower box threshold for better detection
     text_det_unclip_ratio: float = Field(default=1.2)  # Smaller unclip for tighter text boxes
 
+    # Layout Detection Model Configuration
+    # Available models:
+    # - None (default): Use PP-StructureV3's built-in model (PubLayNet-based)
+    # - "PP-DocLayout-S": Better for Chinese docs, papers, contracts, exams (23 categories)
+    # - "picodet_lcnet_x1_0_fgd_layout_cdla": CDLA-based model for Chinese document layout
+    layout_detection_model_name: Optional[str] = Field(
+        default="PP-DocLayout-S",
+        description="Layout detection model name. Set to 'PP-DocLayout-S' for better Chinese document support."
+    )
+    layout_detection_model_dir: Optional[str] = Field(
+        default=None,
+        description="Custom layout detection model directory. If None, downloads official model."
+    )
+
+    # ===== Gap Filling Configuration =====
+    # Supplements PP-StructureV3 output with raw OCR regions when detection is incomplete
+    gap_filling_enabled: bool = Field(default=True)  # Enable gap filling for OCR track
+    gap_filling_coverage_threshold: float = Field(default=0.7)  # Activate when coverage < 70%
+    gap_filling_iou_threshold: float = Field(default=0.15)  # IoU threshold for coverage detection
+    gap_filling_confidence_threshold: float = Field(default=0.3)  # Min confidence for raw OCR regions
+    gap_filling_dedup_iou_threshold: float = Field(default=0.5)  # IoU threshold for deduplication
+
+    # ===== Debug Configuration =====
+    # Enable debug outputs for PP-StructureV3 analysis
+    pp_structure_debug_enabled: bool = Field(default=True)  # Save debug files for PP-StructureV3
+    pp_structure_debug_visualization: bool = Field(default=True)  # Generate visualization images
+
     # Performance tuning
     use_fp16_inference: bool = Field(default=False)  # Half-precision (if supported)
     enable_cudnn_benchmark: bool = Field(default=True)  # Optimize convolution algorithms

backend/app/routers/tasks.py

@@ -68,7 +68,7 @@ def process_task_ocr(
     use_dual_track: bool = True,
     force_track: Optional[str] = None,
     language: str = 'ch',
-    pp_structure_params: Optional[dict] = None
+    layout_model: Optional[str] = "chinese"
 ):
     """
     Background task to process OCR for a task with dual-track support.
@@ -84,7 +84,7 @@ def process_task_ocr(
         use_dual_track: Enable dual-track processing
         force_track: Force specific track ('ocr' or 'direct')
         language: OCR language code
-        pp_structure_params: Optional custom PP-StructureV3 parameters (dict)
+        layout_model: Layout detection model ('chinese', 'default', 'cdla')
     """
     from app.core.database import SessionLocal
     from app.models.task import Task
@@ -143,7 +143,7 @@ def process_task_ocr(
                 output_dir=result_dir,
                 use_dual_track=use_dual_track,
                 force_track=force_track,
-                pp_structure_params=pp_structure_params
+                layout_model=layout_model
             )
         else:
             # Fall back to traditional processing (no force_track support)
@@ -152,7 +152,7 @@ def process_task_ocr(
                 lang=language,
                 detect_layout=True,
                 output_dir=result_dir,
-                pp_structure_params=pp_structure_params
+                layout_model=layout_model
             )
 
         # Calculate processing time
@@ -717,14 +717,14 @@ async def start_task(
     current_user: User = Depends(get_current_user)
 ):
     """
-    Start processing a pending task with dual-track support and optional PP-StructureV3 parameter tuning
+    Start processing a pending task with dual-track support and layout model selection
 
     - **task_id**: Task UUID
     - **options**: Processing options (in request body):
         - **use_dual_track**: Enable intelligent track selection (default: true)
         - **force_track**: Force specific processing track ('ocr' or 'direct')
         - **language**: OCR language code (default: 'ch')
-        - **pp_structure_params**: Fine-tuning parameters for PP-StructureV3 (OCR track only)
+        - **layout_model**: Layout detection model ('chinese', 'default', 'cdla')
     """
     try:
         # Parse processing options with defaults
@@ -735,11 +735,9 @@ async def start_task(
         force_track = options.force_track.value if options.force_track else None
         language = options.language
 
-        # Extract and convert PP-StructureV3 parameters to dict
-        pp_structure_params = None
-        if options.pp_structure_params:
-            pp_structure_params = options.pp_structure_params.model_dump(exclude_none=True)
-            logger.info(f"Using custom PP-StructureV3 parameters: {pp_structure_params}")
+        # Extract layout model (default to 'chinese' for best Chinese document support)
+        layout_model = options.layout_model.value if options.layout_model else "chinese"
+        logger.info(f"Using layout model: {layout_model}")
 
         # Get task details
         task = task_service.get_task_by_id(
@@ -777,7 +775,7 @@ async def start_task(
             status=TaskStatus.PROCESSING
         )
 
-        # Start OCR processing in background with dual-track parameters and custom PP-StructureV3 params
+        # Start OCR processing in background with dual-track parameters and layout model
         background_tasks.add_task(
             process_task_ocr,
             task_id=task_id,
@@ -787,13 +785,11 @@ async def start_task(
             use_dual_track=use_dual_track,
             force_track=force_track,
             language=language,
-            pp_structure_params=pp_structure_params
+            layout_model=layout_model
         )
 
         logger.info(f"Started OCR processing task {task_id} for user {current_user.email}")
-        logger.info(f"Options: dual_track={use_dual_track}, force_track={force_track}, lang={language}")
-        if pp_structure_params:
-            logger.info(f"Custom PP-StructureV3 params: {pp_structure_params}")
+        logger.info(f"Options: dual_track={use_dual_track}, force_track={force_track}, lang={language}, layout_model={layout_model}")
         return task
 
     except HTTPException:

backend/app/schemas/task.py

@@ -24,6 +24,19 @@ class ProcessingTrackEnum(str, Enum):
     AUTO = "auto"         # Auto-detect best track
 
 
+class LayoutModelEnum(str, Enum):
+    """Layout detection model selection for OCR track.
+
+    Different models are optimized for different document types:
+    - CHINESE: PP-DocLayout-S, optimized for Chinese documents (forms, contracts, invoices)
+    - DEFAULT: PubLayNet-based, optimized for English academic papers
+    - CDLA: CDLA model, specialized Chinese document layout analysis
+    """
+    CHINESE = "chinese"   # PP-DocLayout-S - Best for Chinese documents (recommended)
+    DEFAULT = "default"   # PubLayNet-based - Best for English documents
+    CDLA = "cdla"         # CDLA model - Alternative for Chinese layout
+
+
 class TaskCreate(BaseModel):
     """Task creation request"""
     filename: Optional[str] = Field(None, description="Original filename")
@@ -132,7 +145,11 @@ class UploadResponse(BaseModel):
 # ===== Dual-Track Processing Schemas =====
 
 class PPStructureV3Params(BaseModel):
-    """PP-StructureV3 fine-tuning parameters for OCR track"""
+    """PP-StructureV3 fine-tuning parameters for OCR track.
+
+    DEPRECATED: This class is deprecated and will be removed in a future version.
+    Use `layout_model` parameter in ProcessingOptions instead.
+    """
     layout_detection_threshold: Optional[float] = Field(
         None, ge=0, le=1,
         description="Layout block detection score threshold (lower=more blocks, higher=high confidence only)"
@@ -172,10 +189,10 @@ class ProcessingOptions(BaseModel):
     include_images: bool = Field(default=True, description="Extract and save images")
     confidence_threshold: Optional[float] = Field(None, ge=0, le=1, description="OCR confidence threshold")
 
-    # PP-StructureV3 fine-tuning parameters (OCR track only)
-    pp_structure_params: Optional[PPStructureV3Params] = Field(
-        None,
-        description="Fine-tuning parameters for PP-StructureV3 (OCR track only)"
+    # Layout model selection (OCR track only)
+    layout_model: Optional[LayoutModelEnum] = Field(
+        default=LayoutModelEnum.CHINESE,
+        description="Layout detection model: 'chinese' (recommended for Chinese docs), 'default' (English docs), 'cdla' (Chinese layout)"
     )
 
 

backend/app/services/gap_filling_service.py

@@ -0,0 +1,649 @@
+"""
+Gap Filling Service for OCR Track
+
+This service detects and fills gaps in PP-StructureV3 output by supplementing
+with Raw OCR text regions when significant content loss is detected.
+
+The hybrid approach uses Raw OCR's comprehensive text detection to compensate
+for PP-StructureV3's layout model limitations on certain document types.
+"""
+
+import logging
+from typing import Dict, List, Optional, Tuple, Set, Any
+from dataclasses import dataclass
+
+from app.models.unified_document import (
+    DocumentElement, BoundingBox, ElementType, Dimensions
+)
+from app.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+# Element types that should NOT be supplemented (preserve structural integrity)
+SKIP_ELEMENT_TYPES: Set[ElementType] = {
+    ElementType.TABLE,
+    ElementType.IMAGE,
+    ElementType.FIGURE,
+    ElementType.CHART,
+    ElementType.DIAGRAM,
+    ElementType.HEADER,
+    ElementType.FOOTER,
+    ElementType.FORMULA,
+    ElementType.CODE,
+    ElementType.BARCODE,
+    ElementType.QR_CODE,
+    ElementType.LOGO,
+    ElementType.STAMP,
+    ElementType.SIGNATURE,
+}
+
+
+@dataclass
+class TextRegion:
+    """Represents a raw OCR text region."""
+    text: str
+    bbox: List[float]  # [x0, y0, x1, y1] or polygon format
+    confidence: float
+    page: int = 0
+
+    @property
+    def normalized_bbox(self) -> Tuple[float, float, float, float]:
+        """Get normalized bbox as (x0, y0, x1, y1)."""
+        if not self.bbox:
+            return (0, 0, 0, 0)
+
+        # Check if bbox is nested list format [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
+        # This is common PaddleOCR polygon format
+        if len(self.bbox) >= 1 and isinstance(self.bbox[0], (list, tuple)):
+            # Nested format: extract all x and y coordinates
+            xs = [pt[0] for pt in self.bbox if len(pt) >= 2]
+            ys = [pt[1] for pt in self.bbox if len(pt) >= 2]
+            if xs and ys:
+                return (min(xs), min(ys), max(xs), max(ys))
+            return (0, 0, 0, 0)
+
+        # Flat format
+        if len(self.bbox) == 4:
+            # Simple [x0, y0, x1, y1] format
+            return (float(self.bbox[0]), float(self.bbox[1]),
+                    float(self.bbox[2]), float(self.bbox[3]))
+        elif len(self.bbox) >= 8:
+            # Flat polygon format: [x1, y1, x2, y2, x3, y3, x4, y4]
+            xs = [self.bbox[i] for i in range(0, len(self.bbox), 2)]
+            ys = [self.bbox[i] for i in range(1, len(self.bbox), 2)]
+            return (min(xs), min(ys), max(xs), max(ys))
+
+        return (0, 0, 0, 0)
+
+    @property
+    def center(self) -> Tuple[float, float]:
+        """Get center point of the bbox."""
+        x0, y0, x1, y1 = self.normalized_bbox
+        return ((x0 + x1) / 2, (y0 + y1) / 2)
+
+
+class GapFillingService:
+    """
+    Service for detecting and filling gaps in PP-StructureV3 output.
+
+    This service:
+    1. Calculates coverage of PP-StructureV3 elements over raw OCR regions
+    2. Identifies uncovered raw OCR regions
+    3. Supplements uncovered regions as TEXT elements
+    4. Deduplicates against existing PP-StructureV3 TEXT elements
+    5. Recalculates reading order for the combined result
+    """
+
+    def __init__(
+        self,
+        coverage_threshold: float = None,
+        iou_threshold: float = None,
+        confidence_threshold: float = None,
+        dedup_iou_threshold: float = None,
+        enabled: bool = None
+    ):
+        """
+        Initialize the gap filling service.
+
+        Args:
+            coverage_threshold: Coverage ratio below which gap filling activates (default: 0.7)
+            iou_threshold: IoU threshold for coverage detection (default: 0.15)
+            confidence_threshold: Minimum confidence for raw OCR regions (default: 0.3)
+            dedup_iou_threshold: IoU threshold for deduplication (default: 0.5)
+            enabled: Whether gap filling is enabled (default: True)
+        """
+        self.coverage_threshold = coverage_threshold if coverage_threshold is not None else getattr(
+            settings, 'gap_filling_coverage_threshold', 0.7
+        )
+        self.iou_threshold = iou_threshold if iou_threshold is not None else getattr(
+            settings, 'gap_filling_iou_threshold', 0.15
+        )
+        self.confidence_threshold = confidence_threshold if confidence_threshold is not None else getattr(
+            settings, 'gap_filling_confidence_threshold', 0.3
+        )
+        self.dedup_iou_threshold = dedup_iou_threshold if dedup_iou_threshold is not None else getattr(
+            settings, 'gap_filling_dedup_iou_threshold', 0.5
+        )
+        self.enabled = enabled if enabled is not None else getattr(
+            settings, 'gap_filling_enabled', True
+        )
+
+    def should_activate(
+        self,
+        raw_ocr_regions: List[TextRegion],
+        pp_structure_elements: List[DocumentElement]
+    ) -> Tuple[bool, float]:
+        """
+        Determine if gap filling should be activated.
+
+        Gap filling activates when:
+        1. Coverage ratio is below threshold (default: 70%)
+        2. OR element count disparity is significant
+
+        Args:
+            raw_ocr_regions: List of raw OCR text regions
+            pp_structure_elements: List of PP-StructureV3 elements
+
+        Returns:
+            Tuple of (should_activate, coverage_ratio)
+        """
+        if not self.enabled:
+            return False, 1.0
+
+        if not raw_ocr_regions:
+            return False, 1.0
+
+        # Calculate coverage
+        covered_count = 0
+        for region in raw_ocr_regions:
+            if self._is_region_covered(region, pp_structure_elements):
+                covered_count += 1
+
+        coverage_ratio = covered_count / len(raw_ocr_regions)
+
+        # Check activation conditions
+        should_activate = coverage_ratio < self.coverage_threshold
+
+        if should_activate:
+            logger.info(
+                f"Gap filling activated: coverage={coverage_ratio:.2%} < threshold={self.coverage_threshold:.0%}, "
+                f"raw_regions={len(raw_ocr_regions)}, pp_elements={len(pp_structure_elements)}"
+            )
+        else:
+            logger.debug(
+                f"Gap filling not needed: coverage={coverage_ratio:.2%} >= threshold={self.coverage_threshold:.0%}"
+            )
+
+        return should_activate, coverage_ratio
+
+    def find_uncovered_regions(
+        self,
+        raw_ocr_regions: List[TextRegion],
+        pp_structure_elements: List[DocumentElement]
+    ) -> List[TextRegion]:
+        """
+        Find raw OCR regions not covered by PP-StructureV3 elements.
+
+        A region is considered covered if:
+        1. Its center point falls inside any PP-StructureV3 element bbox, OR
+        2. IoU with any PP-StructureV3 element exceeds iou_threshold
+
+        Args:
+            raw_ocr_regions: List of raw OCR text regions
+            pp_structure_elements: List of PP-StructureV3 elements
+
+        Returns:
+            List of uncovered raw OCR regions
+        """
+        uncovered = []
+
+        for region in raw_ocr_regions:
+            # Skip low confidence regions
+            if region.confidence < self.confidence_threshold:
+                continue
+
+            if not self._is_region_covered(region, pp_structure_elements):
+                uncovered.append(region)
+
+        logger.debug(f"Found {len(uncovered)} uncovered regions out of {len(raw_ocr_regions)}")
+        return uncovered
+
+    def _is_region_covered(
+        self,
+        region: TextRegion,
+        pp_structure_elements: List[DocumentElement]
+    ) -> bool:
+        """
+        Check if a raw OCR region is covered by any PP-StructureV3 element.
+
+        Args:
+            region: Raw OCR text region
+            pp_structure_elements: List of PP-StructureV3 elements
+
+        Returns:
+            True if the region is covered
+        """
+        center_x, center_y = region.center
+        region_bbox = region.normalized_bbox
+
+        for element in pp_structure_elements:
+            elem_bbox = (
+                element.bbox.x0, element.bbox.y0,
+                element.bbox.x1, element.bbox.y1
+            )
+
+            # Check 1: Center point falls inside element bbox
+            if self._point_in_bbox(center_x, center_y, elem_bbox):
+                return True
+
+            # Check 2: IoU exceeds threshold
+            iou = self._calculate_iou(region_bbox, elem_bbox)
+            if iou > self.iou_threshold:
+                return True
+
+        return False
+
+    def deduplicate_regions(
+        self,
+        uncovered_regions: List[TextRegion],
+        pp_structure_elements: List[DocumentElement]
+    ) -> List[TextRegion]:
+        """
+        Remove regions that highly overlap with existing PP-StructureV3 TEXT elements.
+
+        Args:
+            uncovered_regions: List of uncovered raw OCR regions
+            pp_structure_elements: List of PP-StructureV3 elements
+
+        Returns:
+            Deduplicated list of regions
+        """
+        # Get TEXT elements only for deduplication
+        text_elements = [
+            e for e in pp_structure_elements
+            if e.type not in SKIP_ELEMENT_TYPES
+        ]
+
+        deduplicated = []
+        for region in uncovered_regions:
+            region_bbox = region.normalized_bbox
+            is_duplicate = False
+
+            for element in text_elements:
+                elem_bbox = (
+                    element.bbox.x0, element.bbox.y0,
+                    element.bbox.x1, element.bbox.y1
+                )
+
+                iou = self._calculate_iou(region_bbox, elem_bbox)
+                if iou > self.dedup_iou_threshold:
+                    logger.debug(
+                        f"Skipping duplicate region (IoU={iou:.2f}): '{region.text[:30]}...'"
+                    )
+                    is_duplicate = True
+                    break
+
+            if not is_duplicate:
+                deduplicated.append(region)
+
+        removed_count = len(uncovered_regions) - len(deduplicated)
+        if removed_count > 0:
+            logger.debug(f"Removed {removed_count} duplicate regions")
+
+        return deduplicated
+
+    def convert_regions_to_elements(
+        self,
+        regions: List[TextRegion],
+        page_number: int,
+        start_element_id: int = 0
+    ) -> List[DocumentElement]:
+        """
+        Convert raw OCR regions to DocumentElement objects.
+
+        Args:
+            regions: List of raw OCR regions to convert
+            page_number: Page number for the elements
+            start_element_id: Starting ID counter for elements
+
+        Returns:
+            List of DocumentElement objects
+        """
+        elements = []
+
+        for idx, region in enumerate(regions):
+            x0, y0, x1, y1 = region.normalized_bbox
+
+            element = DocumentElement(
+                element_id=f"gap_fill_{page_number}_{start_element_id + idx}",
+                type=ElementType.TEXT,
+                content=region.text,
+                bbox=BoundingBox(x0=x0, y0=y0, x1=x1, y1=y1),
+                confidence=region.confidence,
+                metadata={
+                    'source': 'gap_filling',
+                    'original_confidence': region.confidence
+                }
+            )
+            elements.append(element)
+
+        return elements
+
+    def recalculate_reading_order(
+        self,
+        elements: List[DocumentElement]
+    ) -> List[int]:
+        """
+        Recalculate reading order for elements based on position.
+
+        Sorts elements by y0 (top to bottom) then x0 (left to right).
+
+        Args:
+            elements: List of DocumentElement objects
+
+        Returns:
+            List of element indices in reading order
+        """
+        # Create indexed list with position info
+        indexed_elements = [
+            (idx, e.bbox.y0, e.bbox.x0)
+            for idx, e in enumerate(elements)
+        ]
+
+        # Sort by y0 then x0
+        indexed_elements.sort(key=lambda x: (x[1], x[2]))
+
+        # Return indices in reading order
+        return [idx for idx, _, _ in indexed_elements]
+
+    def merge_adjacent_regions(
+        self,
+        regions: List[TextRegion],
+        max_horizontal_gap: float = 20.0,
+        max_vertical_gap: float = 5.0
+    ) -> List[TextRegion]:
+        """
+        Merge fragmented adjacent regions on the same line.
+
+        This is optional and can reduce fragmentation from raw OCR.
+
+        Args:
+            regions: List of raw OCR regions
+            max_horizontal_gap: Maximum horizontal gap to merge (pixels)
+            max_vertical_gap: Maximum vertical gap to merge (pixels)
+
+        Returns:
+            List of merged regions
+        """
+        if not regions:
+            return regions
+
+        # Sort by y0, then x0
+        sorted_regions = sorted(
+            regions,
+            key=lambda r: (r.normalized_bbox[1], r.normalized_bbox[0])
+        )
+
+        merged = []
+        current = sorted_regions[0]
+
+        for next_region in sorted_regions[1:]:
+            curr_bbox = current.normalized_bbox
+            next_bbox = next_region.normalized_bbox
+
+            # Check if on same line (vertical overlap)
+            curr_y_center = (curr_bbox[1] + curr_bbox[3]) / 2
+            next_y_center = (next_bbox[1] + next_bbox[3]) / 2
+            vertical_distance = abs(curr_y_center - next_y_center)
+
+            # Check horizontal gap
+            horizontal_gap = next_bbox[0] - curr_bbox[2]
+
+            if (vertical_distance < max_vertical_gap and
+                    0 <= horizontal_gap <= max_horizontal_gap):
+                # Merge regions
+                merged_bbox = [
+                    min(curr_bbox[0], next_bbox[0]),
+                    min(curr_bbox[1], next_bbox[1]),
+                    max(curr_bbox[2], next_bbox[2]),
+                    max(curr_bbox[3], next_bbox[3])
+                ]
+                current = TextRegion(
+                    text=current.text + " " + next_region.text,
+                    bbox=merged_bbox,
+                    confidence=min(current.confidence, next_region.confidence),
+                    page=current.page
+                )
+            else:
+                merged.append(current)
+                current = next_region
+
+        merged.append(current)
+
+        if len(merged) < len(regions):
+            logger.debug(f"Merged {len(regions)} regions into {len(merged)}")
+
+        return merged
+
+    def fill_gaps(
+        self,
+        raw_ocr_regions: List[Dict[str, Any]],
+        pp_structure_elements: List[DocumentElement],
+        page_number: int,
+        ocr_dimensions: Optional[Dict[str, Any]] = None,
+        pp_dimensions: Optional[Dimensions] = None
+    ) -> Tuple[List[DocumentElement], Dict[str, Any]]:
+        """
+        Main entry point: detect gaps and fill with raw OCR regions.
+
+        Args:
+            raw_ocr_regions: Raw OCR results (list of dicts with text, bbox, confidence)
+            pp_structure_elements: PP-StructureV3 elements
+            page_number: Current page number
+            ocr_dimensions: OCR image dimensions for coordinate alignment
+            pp_dimensions: PP-Structure dimensions for coordinate alignment
+
+        Returns:
+            Tuple of (supplemented_elements, statistics)
+        """
+        statistics = {
+            'enabled': self.enabled,
+            'activated': False,
+            'coverage_ratio': 1.0,
+            'raw_ocr_count': len(raw_ocr_regions),
+            'pp_structure_count': len(pp_structure_elements),
+            'uncovered_count': 0,
+            'deduplicated_count': 0,
+            'supplemented_count': 0
+        }
+
+        if not self.enabled:
+            logger.debug("Gap filling is disabled")
+            return [], statistics
+
+        # Convert raw OCR regions to TextRegion objects
+        text_regions = self._convert_raw_ocr_regions(
+            raw_ocr_regions, page_number, ocr_dimensions, pp_dimensions
+        )
+
+        if not text_regions:
+            logger.debug("No valid text regions to process")
+            return [], statistics
+
+        # Check if gap filling should activate
+        should_activate, coverage_ratio = self.should_activate(
+            text_regions, pp_structure_elements
+        )
+        statistics['coverage_ratio'] = coverage_ratio
+        statistics['activated'] = should_activate
+
+        if not should_activate:
+            return [], statistics
+
+        # Find uncovered regions
+        uncovered = self.find_uncovered_regions(text_regions, pp_structure_elements)
+        statistics['uncovered_count'] = len(uncovered)
+
+        if not uncovered:
+            logger.debug("No uncovered regions found")
+            return [], statistics
+
+        # Deduplicate against existing TEXT elements
+        deduplicated = self.deduplicate_regions(uncovered, pp_structure_elements)
+        statistics['deduplicated_count'] = len(deduplicated)
+
+        if not deduplicated:
+            logger.debug("All uncovered regions were duplicates")
+            return [], statistics
+
+        # Optional: Merge adjacent regions
+        # merged = self.merge_adjacent_regions(deduplicated)
+
+        # Convert to DocumentElements
+        start_id = len(pp_structure_elements)
+        supplemented = self.convert_regions_to_elements(
+            deduplicated, page_number, start_id
+        )
+        statistics['supplemented_count'] = len(supplemented)
+
+        logger.info(
+            f"Gap filling complete: supplemented {len(supplemented)} elements "
+            f"(coverage: {coverage_ratio:.2%} -> estimated {(coverage_ratio + len(supplemented)/len(text_regions) if text_regions else 0):.2%})"
+        )
+
+        return supplemented, statistics
+
+    def _convert_raw_ocr_regions(
+        self,
+        raw_regions: List[Dict[str, Any]],
+        page_number: int,
+        ocr_dimensions: Optional[Dict[str, Any]] = None,
+        pp_dimensions: Optional[Dimensions] = None
+    ) -> List[TextRegion]:
+        """
+        Convert raw OCR region dicts to TextRegion objects.
+
+        Handles coordinate alignment if dimensions are provided.
+
+        Args:
+            raw_regions: List of raw OCR region dictionaries
+            page_number: Current page number
+            ocr_dimensions: OCR image dimensions
+            pp_dimensions: PP-Structure dimensions
+
+        Returns:
+            List of TextRegion objects
+        """
+        text_regions = []
+
+        # Calculate scale factors if needed
+        scale_x, scale_y = 1.0, 1.0
+        if ocr_dimensions and pp_dimensions:
+            ocr_width = ocr_dimensions.get('width', 0)
+            ocr_height = ocr_dimensions.get('height', 0)
+
+            if ocr_width > 0 and pp_dimensions.width > 0:
+                scale_x = pp_dimensions.width / ocr_width
+            if ocr_height > 0 and pp_dimensions.height > 0:
+                scale_y = pp_dimensions.height / ocr_height
+
+            if scale_x != 1.0 or scale_y != 1.0:
+                logger.debug(f"Coordinate scaling: x={scale_x:.3f}, y={scale_y:.3f}")
+
+        for region in raw_regions:
+            text = region.get('text', '')
+            if not text or not text.strip():
+                continue
+
+            confidence = region.get('confidence', 0.0)
+            bbox_raw = region.get('bbox', [])
+
+            # Normalize bbox
+            if isinstance(bbox_raw, dict):
+                # Dict format: {x_min, y_min, x_max, y_max}
+                bbox = [
+                    bbox_raw.get('x_min', 0),
+                    bbox_raw.get('y_min', 0),
+                    bbox_raw.get('x_max', 0),
+                    bbox_raw.get('y_max', 0)
+                ]
+            elif isinstance(bbox_raw, (list, tuple)):
+                bbox = list(bbox_raw)
+            else:
+                continue
+
+            # Apply scaling if needed
+            if scale_x != 1.0 or scale_y != 1.0:
+                # Check if nested list format [[x1,y1], [x2,y2], ...]
+                if len(bbox) >= 1 and isinstance(bbox[0], (list, tuple)):
+                    bbox = [
+                        [pt[0] * scale_x, pt[1] * scale_y]
+                        for pt in bbox if len(pt) >= 2
+                    ]
+                elif len(bbox) == 4 and not isinstance(bbox[0], (list, tuple)):
+                    # Simple [x0, y0, x1, y1] format
+                    bbox = [
+                        bbox[0] * scale_x, bbox[1] * scale_y,
+                        bbox[2] * scale_x, bbox[3] * scale_y
+                    ]
+                elif len(bbox) >= 8:
+                    # Flat polygon format [x1, y1, x2, y2, ...]
+                    bbox = [
+                        bbox[i] * (scale_x if i % 2 == 0 else scale_y)
+                        for i in range(len(bbox))
+                    ]
+
+            text_regions.append(TextRegion(
+                text=text,
+                bbox=bbox,
+                confidence=confidence,
+                page=page_number
+            ))
+
+        return text_regions
+
+    @staticmethod
+    def _point_in_bbox(
+        x: float, y: float,
+        bbox: Tuple[float, float, float, float]
+    ) -> bool:
+        """Check if point (x, y) is inside bbox (x0, y0, x1, y1)."""
+        x0, y0, x1, y1 = bbox
+        return x0 <= x <= x1 and y0 <= y <= y1
+
+    @staticmethod
+    def _calculate_iou(
+        bbox1: Tuple[float, float, float, float],
+        bbox2: Tuple[float, float, float, float]
+    ) -> float:
+        """
+        Calculate Intersection over Union (IoU) of two bboxes.
+
+        Args:
+            bbox1: First bbox (x0, y0, x1, y1)
+            bbox2: Second bbox (x0, y0, x1, y1)
+
+        Returns:
+            IoU value between 0 and 1
+        """
+        # Calculate intersection
+        x0 = max(bbox1[0], bbox2[0])
+        y0 = max(bbox1[1], bbox2[1])
+        x1 = min(bbox1[2], bbox2[2])
+        y1 = min(bbox1[3], bbox2[3])
+
+        if x1 <= x0 or y1 <= y0:
+            return 0.0
+
+        intersection = (x1 - x0) * (y1 - y0)
+
+        # Calculate union
+        area1 = (bbox1[2] - bbox1[0]) * (bbox1[3] - bbox1[1])
+        area2 = (bbox2[2] - bbox2[0]) * (bbox2[3] - bbox2[1])
+        union = area1 + area2 - intersection
+
+        if union <= 0:
+            return 0.0
+
+        return intersection / union

backend/app/services/ocr_service.py

@@ -46,6 +46,19 @@ except ImportError as e:
 
 logger = logging.getLogger(__name__)
 
+# Sentinel value for "use PubLayNet default" - explicitly NO model specification
+_USE_PUBLAYNET_DEFAULT = "__USE_PUBLAYNET_DEFAULT__"
+
+# Layout model mapping: user-friendly names to actual model names
+# - "chinese": PP-DocLayout-S - Best for Chinese documents (forms, contracts, invoices)
+# - "default": PubLayNet-based default model - Best for English documents
+# - "cdla": picodet_lcnet_x1_0_fgd_layout_cdla - Alternative for Chinese layout
+LAYOUT_MODEL_MAPPING = {
+    "chinese": "PP-DocLayout-S",
+    "default": _USE_PUBLAYNET_DEFAULT,  # Uses default PubLayNet-based model (no custom model)
+    "cdla": "picodet_lcnet_x1_0_fgd_layout_cdla",
+}
+
 
 class OCRService:
     """
@@ -436,77 +449,45 @@ class OCRService:
 
         return self.ocr_engines[lang]
 
-    def _ensure_structure_engine(self, custom_params: Optional[Dict[str, any]] = None) -> PPStructureV3:
+    def _ensure_structure_engine(self, layout_model: Optional[str] = None) -> PPStructureV3:
         """
         Get or create PP-Structure engine for layout analysis with GPU support.
-        Supports custom parameters that override default settings.
+        Supports layout model selection for different document types.
 
         Args:
-            custom_params: Optional dictionary of custom PP-StructureV3 parameters.
-                          If provided, creates a new engine instance (not cached).
-                          Supported keys: layout_detection_threshold, layout_nms_threshold,
-                          layout_merge_bboxes_mode, layout_unclip_ratio, text_det_thresh,
-                          text_det_box_thresh, text_det_unclip_ratio
+            layout_model: Layout detection model selection:
+                         - "chinese": PP-DocLayout-S (best for Chinese documents)
+                         - "default": PubLayNet-based (best for English documents)
+                         - "cdla": CDLA model (alternative for Chinese layout)
+                         - None: Use config default
 
         Returns:
             PPStructure engine instance
         """
-        # If custom params provided, create a new engine instance (don't use cache)
-        if custom_params:
-            logger.info(f"Creating PP-StructureV3 engine with custom parameters (GPU: {self.use_gpu})")
-            logger.info(f"Custom params: {custom_params}")
+        # Resolve layout model name from user-friendly name
+        resolved_model_name = None
+        use_publaynet_default = False  # Flag to explicitly use PubLayNet default (no model param)
 
-            try:
-                # Base configuration from settings
-                use_chart = settings.enable_chart_recognition
-                use_formula = settings.enable_formula_recognition
-                use_table = settings.enable_table_recognition
+        if layout_model:
+            resolved_model_name = LAYOUT_MODEL_MAPPING.get(layout_model)
+            if layout_model not in LAYOUT_MODEL_MAPPING:
+                logger.warning(f"Unknown layout model '{layout_model}', using config default")
+                resolved_model_name = settings.layout_detection_model_name
+            elif resolved_model_name == _USE_PUBLAYNET_DEFAULT:
+                # User explicitly selected "default" - use PubLayNet without custom model
+                use_publaynet_default = True
+                resolved_model_name = None
+                logger.info(f"Using layout model: {layout_model} -> PubLayNet default (no custom model)")
+            else:
+                logger.info(f"Using layout model: {layout_model} -> {resolved_model_name}")
 
-                # Parameter priority: custom > settings default
-                layout_threshold = custom_params.get('layout_detection_threshold', settings.layout_detection_threshold)
-                layout_nms = custom_params.get('layout_nms_threshold', settings.layout_nms_threshold)
-                layout_merge = custom_params.get('layout_merge_bboxes_mode', settings.layout_merge_mode)
-                layout_unclip = custom_params.get('layout_unclip_ratio', settings.layout_unclip_ratio)
-                text_thresh = custom_params.get('text_det_thresh', settings.text_det_thresh)
-                text_box_thresh = custom_params.get('text_det_box_thresh', settings.text_det_box_thresh)
-                text_unclip = custom_params.get('text_det_unclip_ratio', settings.text_det_unclip_ratio)
+        # Check if we need to recreate the engine due to different model
+        current_model = getattr(self, '_current_layout_model', None)
+        if self.structure_engine is not None and layout_model and layout_model != current_model:
+            logger.info(f"Layout model changed from {current_model} to {layout_model}, recreating engine")
+            self.structure_engine = None  # Force recreation
 
-                logger.info(f"PP-StructureV3 config: table={use_table}, formula={use_formula}, chart={use_chart}")
-                logger.info(f"Layout config: threshold={layout_threshold}, nms={layout_nms}, merge={layout_merge}, unclip={layout_unclip}")
-                logger.info(f"Text detection: thresh={text_thresh}, box_thresh={text_box_thresh}, unclip={text_unclip}")
-
-                # Create temporary engine with custom params (not cached)
-                custom_engine = PPStructureV3(
-                    use_doc_orientation_classify=False,
-                    use_doc_unwarping=False,
-                    use_textline_orientation=False,
-                    use_table_recognition=use_table,
-                    use_formula_recognition=use_formula,
-                    use_chart_recognition=use_chart,
-                    layout_threshold=layout_threshold,
-                    layout_nms=layout_nms,
-                    layout_unclip_ratio=layout_unclip,
-                    layout_merge_bboxes_mode=layout_merge,
-                    text_det_thresh=text_thresh,
-                    text_det_box_thresh=text_box_thresh,
-                    text_det_unclip_ratio=text_unclip,
-                )
-
-                logger.info(f"PP-StructureV3 engine with custom params ready (PaddlePaddle {paddle.__version__}, {'GPU' if self.use_gpu else 'CPU'} mode)")
-
-                # Check GPU memory after loading
-                if self.use_gpu and settings.enable_memory_optimization:
-                    self._check_gpu_memory_usage()
-
-                return custom_engine
-
-            except Exception as e:
-                logger.error(f"Failed to create PP-StructureV3 engine with custom params: {e}")
-                # Fall back to default cached engine
-                logger.warning("Falling back to default cached engine")
-                custom_params = None  # Clear custom params to use cached engine
-
-        # Use cached default engine
+        # Use cached engine or create new one
         if self.structure_engine is None:
             logger.info(f"Initializing PP-StructureV3 engine (GPU: {self.use_gpu})")
 
@@ -524,28 +505,51 @@ class OCRService:
                 text_box_thresh = settings.text_det_box_thresh
                 text_unclip = settings.text_det_unclip_ratio
 
+                # Layout model configuration:
+                # - If use_publaynet_default: don't specify any model (use PubLayNet default)
+                # - If resolved_model_name: use the specified model
+                # - Otherwise: use config default
+                if use_publaynet_default:
+                    layout_model_name = None  # Explicitly no model = PubLayNet default
+                elif resolved_model_name:
+                    layout_model_name = resolved_model_name
+                else:
+                    layout_model_name = settings.layout_detection_model_name
+                layout_model_dir = settings.layout_detection_model_dir
+
                 logger.info(f"PP-StructureV3 config: table={use_table}, formula={use_formula}, chart={use_chart}")
+                logger.info(f"Layout model: name={layout_model_name}, dir={layout_model_dir}")
                 logger.info(f"Layout config: threshold={layout_threshold}, nms={layout_nms}, merge={layout_merge}, unclip={layout_unclip}")
                 logger.info(f"Text detection: thresh={text_thresh}, box_thresh={text_box_thresh}, unclip={text_unclip}")
 
-                self.structure_engine = PPStructureV3(
-                    use_doc_orientation_classify=False,
-                    use_doc_unwarping=False,
-                    use_textline_orientation=False,
-                    use_table_recognition=use_table,
-                    use_formula_recognition=use_formula,
-                    use_chart_recognition=use_chart,
-                    layout_threshold=layout_threshold,
-                    layout_nms=layout_nms,
-                    layout_unclip_ratio=layout_unclip,
-                    layout_merge_bboxes_mode=layout_merge,  # Use 'small' to minimize merging
-                    text_det_thresh=text_thresh,
-                    text_det_box_thresh=text_box_thresh,
-                    text_det_unclip_ratio=text_unclip,
-                )
+                # Build PPStructureV3 kwargs
+                pp_kwargs = {
+                    'use_doc_orientation_classify': False,
+                    'use_doc_unwarping': False,
+                    'use_textline_orientation': False,
+                    'use_table_recognition': use_table,
+                    'use_formula_recognition': use_formula,
+                    'use_chart_recognition': use_chart,
+                    'layout_threshold': layout_threshold,
+                    'layout_nms': layout_nms,
+                    'layout_unclip_ratio': layout_unclip,
+                    'layout_merge_bboxes_mode': layout_merge,
+                    'text_det_thresh': text_thresh,
+                    'text_det_box_thresh': text_box_thresh,
+                    'text_det_unclip_ratio': text_unclip,
+                }
+
+                # Add layout model configuration if specified
+                if layout_model_name:
+                    pp_kwargs['layout_detection_model_name'] = layout_model_name
+                if layout_model_dir:
+                    pp_kwargs['layout_detection_model_dir'] = layout_model_dir
+
+                self.structure_engine = PPStructureV3(**pp_kwargs)
 
                 # Track model loading for cache management
                 self._model_last_used['structure'] = datetime.now()
+                self._current_layout_model = layout_model  # Track current model for recreation check
 
                 logger.info(f"PP-StructureV3 engine ready (PaddlePaddle {paddle.__version__}, {'GPU' if self.use_gpu else 'CPU'} mode)")
 
@@ -565,17 +569,27 @@ class OCRService:
                     use_formula = settings.enable_formula_recognition
                     use_table = settings.enable_table_recognition
                     layout_threshold = settings.layout_detection_threshold
+                    layout_model_name = settings.layout_detection_model_name
+                    layout_model_dir = settings.layout_detection_model_dir
 
-                    self.structure_engine = PPStructureV3(
-                        use_doc_orientation_classify=False,
-                        use_doc_unwarping=False,
-                        use_textline_orientation=False,
-                        use_table_recognition=use_table,
-                        use_formula_recognition=use_formula,
-                        use_chart_recognition=use_chart,
-                        layout_threshold=layout_threshold,
-                    )
-                    logger.info("PP-StructureV3 engine ready (CPU mode - fallback)")
+                    # Build CPU fallback kwargs
+                    cpu_kwargs = {
+                        'use_doc_orientation_classify': False,
+                        'use_doc_unwarping': False,
+                        'use_textline_orientation': False,
+                        'use_table_recognition': use_table,
+                        'use_formula_recognition': use_formula,
+                        'use_chart_recognition': use_chart,
+                        'layout_threshold': layout_threshold,
+                    }
+                    if layout_model_name:
+                        cpu_kwargs['layout_detection_model_name'] = layout_model_name
+                    if layout_model_dir:
+                        cpu_kwargs['layout_detection_model_dir'] = layout_model_dir
+
+                    self.structure_engine = PPStructureV3(**cpu_kwargs)
+                    self._current_layout_model = layout_model  # Track current model for recreation check
+                    logger.info(f"PP-StructureV3 engine ready (CPU mode - fallback, layout_model={layout_model_name})")
                 else:
                     raise
 
@@ -813,7 +827,7 @@ class OCRService:
         confidence_threshold: Optional[float] = None,
         output_dir: Optional[Path] = None,
         current_page: int = 0,
-        pp_structure_params: Optional[Dict[str, any]] = None
+        layout_model: Optional[str] = None
     ) -> Dict:
         """
         Process single image with OCR and layout analysis
@@ -825,7 +839,7 @@ class OCRService:
             confidence_threshold: Minimum confidence threshold (uses default if None)
             output_dir: Optional output directory for saving extracted images
             current_page: Current page number (0-based) for multi-page documents
-            pp_structure_params: Optional custom PP-StructureV3 parameters
+            layout_model: Layout detection model ('chinese', 'default', 'cdla')
 
         Returns:
             Dictionary with OCR results and metadata
@@ -894,7 +908,7 @@ class OCRService:
                         confidence_threshold=confidence_threshold,
                         output_dir=output_dir,
                         current_page=page_num - 1,  # Convert to 0-based page number for layout data
-                        pp_structure_params=pp_structure_params
+                        layout_model=layout_model
                     )
 
                     # Accumulate results
@@ -1040,7 +1054,7 @@ class OCRService:
                     image_path,
                     output_dir=output_dir,
                     current_page=current_page,
-                    pp_structure_params=pp_structure_params
+                    layout_model=layout_model
                 )
 
             # Generate Markdown
@@ -1078,6 +1092,38 @@ class OCRService:
                     'height': ocr_height
                 }]
 
+                # Generate PP-StructureV3 debug outputs if enabled
+                if settings.pp_structure_debug_enabled and output_dir:
+                    try:
+                        from app.services.pp_structure_debug import PPStructureDebug
+                        debug_service = PPStructureDebug(output_dir)
+
+                        # Save raw results as JSON
+                        debug_service.save_raw_results(
+                            pp_structure_results={
+                                'elements': layout_data.get('elements', []),
+                                'total_elements': layout_data.get('total_elements', 0),
+                                'element_types': layout_data.get('element_types', {}),
+                                'reading_order': layout_data.get('reading_order', []),
+                                'enhanced': True,
+                                'has_parsing_res_list': True
+                            },
+                            raw_ocr_regions=text_regions,
+                            filename_prefix=image_path.stem
+                        )
+
+                        # Generate visualization if enabled
+                        if settings.pp_structure_debug_visualization:
+                            debug_service.generate_visualization(
+                                image_path=image_path,
+                                pp_structure_elements=layout_data.get('elements', []),
+                                raw_ocr_regions=text_regions,
+                                filename_prefix=image_path.stem
+                            )
+                            logger.info(f"Generated PP-StructureV3 debug outputs for {image_path.name}")
+                    except Exception as debug_error:
+                        logger.warning(f"Failed to generate debug outputs: {debug_error}")
+
             logger.info(
                 f"OCR completed: {image_path.name} - "
                 f"{len(text_regions)} regions, "
@@ -1164,7 +1210,7 @@ class OCRService:
         image_path: Path,
         output_dir: Optional[Path] = None,
         current_page: int = 0,
-        pp_structure_params: Optional[Dict[str, any]] = None
+        layout_model: Optional[str] = None
     ) -> Tuple[Optional[Dict], List[Dict]]:
         """
         Analyze document layout using PP-StructureV3 with enhanced element extraction
@@ -1173,7 +1219,7 @@ class OCRService:
             image_path: Path to image file
             output_dir: Optional output directory for saving extracted images (defaults to image_path.parent)
             current_page: Current page number (0-based) for multi-page documents
-            pp_structure_params: Optional custom PP-StructureV3 parameters
+            layout_model: Layout detection model ('chinese', 'default', 'cdla')
 
         Returns:
             Tuple of (layout_data, images_metadata)
@@ -1191,7 +1237,7 @@ class OCRService:
                     f"Mode: {'CPU fallback' if self._cpu_fallback_active else 'GPU'}"
                 )
 
-            structure_engine = self._ensure_structure_engine(pp_structure_params)
+            structure_engine = self._ensure_structure_engine(layout_model)
 
             # Try enhanced processing first
             try:
@@ -1425,7 +1471,7 @@ class OCRService:
         confidence_threshold: Optional[float] = None,
         output_dir: Optional[Path] = None,
         force_track: Optional[str] = None,
-        pp_structure_params: Optional[Dict[str, any]] = None
+        layout_model: Optional[str] = None
     ) -> Union[UnifiedDocument, Dict]:
         """
         Process document using dual-track approach.
@@ -1437,7 +1483,7 @@ class OCRService:
             confidence_threshold: Minimum confidence threshold
             output_dir: Optional output directory for extracted images
             force_track: Force specific track ("ocr" or "direct"), None for auto-detection
-            pp_structure_params: Optional custom PP-StructureV3 parameters (used for OCR track only)
+            layout_model: Layout detection model ('chinese', 'default', 'cdla') (used for OCR track only)
 
         Returns:
             UnifiedDocument if dual-track is enabled, Dict otherwise
@@ -1445,7 +1491,7 @@ class OCRService:
         if not self.dual_track_enabled:
             # Fallback to traditional OCR processing
             return self.process_file_traditional(
-                file_path, lang, detect_layout, confidence_threshold, output_dir, pp_structure_params
+                file_path, lang, detect_layout, confidence_threshold, output_dir, layout_model
             )
 
         start_time = datetime.now()
@@ -1517,7 +1563,7 @@ class OCRService:
                             ocr_result = self.process_file_traditional(
                                 actual_file_path, lang, detect_layout=True,
                                 confidence_threshold=confidence_threshold,
-                                output_dir=output_dir, pp_structure_params=pp_structure_params
+                                output_dir=output_dir, layout_model=layout_model
                             )
 
                             # Convert OCR result to extract images
@@ -1550,7 +1596,7 @@ class OCRService:
                 # Use OCR for scanned documents, images, etc.
                 logger.info("Using OCR track (PaddleOCR)")
                 ocr_result = self.process_file_traditional(
-                    file_path, lang, detect_layout, confidence_threshold, output_dir, pp_structure_params
+                    file_path, lang, detect_layout, confidence_threshold, output_dir, layout_model
                 )
 
                 # Convert OCR result to UnifiedDocument using the converter
@@ -1580,7 +1626,7 @@ class OCRService:
             logger.error(f"Error in dual-track processing: {e}")
             # Fallback to traditional OCR
             return self.process_file_traditional(
-                file_path, lang, detect_layout, confidence_threshold, output_dir, pp_structure_params
+                file_path, lang, detect_layout, confidence_threshold, output_dir, layout_model
             )
 
     def _merge_ocr_images_into_direct(
@@ -1659,7 +1705,7 @@ class OCRService:
         detect_layout: bool = True,
         confidence_threshold: Optional[float] = None,
         output_dir: Optional[Path] = None,
-        pp_structure_params: Optional[Dict[str, any]] = None
+        layout_model: Optional[str] = None
     ) -> Dict:
         """
         Traditional OCR processing (legacy method).
@@ -1670,7 +1716,7 @@ class OCRService:
             detect_layout: Whether to perform layout analysis
             confidence_threshold: Minimum confidence threshold
             output_dir: Optional output directory
-            pp_structure_params: Optional custom PP-StructureV3 parameters
+            layout_model: Layout detection model ('chinese', 'default', 'cdla')
 
         Returns:
             Dictionary with OCR results in legacy format
@@ -1683,7 +1729,7 @@ class OCRService:
             all_results = []
             for i, image_path in enumerate(image_paths):
                 result = self.process_image(
-                    image_path, lang, detect_layout, confidence_threshold, output_dir, i, pp_structure_params
+                    image_path, lang, detect_layout, confidence_threshold, output_dir, i, layout_model
                 )
                 all_results.append(result)
 
@@ -1699,7 +1745,7 @@ class OCRService:
         else:
             # Single image or other file
             return self.process_image(
-                file_path, lang, detect_layout, confidence_threshold, output_dir, 0, pp_structure_params
+                file_path, lang, detect_layout, confidence_threshold, output_dir, 0, layout_model
             )
 
     def _combine_results(self, results: List[Dict]) -> Dict:
@@ -1784,7 +1830,7 @@ class OCRService:
         output_dir: Optional[Path] = None,
         use_dual_track: bool = True,
         force_track: Optional[str] = None,
-        pp_structure_params: Optional[Dict[str, any]] = None
+        layout_model: Optional[str] = None
     ) -> Union[UnifiedDocument, Dict]:
         """
         Main processing method with dual-track support.
@@ -1797,7 +1843,7 @@ class OCRService:
             output_dir: Optional output directory
             use_dual_track: Whether to use dual-track processing (default True)
             force_track: Force specific track ("ocr" or "direct")
-            pp_structure_params: Optional custom PP-StructureV3 parameters (used for OCR track only)
+            layout_model: Layout detection model ('chinese', 'default', 'cdla') (used for OCR track only)
 
         Returns:
             UnifiedDocument if dual-track is enabled and use_dual_track=True,
@@ -1809,12 +1855,12 @@ class OCRService:
         if (use_dual_track or force_track) and self.dual_track_enabled:
             # Use dual-track processing (or forced track)
             return self.process_with_dual_track(
-                file_path, lang, detect_layout, confidence_threshold, output_dir, force_track, pp_structure_params
+                file_path, lang, detect_layout, confidence_threshold, output_dir, force_track, layout_model
             )
         else:
             # Use traditional OCR processing (no force_track support)
             return self.process_file_traditional(
-                file_path, lang, detect_layout, confidence_threshold, output_dir, pp_structure_params
+                file_path, lang, detect_layout, confidence_threshold, output_dir, layout_model
             )
 
     def process_legacy(

backend/app/services/ocr_to_unified_converter.py

@@ -3,6 +3,9 @@ OCR to UnifiedDocument Converter
 
 Converts PP-StructureV3 OCR results to UnifiedDocument format, preserving
 all structure information and metadata.
+
+Includes gap filling support to supplement PP-StructureV3 output with raw OCR
+regions when significant content loss is detected.
 """
 
 import logging
@@ -16,10 +19,165 @@ from app.models.unified_document import (
     BoundingBox, StyleInfo, TableData, ElementType,
     ProcessingTrack, TableCell, Dimensions
 )
+from app.services.gap_filling_service import GapFillingService
 
 logger = logging.getLogger(__name__)
 
 
+def trim_empty_columns(table_dict: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Remove empty columns from a table dictionary.
+
+    A column is considered empty if ALL cells in that column have content that is
+    empty or whitespace-only (using .strip() to determine emptiness).
+
+    This function:
+    1. Identifies columns where every cell's content is empty/whitespace
+    2. Removes identified empty columns
+    3. Updates cols/columns value
+    4. Recalculates each cell's col index
+    5. Adjusts col_span when spans cross removed columns
+    6. Removes cells entirely when their complete span falls within removed columns
+    7. Preserves original bbox (no layout drift)
+
+    Args:
+        table_dict: Table dictionary with keys: rows, cols/columns, cells
+
+    Returns:
+        Cleaned table dictionary with empty columns removed
+    """
+    cells = table_dict.get('cells', [])
+    if not cells:
+        return table_dict
+
+    # Get original column count
+    original_cols = table_dict.get('cols', table_dict.get('columns', 0))
+    if original_cols == 0:
+        # Calculate from cells if not provided
+        max_col = 0
+        for cell in cells:
+            cell_col = cell.get('col', 0) if isinstance(cell, dict) else getattr(cell, 'col', 0)
+            cell_span = cell.get('col_span', 1) if isinstance(cell, dict) else getattr(cell, 'col_span', 1)
+            max_col = max(max_col, cell_col + cell_span)
+        original_cols = max_col
+
+    if original_cols == 0:
+        return table_dict
+
+    # Build a map: column_index -> list of cell contents
+    # For cells with col_span > 1, we only check their primary column
+    column_contents: Dict[int, List[str]] = {i: [] for i in range(original_cols)}
+
+    for cell in cells:
+        if isinstance(cell, dict):
+            col = cell.get('col', 0)
+            col_span = cell.get('col_span', 1)
+            content = cell.get('content', '')
+        else:
+            col = getattr(cell, 'col', 0)
+            col_span = getattr(cell, 'col_span', 1)
+            content = getattr(cell, 'content', '')
+
+        # Mark content for each column this cell spans
+        for c in range(col, min(col + col_span, original_cols)):
+            if c in column_contents:
+                column_contents[c].append(str(content).strip() if content else '')
+
+    # Identify empty columns (all content is empty/whitespace)
+    empty_columns = set()
+    for col_idx, contents in column_contents.items():
+        # A column is empty if ALL cells in it have empty content
+        # Note: If a column has no cells at all, it's considered empty
+        if all(c == '' for c in contents):
+            empty_columns.add(col_idx)
+
+    if not empty_columns:
+        # No empty columns to remove, just ensure cols is set
+        result = dict(table_dict)
+        if result.get('cols', result.get('columns', 0)) == 0:
+            result['cols'] = original_cols
+            if 'columns' in result:
+                result['columns'] = original_cols
+        return result
+
+    logger.debug(f"Removing empty columns: {sorted(empty_columns)} from table with {original_cols} cols")
+
+    # Build column mapping: old_col -> new_col (or None if removed)
+    col_mapping: Dict[int, Optional[int]] = {}
+    new_col = 0
+    for old_col in range(original_cols):
+        if old_col in empty_columns:
+            col_mapping[old_col] = None
+        else:
+            col_mapping[old_col] = new_col
+            new_col += 1
+
+    new_cols = new_col
+
+    # Process cells
+    new_cells = []
+    for cell in cells:
+        if isinstance(cell, dict):
+            old_col = cell.get('col', 0)
+            old_col_span = cell.get('col_span', 1)
+        else:
+            old_col = getattr(cell, 'col', 0)
+            old_col_span = getattr(cell, 'col_span', 1)
+
+        # Calculate new col and col_span
+        # Find the first non-removed column in this cell's span
+        new_start_col = None
+        new_end_col = None
+
+        for c in range(old_col, min(old_col + old_col_span, original_cols)):
+            mapped = col_mapping.get(c)
+            if mapped is not None:
+                if new_start_col is None:
+                    new_start_col = mapped
+                new_end_col = mapped
+
+        # If entire span falls within removed columns, skip this cell
+        if new_start_col is None:
+            logger.debug(f"Removing cell at row={cell.get('row', 0) if isinstance(cell, dict) else cell.row}, "
+                        f"col={old_col} (entire span in removed columns)")
+            continue
+
+        new_col_span = new_end_col - new_start_col + 1
+
+        # Create new cell
+        if isinstance(cell, dict):
+            new_cell = dict(cell)
+            new_cell['col'] = new_start_col
+            new_cell['col_span'] = new_col_span
+        else:
+            # Handle TableCell objects
+            new_cell = {
+                'row': cell.row,
+                'col': new_start_col,
+                'row_span': cell.row_span,
+                'col_span': new_col_span,
+                'content': cell.content
+            }
+            if hasattr(cell, 'bbox') and cell.bbox:
+                new_cell['bbox'] = cell.bbox
+            if hasattr(cell, 'style') and cell.style:
+                new_cell['style'] = cell.style
+
+        new_cells.append(new_cell)
+
+    # Build result
+    result = dict(table_dict)
+    result['cells'] = new_cells
+    result['cols'] = new_cols
+    if 'columns' in result:
+        result['columns'] = new_cols
+
+    logger.info(f"Trimmed table: {original_cols} -> {new_cols} columns, "
+               f"{len(cells)} -> {len(new_cells)} cells")
+
+    return result
+
+
 class OCRToUnifiedConverter:
     """
     Converter for transforming PP-StructureV3 OCR results to UnifiedDocument format.
@@ -30,11 +188,19 @@ class OCRToUnifiedConverter:
     - Multi-page document assembly
     - Metadata preservation
     - Structure relationship mapping
+    - Gap filling with raw OCR regions (when PP-StructureV3 misses content)
     """
 
-    def __init__(self):
-        """Initialize the converter."""
+    def __init__(self, enable_gap_filling: bool = True):
+        """
+        Initialize the converter.
+
+        Args:
+            enable_gap_filling: Whether to enable gap filling with raw OCR regions
+        """
         self.element_counter = 0
+        self.gap_filling_service = GapFillingService() if enable_gap_filling else None
+        self.gap_filling_stats: Dict[str, Any] = {}
 
     def convert(
         self,
@@ -120,13 +286,21 @@ class OCRToUnifiedConverter:
         Extract pages from OCR results.
 
         Handles both enhanced PP-StructureV3 results (with parsing_res_list)
-        and traditional markdown results.
+        and traditional markdown results. Applies gap filling when enabled.
         """
         pages = []
 
+        # Extract raw OCR text regions for gap filling
+        raw_text_regions = ocr_results.get('text_regions', [])
+        ocr_dimensions = ocr_results.get('ocr_dimensions', {})
+
         # Check if we have enhanced results from PPStructureEnhanced
         if 'enhanced_results' in ocr_results:
-            pages = self._extract_from_enhanced_results(ocr_results['enhanced_results'])
+            pages = self._extract_from_enhanced_results(
+                ocr_results['enhanced_results'],
+                raw_text_regions=raw_text_regions,
+                ocr_dimensions=ocr_dimensions
+            )
         # Check for traditional OCR results with text_regions at top level (from process_file_traditional)
         elif 'text_regions' in ocr_results:
             pages = self._extract_from_traditional_ocr(ocr_results)
@@ -143,9 +317,21 @@ class OCRToUnifiedConverter:
 
     def _extract_from_enhanced_results(
         self,
-        enhanced_results: List[Dict[str, Any]]
+        enhanced_results: List[Dict[str, Any]],
+        raw_text_regions: Optional[List[Dict[str, Any]]] = None,
+        ocr_dimensions: Optional[Dict[str, Any]] = None
     ) -> List[Page]:
-        """Extract pages from enhanced PP-StructureV3 results."""
+        """
+        Extract pages from enhanced PP-StructureV3 results.
+
+        Applies gap filling when enabled to supplement PP-StructureV3 output
+        with raw OCR regions that were not detected by the layout model.
+
+        Args:
+            enhanced_results: PP-StructureV3 enhanced results
+            raw_text_regions: Raw OCR text regions for gap filling
+            ocr_dimensions: OCR image dimensions for coordinate alignment
+        """
         pages = []
 
         for page_idx, page_result in enumerate(enhanced_results):
@@ -158,15 +344,52 @@ class OCRToUnifiedConverter:
                     if element:
                         elements.append(element)
 
+            # Get page dimensions
+            pp_dimensions = Dimensions(
+                width=page_result.get('width', 0),
+                height=page_result.get('height', 0)
+            )
+
+            # Apply gap filling if enabled and raw regions available
+            if self.gap_filling_service and raw_text_regions:
+                # Filter raw regions for current page
+                page_raw_regions = [
+                    r for r in raw_text_regions
+                    if r.get('page', 0) == page_idx or r.get('page', 1) == page_idx + 1
+                ]
+
+                if page_raw_regions:
+                    supplemented, stats = self.gap_filling_service.fill_gaps(
+                        raw_ocr_regions=page_raw_regions,
+                        pp_structure_elements=elements,
+                        page_number=page_idx + 1,
+                        ocr_dimensions=ocr_dimensions,
+                        pp_dimensions=pp_dimensions
+                    )
+
+                    # Store statistics
+                    self.gap_filling_stats[f'page_{page_idx + 1}'] = stats
+
+                    if supplemented:
+                        logger.info(
+                            f"Page {page_idx + 1}: Gap filling added {len(supplemented)} elements "
+                            f"(coverage: {stats.get('coverage_ratio', 0):.2%})"
+                        )
+                        elements.extend(supplemented)
+
+                        # Recalculate reading order for combined elements
+                        reading_order = self.gap_filling_service.recalculate_reading_order(elements)
+                        page_result['reading_order'] = reading_order
+
             # Create page
             page = Page(
                 page_number=page_idx + 1,
-                dimensions=Dimensions(
-                    width=page_result.get('width', 0),
-                    height=page_result.get('height', 0)
-                ),
+                dimensions=pp_dimensions,
                 elements=elements,
-                metadata={'reading_order': page_result.get('reading_order', [])}
+                metadata={
+                    'reading_order': page_result.get('reading_order', []),
+                    'gap_filling': self.gap_filling_stats.get(f'page_{page_idx + 1}', {})
+                }
             )
 
             pages.append(page)
@@ -500,6 +723,9 @@ class OCRToUnifiedConverter:
     ) -> Optional[DocumentElement]:
         """Convert table data to DocumentElement."""
         try:
+            # Clean up empty columns before building TableData
+            table_dict = trim_empty_columns(table_dict)
+
             # Extract bbox
             bbox_data = table_dict.get('bbox', [0, 0, 0, 0])
             bbox = BoundingBox(
@@ -587,14 +813,22 @@ class OCRToUnifiedConverter:
                 cells = []
                 headers = []
                 rows = table.find_all('tr')
+                num_rows = len(rows)
 
-                # Track actual column positions accounting for rowspan/colspan
-                # This is a simplified approach - complex spanning may need enhancement
+                # First pass: calculate total columns by finding max column extent
+                # Track cells that span multiple rows: occupied[row][col] = True
+                occupied: Dict[int, Dict[int, bool]] = {r: {} for r in range(num_rows)}
+
+                # Parse all cells with proper rowspan/colspan handling
                 for row_idx, row in enumerate(rows):
                     row_cells = row.find_all(['td', 'th'])
                     col_idx = 0
 
                     for cell in row_cells:
+                        # Skip columns that are occupied by rowspan from previous rows
+                        while occupied[row_idx].get(col_idx, False):
+                            col_idx += 1
+
                         cell_content = cell.get_text(strip=True)
                         rowspan = int(cell.get('rowspan', 1))
                         colspan = int(cell.get('colspan', 1))
@@ -611,26 +845,66 @@ class OCRToUnifiedConverter:
                         if cell.name == 'th' or row_idx == 0:
                             headers.append(cell_content)
 
+                        # Mark cells as occupied for rowspan/colspan
+                        for r in range(row_idx, min(row_idx + rowspan, num_rows)):
+                            for c in range(col_idx, col_idx + colspan):
+                                if r not in occupied:
+                                    occupied[r] = {}
+                                occupied[r][c] = True
+
                         # Advance column index by colspan
                         col_idx += colspan
 
-                # Calculate actual dimensions
-                num_rows = len(rows)
-                num_cols = max(
-                    sum(int(cell.get('colspan', 1)) for cell in row.find_all(['td', 'th']))
-                    for row in rows
-                ) if rows else 0
+                # Calculate actual column count from occupied cells
+                num_cols = 0
+                for r in range(num_rows):
+                    if occupied[r]:
+                        max_col_in_row = max(occupied[r].keys()) + 1
+                        num_cols = max(num_cols, max_col_in_row)
 
                 logger.debug(
                     f"Parsed HTML table: {num_rows} rows, {num_cols} cols, {len(cells)} cells"
                 )
 
+                # Build table dict for cleanup
+                table_dict = {
+                    'rows': num_rows,
+                    'cols': num_cols,
+                    'cells': [
+                        {
+                            'row': c.row,
+                            'col': c.col,
+                            'row_span': c.row_span,
+                            'col_span': c.col_span,
+                            'content': c.content
+                        }
+                        for c in cells
+                    ],
+                    'headers': headers if headers else None,
+                    'caption': extracted_text if extracted_text else None
+                }
+
+                # Clean up empty columns
+                table_dict = trim_empty_columns(table_dict)
+
+                # Convert cleaned cells back to TableCell objects
+                cleaned_cells = [
+                    TableCell(
+                        row=c['row'],
+                        col=c['col'],
+                        row_span=c.get('row_span', 1),
+                        col_span=c.get('col_span', 1),
+                        content=c.get('content', '')
+                    )
+                    for c in table_dict.get('cells', [])
+                ]
+
                 return TableData(
-                    rows=num_rows,
-                    cols=num_cols,
-                    cells=cells,
-                    headers=headers if headers else None,
-                    caption=extracted_text if extracted_text else None
+                    rows=table_dict.get('rows', num_rows),
+                    cols=table_dict.get('cols', num_cols),
+                    cells=cleaned_cells,
+                    headers=table_dict.get('headers'),
+                    caption=table_dict.get('caption')
                 )
 
             except ImportError:

backend/app/services/pp_structure_debug.py

@@ -0,0 +1,344 @@
+"""
+PP-StructureV3 Debug Service
+
+Provides debugging tools for visualizing and saving PP-StructureV3 results:
+- Save raw results as JSON for inspection
+- Generate visualization images showing detected bboxes
+- Compare raw OCR regions with PP-StructureV3 elements
+"""
+
+import json
+import logging
+from pathlib import Path
+from typing import Dict, List, Any, Optional, Tuple
+from datetime import datetime
+
+from PIL import Image, ImageDraw, ImageFont
+
+logger = logging.getLogger(__name__)
+
+# Color palette for different element types (RGB)
+ELEMENT_COLORS: Dict[str, Tuple[int, int, int]] = {
+    'text': (0, 128, 0),       # Green
+    'title': (0, 0, 255),      # Blue
+    'table': (255, 0, 0),      # Red
+    'figure': (255, 165, 0),   # Orange
+    'image': (255, 165, 0),    # Orange
+    'header': (128, 0, 128),   # Purple
+    'footer': (128, 0, 128),   # Purple
+    'equation': (0, 255, 255), # Cyan
+    'chart': (255, 192, 203),  # Pink
+    'list': (139, 69, 19),     # Brown
+    'reference': (128, 128, 128),  # Gray
+    'default': (255, 0, 255),  # Magenta for unknown types
+}
+
+# Color for raw OCR regions
+RAW_OCR_COLOR = (255, 215, 0)  # Gold
+
+
+class PPStructureDebug:
+    """Debug service for PP-StructureV3 analysis results."""
+
+    def __init__(self, output_dir: Path):
+        """
+        Initialize debug service.
+
+        Args:
+            output_dir: Directory to save debug outputs
+        """
+        self.output_dir = Path(output_dir)
+        self.output_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_raw_results(
+        self,
+        pp_structure_results: Dict[str, Any],
+        raw_ocr_regions: List[Dict[str, Any]],
+        filename_prefix: str = "debug"
+    ) -> Dict[str, Path]:
+        """
+        Save raw PP-StructureV3 results and OCR regions as JSON files.
+
+        Args:
+            pp_structure_results: Raw PP-StructureV3 analysis results
+            raw_ocr_regions: Raw OCR text regions
+            filename_prefix: Prefix for output files
+
+        Returns:
+            Dictionary with paths to saved files
+        """
+        saved_files = {}
+
+        # Save PP-StructureV3 results
+        pp_json_path = self.output_dir / f"{filename_prefix}_pp_structure_raw.json"
+        try:
+            # Convert any non-serializable types
+            serializable_results = self._make_serializable(pp_structure_results)
+            with open(pp_json_path, 'w', encoding='utf-8') as f:
+                json.dump(serializable_results, f, ensure_ascii=False, indent=2)
+            saved_files['pp_structure'] = pp_json_path
+            logger.info(f"Saved PP-StructureV3 raw results to {pp_json_path}")
+        except Exception as e:
+            logger.error(f"Failed to save PP-StructureV3 results: {e}")
+
+        # Save raw OCR regions
+        ocr_json_path = self.output_dir / f"{filename_prefix}_raw_ocr_regions.json"
+        try:
+            serializable_ocr = self._make_serializable(raw_ocr_regions)
+            with open(ocr_json_path, 'w', encoding='utf-8') as f:
+                json.dump(serializable_ocr, f, ensure_ascii=False, indent=2)
+            saved_files['raw_ocr'] = ocr_json_path
+            logger.info(f"Saved raw OCR regions to {ocr_json_path}")
+        except Exception as e:
+            logger.error(f"Failed to save raw OCR regions: {e}")
+
+        # Save summary comparison
+        summary_path = self.output_dir / f"{filename_prefix}_debug_summary.json"
+        try:
+            summary = self._generate_summary(pp_structure_results, raw_ocr_regions)
+            with open(summary_path, 'w', encoding='utf-8') as f:
+                json.dump(summary, f, ensure_ascii=False, indent=2)
+            saved_files['summary'] = summary_path
+            logger.info(f"Saved debug summary to {summary_path}")
+        except Exception as e:
+            logger.error(f"Failed to save debug summary: {e}")
+
+        return saved_files
+
+    def generate_visualization(
+        self,
+        image_path: Path,
+        pp_structure_elements: List[Dict[str, Any]],
+        raw_ocr_regions: Optional[List[Dict[str, Any]]] = None,
+        filename_prefix: str = "debug",
+        show_labels: bool = True,
+        show_raw_ocr: bool = True
+    ) -> Optional[Path]:
+        """
+        Generate visualization image showing detected elements.
+
+        Args:
+            image_path: Path to original image
+            pp_structure_elements: PP-StructureV3 detected elements
+            raw_ocr_regions: Optional raw OCR regions to overlay
+            filename_prefix: Prefix for output file
+            show_labels: Whether to show element type labels
+            show_raw_ocr: Whether to show raw OCR regions
+
+        Returns:
+            Path to generated visualization image
+        """
+        try:
+            # Load original image
+            img = Image.open(image_path)
+            if img.mode != 'RGB':
+                img = img.convert('RGB')
+
+            # Create copy for drawing
+            viz_img = img.copy()
+            draw = ImageDraw.Draw(viz_img)
+
+            # Try to load a font, fall back to default
+            try:
+                font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", 14)
+                small_font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", 10)
+            except (IOError, OSError):
+                try:
+                    font = ImageFont.truetype("/home/egg/project/Tool_OCR/backend/fonts/NotoSansSC-Regular.ttf", 14)
+                    small_font = ImageFont.truetype("/home/egg/project/Tool_OCR/backend/fonts/NotoSansSC-Regular.ttf", 10)
+                except (IOError, OSError):
+                    font = ImageFont.load_default()
+                    small_font = font
+
+            # Draw raw OCR regions first (so PP-Structure boxes are on top)
+            if show_raw_ocr and raw_ocr_regions:
+                for idx, region in enumerate(raw_ocr_regions):
+                    bbox = self._normalize_bbox(region.get('bbox', []))
+                    if bbox:
+                        # Draw with dashed style simulation (draw thin lines)
+                        x0, y0, x1, y1 = bbox
+                        draw.rectangle([x0, y0, x1, y1], outline=RAW_OCR_COLOR, width=1)
+
+                        # Add small label
+                        if show_labels:
+                            confidence = region.get('confidence', 0)
+                            label = f"OCR:{confidence:.2f}"
+                            draw.text((x0, y0 - 12), label, fill=RAW_OCR_COLOR, font=small_font)
+
+            # Draw PP-StructureV3 elements
+            for idx, elem in enumerate(pp_structure_elements):
+                elem_type = elem.get('type', 'default')
+                if hasattr(elem_type, 'value'):
+                    elem_type = elem_type.value
+                elem_type = str(elem_type).lower()
+
+                color = ELEMENT_COLORS.get(elem_type, ELEMENT_COLORS['default'])
+                bbox = self._normalize_bbox(elem.get('bbox', []))
+
+                if bbox:
+                    x0, y0, x1, y1 = bbox
+                    # Draw thicker rectangle for PP-Structure elements
+                    draw.rectangle([x0, y0, x1, y1], outline=color, width=3)
+
+                    # Add label
+                    if show_labels:
+                        label = f"{idx}:{elem_type}"
+                        # Draw label background
+                        text_bbox = draw.textbbox((x0, y0 - 18), label, font=font)
+                        draw.rectangle(text_bbox, fill=(255, 255, 255, 200))
+                        draw.text((x0, y0 - 18), label, fill=color, font=font)
+
+            # Add legend
+            self._draw_legend(draw, img.width, font)
+
+            # Add image info
+            info_text = f"PP-Structure: {len(pp_structure_elements)} elements"
+            if raw_ocr_regions:
+                info_text += f" | Raw OCR: {len(raw_ocr_regions)} regions"
+            info_text += f" | Size: {img.width}x{img.height}"
+            draw.text((10, img.height - 25), info_text, fill=(0, 0, 0), font=font)
+
+            # Save visualization
+            viz_path = self.output_dir / f"{filename_prefix}_pp_structure_viz.png"
+            viz_img.save(viz_path, 'PNG')
+            logger.info(f"Saved visualization to {viz_path}")
+
+            return viz_path
+
+        except Exception as e:
+            logger.error(f"Failed to generate visualization: {e}")
+            import traceback
+            traceback.print_exc()
+            return None
+
+    def _draw_legend(self, draw: ImageDraw, img_width: int, font: ImageFont):
+        """Draw a legend showing element type colors."""
+        legend_x = img_width - 150
+        legend_y = 10
+
+        # Draw legend background
+        draw.rectangle(
+            [legend_x - 5, legend_y - 5, img_width - 5, legend_y + len(ELEMENT_COLORS) * 18 + 25],
+            fill=(255, 255, 255, 230),
+            outline=(0, 0, 0)
+        )
+
+        draw.text((legend_x, legend_y), "Legend:", fill=(0, 0, 0), font=font)
+        legend_y += 20
+
+        for elem_type, color in ELEMENT_COLORS.items():
+            if elem_type == 'default':
+                continue
+            draw.rectangle([legend_x, legend_y + 2, legend_x + 12, legend_y + 14], fill=color)
+            draw.text((legend_x + 18, legend_y), elem_type, fill=(0, 0, 0), font=font)
+            legend_y += 18
+
+        # Add raw OCR legend entry
+        draw.rectangle([legend_x, legend_y + 2, legend_x + 12, legend_y + 14], fill=RAW_OCR_COLOR)
+        draw.text((legend_x + 18, legend_y), "raw_ocr", fill=(0, 0, 0), font=font)
+
+    def _normalize_bbox(self, bbox: Any) -> Optional[Tuple[float, float, float, float]]:
+        """Normalize bbox to (x0, y0, x1, y1) format."""
+        if not bbox:
+            return None
+
+        try:
+            # Handle nested list format [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
+            if isinstance(bbox, (list, tuple)) and len(bbox) >= 1:
+                if isinstance(bbox[0], (list, tuple)):
+                    xs = [pt[0] for pt in bbox if len(pt) >= 2]
+                    ys = [pt[1] for pt in bbox if len(pt) >= 2]
+                    if xs and ys:
+                        return (min(xs), min(ys), max(xs), max(ys))
+
+            # Handle flat list [x0, y0, x1, y1]
+            if isinstance(bbox, (list, tuple)) and len(bbox) == 4:
+                return (float(bbox[0]), float(bbox[1]), float(bbox[2]), float(bbox[3]))
+
+            # Handle flat polygon [x1, y1, x2, y2, ...]
+            if isinstance(bbox, (list, tuple)) and len(bbox) >= 8:
+                xs = [bbox[i] for i in range(0, len(bbox), 2)]
+                ys = [bbox[i] for i in range(1, len(bbox), 2)]
+                return (min(xs), min(ys), max(xs), max(ys))
+
+            # Handle dict format
+            if isinstance(bbox, dict):
+                return (
+                    float(bbox.get('x0', bbox.get('x_min', 0))),
+                    float(bbox.get('y0', bbox.get('y_min', 0))),
+                    float(bbox.get('x1', bbox.get('x_max', 0))),
+                    float(bbox.get('y1', bbox.get('y_max', 0)))
+                )
+
+        except (TypeError, ValueError, IndexError) as e:
+            logger.warning(f"Failed to normalize bbox {bbox}: {e}")
+
+        return None
+
+    def _generate_summary(
+        self,
+        pp_structure_results: Dict[str, Any],
+        raw_ocr_regions: List[Dict[str, Any]]
+    ) -> Dict[str, Any]:
+        """Generate summary comparing PP-Structure and raw OCR."""
+        pp_elements = pp_structure_results.get('elements', [])
+
+        # Count element types
+        type_counts = {}
+        for elem in pp_elements:
+            elem_type = elem.get('type', 'unknown')
+            if hasattr(elem_type, 'value'):
+                elem_type = elem_type.value
+            type_counts[str(elem_type)] = type_counts.get(str(elem_type), 0) + 1
+
+        # Calculate bounding box coverage
+        pp_bbox_area = 0
+        ocr_bbox_area = 0
+
+        for elem in pp_elements:
+            bbox = self._normalize_bbox(elem.get('bbox'))
+            if bbox:
+                pp_bbox_area += (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
+
+        for region in raw_ocr_regions:
+            bbox = self._normalize_bbox(region.get('bbox'))
+            if bbox:
+                ocr_bbox_area += (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
+
+        return {
+            'timestamp': datetime.now().isoformat(),
+            'pp_structure': {
+                'total_elements': len(pp_elements),
+                'element_types': type_counts,
+                'total_bbox_area': pp_bbox_area,
+                'has_parsing_res_list': pp_structure_results.get('has_parsing_res_list', False)
+            },
+            'raw_ocr': {
+                'total_regions': len(raw_ocr_regions),
+                'total_bbox_area': ocr_bbox_area,
+                'avg_confidence': sum(r.get('confidence', 0) for r in raw_ocr_regions) / len(raw_ocr_regions) if raw_ocr_regions else 0
+            },
+            'comparison': {
+                'element_count_ratio': len(pp_elements) / len(raw_ocr_regions) if raw_ocr_regions else 0,
+                'area_ratio': pp_bbox_area / ocr_bbox_area if ocr_bbox_area > 0 else 0,
+                'potential_gap': len(raw_ocr_regions) - len(pp_elements) if raw_ocr_regions else 0
+            }
+        }
+
+    def _make_serializable(self, obj: Any) -> Any:
+        """Convert object to JSON-serializable format."""
+        if obj is None:
+            return None
+        if isinstance(obj, (str, int, float, bool)):
+            return obj
+        if isinstance(obj, (list, tuple)):
+            return [self._make_serializable(item) for item in obj]
+        if isinstance(obj, dict):
+            return {str(k): self._make_serializable(v) for k, v in obj.items()}
+        if hasattr(obj, 'value'):
+            return obj.value
+        if hasattr(obj, '__dict__'):
+            return self._make_serializable(obj.__dict__)
+        if hasattr(obj, 'tolist'):  # numpy array
+            return obj.tolist()
+        return str(obj)

backend/tests/api/test_layout_model_api.py

@@ -0,0 +1,332 @@
+"""
+API integration tests for Layout Model Selection feature.
+
+This replaces the deprecated PP-StructureV3 parameter tests.
+"""
+
+import pytest
+from fastapi.testclient import TestClient
+from unittest.mock import patch
+from app.main import app
+from app.core.database import get_db
+from app.models.user import User
+from app.models.task import Task, TaskStatus, TaskFile
+
+
+@pytest.fixture
+def client():
+    """Create test client"""
+    return TestClient(app)
+
+
+@pytest.fixture
+def test_user(db_session):
+    """Create test user"""
+    user = User(
+        email="test@example.com",
+        hashed_password="test_hash",
+        is_active=True
+    )
+    db_session.add(user)
+    db_session.commit()
+    db_session.refresh(user)
+    return user
+
+
+@pytest.fixture
+def test_task(db_session, test_user):
+    """Create test task with uploaded file"""
+    task = Task(
+        user_id=test_user.id,
+        task_id="test-task-123",
+        filename="test.pdf",
+        status=TaskStatus.PENDING
+    )
+    db_session.add(task)
+    db_session.commit()
+    db_session.refresh(task)
+
+    # Add task file
+    task_file = TaskFile(
+        task_id=task.id,
+        original_name="test.pdf",
+        stored_path="/tmp/test.pdf",
+        file_size=1024,
+        mime_type="application/pdf"
+    )
+    db_session.add(task_file)
+    db_session.commit()
+
+    return task
+
+
+class TestLayoutModelSchema:
+    """Test LayoutModel and ProcessingOptions schema validation"""
+
+    def test_processing_options_accepts_layout_model(self):
+        """Verify ProcessingOptions schema accepts layout_model parameter"""
+        from app.schemas.task import ProcessingOptions, LayoutModelEnum
+
+        options = ProcessingOptions(
+            use_dual_track=True,
+            language='ch',
+            layout_model=LayoutModelEnum.CHINESE
+        )
+
+        assert options.layout_model == LayoutModelEnum.CHINESE
+
+    def test_layout_model_enum_values(self):
+        """Verify all layout model enum values are valid"""
+        from app.schemas.task import LayoutModelEnum
+
+        assert LayoutModelEnum.CHINESE.value == "chinese"
+        assert LayoutModelEnum.DEFAULT.value == "default"
+        assert LayoutModelEnum.CDLA.value == "cdla"
+
+    def test_default_layout_model_is_chinese(self):
+        """Verify default layout model is 'chinese' for best Chinese document support"""
+        from app.schemas.task import ProcessingOptions
+
+        options = ProcessingOptions()
+
+        # Default should be chinese
+        assert options.layout_model.value == "chinese"
+
+    def test_layout_model_string_values_accepted(self):
+        """Verify string values are accepted for layout_model"""
+        from app.schemas.task import ProcessingOptions
+
+        # String values should be converted to enum
+        options = ProcessingOptions(layout_model="default")
+        assert options.layout_model.value == "default"
+
+        options = ProcessingOptions(layout_model="cdla")
+        assert options.layout_model.value == "cdla"
+
+    def test_invalid_layout_model_rejected(self):
+        """Verify invalid layout model values are rejected"""
+        from app.schemas.task import ProcessingOptions
+        from pydantic import ValidationError
+
+        with pytest.raises(ValidationError):
+            ProcessingOptions(layout_model="invalid_model")
+
+
+class TestStartTaskEndpoint:
+    """Test /tasks/{task_id}/start endpoint with layout_model parameter"""
+
+    @patch('app.routers.tasks.process_task_ocr')
+    def test_start_task_with_layout_model(self, mock_process_ocr, client, test_task, db_session):
+        """Verify layout_model is accepted and passed to OCR service"""
+
+        # Override get_db dependency
+        def override_get_db():
+            try:
+                yield db_session
+            finally:
+                pass
+
+        # Override auth dependency
+        def override_get_current_user():
+            return test_task.user
+
+        app.dependency_overrides[get_db] = override_get_db
+        from app.core.deps import get_current_user
+        app.dependency_overrides[get_current_user] = override_get_current_user
+
+        # Request body with layout_model
+        request_body = {
+            "use_dual_track": True,
+            "language": "ch",
+            "layout_model": "chinese"
+        }
+
+        # Make API call
+        response = client.post(
+            f"/api/v2/tasks/{test_task.task_id}/start",
+            json=request_body
+        )
+
+        # Verify response
+        assert response.status_code == 200
+        data = response.json()
+        assert data['status'] == 'processing'
+
+        # Verify background task was called with layout_model
+        mock_process_ocr.assert_called_once()
+        call_kwargs = mock_process_ocr.call_args[1]
+
+        assert 'layout_model' in call_kwargs
+        assert call_kwargs['layout_model'] == 'chinese'
+
+        # Clean up
+        app.dependency_overrides.clear()
+
+    @patch('app.routers.tasks.process_task_ocr')
+    def test_start_task_with_default_model(self, mock_process_ocr, client, test_task, db_session):
+        """Verify 'default' layout model is accepted"""
+
+        def override_get_db():
+            try:
+                yield db_session
+            finally:
+                pass
+
+        def override_get_current_user():
+            return test_task.user
+
+        app.dependency_overrides[get_db] = override_get_db
+        from app.core.deps import get_current_user
+        app.dependency_overrides[get_current_user] = override_get_current_user
+
+        request_body = {
+            "use_dual_track": True,
+            "layout_model": "default"
+        }
+
+        response = client.post(
+            f"/api/v2/tasks/{test_task.task_id}/start",
+            json=request_body
+        )
+
+        assert response.status_code == 200
+
+        mock_process_ocr.assert_called_once()
+        call_kwargs = mock_process_ocr.call_args[1]
+        assert call_kwargs['layout_model'] == 'default'
+
+        app.dependency_overrides.clear()
+
+    @patch('app.routers.tasks.process_task_ocr')
+    def test_start_task_with_cdla_model(self, mock_process_ocr, client, test_task, db_session):
+        """Verify 'cdla' layout model is accepted"""
+
+        def override_get_db():
+            try:
+                yield db_session
+            finally:
+                pass
+
+        def override_get_current_user():
+            return test_task.user
+
+        app.dependency_overrides[get_db] = override_get_db
+        from app.core.deps import get_current_user
+        app.dependency_overrides[get_current_user] = override_get_current_user
+
+        request_body = {
+            "use_dual_track": True,
+            "layout_model": "cdla"
+        }
+
+        response = client.post(
+            f"/api/v2/tasks/{test_task.task_id}/start",
+            json=request_body
+        )
+
+        assert response.status_code == 200
+
+        mock_process_ocr.assert_called_once()
+        call_kwargs = mock_process_ocr.call_args[1]
+        assert call_kwargs['layout_model'] == 'cdla'
+
+        app.dependency_overrides.clear()
+
+    @patch('app.routers.tasks.process_task_ocr')
+    def test_start_task_without_layout_model_uses_default(self, mock_process_ocr, client, test_task, db_session):
+        """Verify task can start without layout_model (uses 'chinese' as default)"""
+
+        def override_get_db():
+            try:
+                yield db_session
+            finally:
+                pass
+
+        def override_get_current_user():
+            return test_task.user
+
+        app.dependency_overrides[get_db] = override_get_db
+        from app.core.deps import get_current_user
+        app.dependency_overrides[get_current_user] = override_get_current_user
+
+        # Request without layout_model
+        request_body = {
+            "use_dual_track": True,
+            "language": "ch"
+        }
+
+        response = client.post(
+            f"/api/v2/tasks/{test_task.task_id}/start",
+            json=request_body
+        )
+
+        assert response.status_code == 200
+
+        mock_process_ocr.assert_called_once()
+        call_kwargs = mock_process_ocr.call_args[1]
+
+        # layout_model should default to 'chinese'
+        assert call_kwargs['layout_model'] == 'chinese'
+
+        app.dependency_overrides.clear()
+
+    def test_start_task_with_invalid_layout_model(self, client, test_task, db_session):
+        """Verify invalid layout_model returns 422 validation error"""
+
+        def override_get_db():
+            try:
+                yield db_session
+            finally:
+                pass
+
+        def override_get_current_user():
+            return test_task.user
+
+        app.dependency_overrides[get_db] = override_get_db
+        from app.core.deps import get_current_user
+        app.dependency_overrides[get_current_user] = override_get_current_user
+
+        # Request with invalid layout_model
+        request_body = {
+            "use_dual_track": True,
+            "layout_model": "invalid_model"
+        }
+
+        response = client.post(
+            f"/api/v2/tasks/{test_task.task_id}/start",
+            json=request_body
+        )
+
+        # Should return validation error
+        assert response.status_code == 422
+
+        app.dependency_overrides.clear()
+
+
+class TestOpenAPISchema:
+    """Test OpenAPI schema includes layout_model parameter"""
+
+    def test_openapi_schema_includes_layout_model(self, client):
+        """Verify OpenAPI schema documents layout_model parameter"""
+        response = client.get("/openapi.json")
+        assert response.status_code == 200
+
+        schema = response.json()
+
+        # Check LayoutModelEnum schema exists
+        assert 'LayoutModelEnum' in schema['components']['schemas']
+
+        model_schema = schema['components']['schemas']['LayoutModelEnum']
+
+        # Verify all 3 model options are documented
+        assert 'chinese' in model_schema['enum']
+        assert 'default' in model_schema['enum']
+        assert 'cdla' in model_schema['enum']
+
+        # Verify ProcessingOptions includes layout_model
+        options_schema = schema['components']['schemas']['ProcessingOptions']
+        assert 'layout_model' in options_schema['properties']
+
+
+if __name__ == '__main__':
+    pytest.main([__file__, '-v'])

backend/tests/services/test_layout_model.py

@@ -0,0 +1,244 @@
+"""
+Unit tests for Layout Model Selection feature in OCR Service.
+
+This replaces the deprecated PP-StructureV3 parameter tests.
+"""
+
+import pytest
+import sys
+from pathlib import Path
+from unittest.mock import Mock, patch, MagicMock
+
+# Mock all external dependencies before importing OCRService
+sys.modules['paddleocr'] = MagicMock()
+sys.modules['PIL'] = MagicMock()
+sys.modules['pdf2image'] = MagicMock()
+
+# Mock paddle with version attribute
+paddle_mock = MagicMock()
+paddle_mock.__version__ = '2.5.0'
+paddle_mock.device.get_device.return_value = 'cpu'
+paddle_mock.device.get_available_device.return_value = 'cpu'
+sys.modules['paddle'] = paddle_mock
+
+# Mock torch
+torch_mock = MagicMock()
+torch_mock.cuda.is_available.return_value = False
+sys.modules['torch'] = torch_mock
+
+from app.services.ocr_service import OCRService, LAYOUT_MODEL_MAPPING, _USE_PUBLAYNET_DEFAULT
+from app.core.config import settings
+
+
+class TestLayoutModelMapping:
+    """Test layout model name mapping"""
+
+    def test_layout_model_mapping_exists(self):
+        """Verify LAYOUT_MODEL_MAPPING constant exists and has correct values"""
+        assert 'chinese' in LAYOUT_MODEL_MAPPING
+        assert 'default' in LAYOUT_MODEL_MAPPING
+        assert 'cdla' in LAYOUT_MODEL_MAPPING
+
+    def test_chinese_model_maps_to_pp_doclayout(self):
+        """Verify 'chinese' maps to PP-DocLayout-S"""
+        assert LAYOUT_MODEL_MAPPING['chinese'] == 'PP-DocLayout-S'
+
+    def test_default_model_maps_to_publaynet_sentinel(self):
+        """Verify 'default' maps to sentinel value for PubLayNet default"""
+        # The 'default' model uses a sentinel value that signals "use PubLayNet default (no custom model)"
+        assert LAYOUT_MODEL_MAPPING['default'] == _USE_PUBLAYNET_DEFAULT
+
+    def test_cdla_model_maps_to_picodet(self):
+        """Verify 'cdla' maps to picodet_lcnet_x1_0_fgd_layout_cdla"""
+        assert LAYOUT_MODEL_MAPPING['cdla'] == 'picodet_lcnet_x1_0_fgd_layout_cdla'
+
+
+class TestLayoutModelEngine:
+    """Test engine creation with different layout models"""
+
+    def test_chinese_model_creates_engine_with_pp_doclayout(self):
+        """Verify 'chinese' layout model uses PP-DocLayout-S"""
+        ocr_service = OCRService()
+
+        with patch.object(ocr_service, 'structure_engine', None):
+            with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+                mock_engine = Mock()
+                mock_ppstructure.return_value = mock_engine
+
+                engine = ocr_service._ensure_structure_engine(layout_model='chinese')
+
+                mock_ppstructure.assert_called_once()
+                call_kwargs = mock_ppstructure.call_args[1]
+
+                assert call_kwargs.get('layout_detection_model_name') == 'PP-DocLayout-S'
+
+    def test_default_model_creates_engine_without_model_name(self):
+        """Verify 'default' layout model does not specify model name (uses default)"""
+        ocr_service = OCRService()
+
+        with patch.object(ocr_service, 'structure_engine', None):
+            with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+                mock_engine = Mock()
+                mock_ppstructure.return_value = mock_engine
+
+                engine = ocr_service._ensure_structure_engine(layout_model='default')
+
+                mock_ppstructure.assert_called_once()
+                call_kwargs = mock_ppstructure.call_args[1]
+
+                # For 'default', layout_detection_model_name should be None or not set
+                assert call_kwargs.get('layout_detection_model_name') is None
+
+    def test_cdla_model_creates_engine_with_picodet(self):
+        """Verify 'cdla' layout model uses picodet_lcnet_x1_0_fgd_layout_cdla"""
+        ocr_service = OCRService()
+
+        with patch.object(ocr_service, 'structure_engine', None):
+            with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+                mock_engine = Mock()
+                mock_ppstructure.return_value = mock_engine
+
+                engine = ocr_service._ensure_structure_engine(layout_model='cdla')
+
+                mock_ppstructure.assert_called_once()
+                call_kwargs = mock_ppstructure.call_args[1]
+
+                assert call_kwargs.get('layout_detection_model_name') == 'picodet_lcnet_x1_0_fgd_layout_cdla'
+
+    def test_none_layout_model_uses_chinese_default(self):
+        """Verify None layout_model defaults to 'chinese' model"""
+        ocr_service = OCRService()
+
+        with patch.object(ocr_service, 'structure_engine', None):
+            with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+                mock_engine = Mock()
+                mock_ppstructure.return_value = mock_engine
+
+                # Pass None for layout_model
+                engine = ocr_service._ensure_structure_engine(layout_model=None)
+
+                mock_ppstructure.assert_called_once()
+                call_kwargs = mock_ppstructure.call_args[1]
+
+                # Should use 'chinese' model as default
+                assert call_kwargs.get('layout_detection_model_name') == 'PP-DocLayout-S'
+
+
+class TestLayoutModelCaching:
+    """Test engine caching behavior with layout models"""
+
+    def test_same_layout_model_uses_cached_engine(self):
+        """Verify same layout model reuses cached engine"""
+        ocr_service = OCRService()
+
+        with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+            mock_engine = Mock()
+            mock_ppstructure.return_value = mock_engine
+
+            # First call with 'chinese'
+            engine1 = ocr_service._ensure_structure_engine(layout_model='chinese')
+
+            # Second call with same model should use cache
+            engine2 = ocr_service._ensure_structure_engine(layout_model='chinese')
+
+            # Verify only one engine was created
+            assert mock_ppstructure.call_count == 1
+            assert engine1 is engine2
+
+    def test_different_layout_model_creates_new_engine(self):
+        """Verify different layout model creates new engine"""
+        ocr_service = OCRService()
+
+        with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+            mock_engine1 = Mock()
+            mock_engine2 = Mock()
+            mock_ppstructure.side_effect = [mock_engine1, mock_engine2]
+
+            # First call with 'chinese'
+            engine1 = ocr_service._ensure_structure_engine(layout_model='chinese')
+
+            # Second call with 'cdla' should create new engine
+            engine2 = ocr_service._ensure_structure_engine(layout_model='cdla')
+
+            # Verify two engines were created
+            assert mock_ppstructure.call_count == 2
+            assert engine1 is not engine2
+
+
+class TestLayoutModelFlow:
+    """Test layout model parameter flow through processing pipeline"""
+
+    def test_layout_model_passed_to_engine_creation(self):
+        """Verify layout_model is passed through to _ensure_structure_engine"""
+        ocr_service = OCRService()
+
+        # Test that _ensure_structure_engine accepts layout_model parameter
+        with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+            mock_engine = Mock()
+            mock_ppstructure.return_value = mock_engine
+
+            # Call with specific layout_model
+            engine = ocr_service._ensure_structure_engine(layout_model='cdla')
+
+            # Verify correct model was requested
+            mock_ppstructure.assert_called_once()
+            call_kwargs = mock_ppstructure.call_args[1]
+            assert call_kwargs.get('layout_detection_model_name') == 'picodet_lcnet_x1_0_fgd_layout_cdla'
+
+    def test_layout_model_default_behavior(self):
+        """Verify default layout model behavior when None is passed"""
+        ocr_service = OCRService()
+
+        with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+            mock_engine = Mock()
+            mock_ppstructure.return_value = mock_engine
+
+            # Call without layout_model (None)
+            engine = ocr_service._ensure_structure_engine(layout_model=None)
+
+            # Should use config default (PP-DocLayout-S)
+            mock_ppstructure.assert_called_once()
+            call_kwargs = mock_ppstructure.call_args[1]
+            assert call_kwargs.get('layout_detection_model_name') == settings.layout_detection_model_name
+
+    def test_layout_model_unknown_value_falls_back(self):
+        """Verify unknown layout model falls back to config default"""
+        ocr_service = OCRService()
+
+        with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+            mock_engine = Mock()
+            mock_ppstructure.return_value = mock_engine
+
+            # Call with unknown layout_model
+            engine = ocr_service._ensure_structure_engine(layout_model='unknown_model')
+
+            # Should use config default
+            mock_ppstructure.assert_called_once()
+            call_kwargs = mock_ppstructure.call_args[1]
+            assert call_kwargs.get('layout_detection_model_name') == settings.layout_detection_model_name
+
+
+class TestLayoutModelLogging:
+    """Test layout model logging"""
+
+    def test_layout_model_is_logged(self):
+        """Verify layout model selection is logged"""
+        ocr_service = OCRService()
+
+        with patch('app.services.ocr_service.PPStructureV3') as mock_ppstructure:
+            with patch('app.services.ocr_service.logger') as mock_logger:
+                mock_engine = Mock()
+                mock_ppstructure.return_value = mock_engine
+
+                # Call with specific layout_model
+                ocr_service._ensure_structure_engine(layout_model='cdla')
+
+                # Verify logging occurred
+                assert mock_logger.info.call_count >= 1
+                # Check that model name was logged
+                log_calls = [str(call) for call in mock_logger.info.call_args_list]
+                assert any('cdla' in str(call).lower() or 'layout' in str(call).lower() for call in log_calls)
+
+
+if __name__ == '__main__':
+    pytest.main([__file__, '-v'])

backend/tests/test_gap_filling.py

@@ -0,0 +1,503 @@
+"""
+Tests for Gap Filling Service
+
+Tests the detection and filling of gaps in PP-StructureV3 output
+using raw OCR text regions.
+"""
+
+import pytest
+from typing import List, Dict, Any
+
+from app.services.gap_filling_service import GapFillingService, TextRegion, SKIP_ELEMENT_TYPES
+from app.models.unified_document import DocumentElement, BoundingBox, ElementType, Dimensions
+
+
+class TestGapFillingService:
+    """Tests for GapFillingService class."""
+
+    @pytest.fixture
+    def service(self) -> GapFillingService:
+        """Create a GapFillingService instance with default settings."""
+        return GapFillingService(
+            coverage_threshold=0.7,
+            iou_threshold=0.15,
+            confidence_threshold=0.3,
+            dedup_iou_threshold=0.5,
+            enabled=True
+        )
+
+    @pytest.fixture
+    def disabled_service(self) -> GapFillingService:
+        """Create a disabled GapFillingService instance."""
+        return GapFillingService(enabled=False)
+
+    @pytest.fixture
+    def sample_raw_regions(self) -> List[TextRegion]:
+        """Create sample raw OCR text regions."""
+        return [
+            TextRegion(text="Header text", bbox=[100, 50, 300, 80], confidence=0.95, page=1),
+            TextRegion(text="Title of document", bbox=[100, 100, 500, 150], confidence=0.92, page=1),
+            TextRegion(text="First paragraph", bbox=[100, 200, 500, 250], confidence=0.90, page=1),
+            TextRegion(text="Second paragraph", bbox=[100, 300, 500, 350], confidence=0.88, page=1),
+            TextRegion(text="Footer note", bbox=[100, 900, 300, 930], confidence=0.85, page=1),
+            # Low confidence region (should be filtered)
+            TextRegion(text="Noise", bbox=[50, 50, 80, 80], confidence=0.1, page=1),
+        ]
+
+    @pytest.fixture
+    def sample_pp_elements(self) -> List[DocumentElement]:
+        """Create sample PP-StructureV3 elements that cover only some regions."""
+        return [
+            DocumentElement(
+                element_id="pp_1",
+                type=ElementType.TITLE,
+                content="Title of document",
+                bbox=BoundingBox(x0=100, y0=100, x1=500, y1=150),
+                confidence=0.95
+            ),
+            DocumentElement(
+                element_id="pp_2",
+                type=ElementType.TEXT,
+                content="First paragraph",
+                bbox=BoundingBox(x0=100, y0=200, x1=500, y1=250),
+                confidence=0.90
+            ),
+            # Note: Header, Second paragraph, and Footer are NOT covered
+        ]
+
+    def test_service_initialization(self, service: GapFillingService):
+        """Test service initializes with correct parameters."""
+        assert service.enabled is True
+        assert service.coverage_threshold == 0.7
+        assert service.iou_threshold == 0.15
+        assert service.confidence_threshold == 0.3
+        assert service.dedup_iou_threshold == 0.5
+
+    def test_disabled_service(self, disabled_service: GapFillingService):
+        """Test disabled service does not activate."""
+        regions = [TextRegion(text="Test", bbox=[0, 0, 100, 100], confidence=0.9, page=1)]
+        elements = []
+
+        should_activate, coverage = disabled_service.should_activate(regions, elements)
+        assert should_activate is False
+        assert coverage == 1.0
+
+    def test_should_activate_low_coverage(
+        self,
+        service: GapFillingService,
+        sample_raw_regions: List[TextRegion],
+        sample_pp_elements: List[DocumentElement]
+    ):
+        """Test activation when coverage is below threshold."""
+        # Filter out low confidence regions
+        valid_regions = [r for r in sample_raw_regions if r.confidence >= 0.3]
+
+        should_activate, coverage = service.should_activate(valid_regions, sample_pp_elements)
+
+        # Only 2 out of 5 valid regions are covered (Title, First paragraph)
+        assert should_activate is True
+        assert coverage < 0.7  # Below threshold
+
+    def test_should_not_activate_high_coverage(self, service: GapFillingService):
+        """Test no activation when coverage is above threshold."""
+        # All regions covered
+        regions = [
+            TextRegion(text="Text 1", bbox=[100, 100, 200, 150], confidence=0.9, page=1),
+            TextRegion(text="Text 2", bbox=[100, 200, 200, 250], confidence=0.9, page=1),
+        ]
+
+        elements = [
+            DocumentElement(
+                element_id="pp_1",
+                type=ElementType.TEXT,
+                content="Text 1",
+                bbox=BoundingBox(x0=50, y0=50, x1=250, y1=200),  # Covers first region
+                confidence=0.95
+            ),
+            DocumentElement(
+                element_id="pp_2",
+                type=ElementType.TEXT,
+                content="Text 2",
+                bbox=BoundingBox(x0=50, y0=180, x1=250, y1=300),  # Covers second region
+                confidence=0.95
+            ),
+        ]
+
+        should_activate, coverage = service.should_activate(regions, elements)
+
+        assert should_activate is False
+        assert coverage >= 0.7
+
+    def test_find_uncovered_regions(
+        self,
+        service: GapFillingService,
+        sample_raw_regions: List[TextRegion],
+        sample_pp_elements: List[DocumentElement]
+    ):
+        """Test finding uncovered regions."""
+        uncovered = service.find_uncovered_regions(sample_raw_regions, sample_pp_elements)
+
+        # Should find Header, Second paragraph, Footer (not Title, First paragraph, or low-confidence Noise)
+        assert len(uncovered) == 3
+
+        uncovered_texts = [r.text for r in uncovered]
+        assert "Header text" in uncovered_texts
+        assert "Second paragraph" in uncovered_texts
+        assert "Footer note" in uncovered_texts
+        assert "Title of document" not in uncovered_texts  # Covered
+        assert "First paragraph" not in uncovered_texts  # Covered
+        assert "Noise" not in uncovered_texts  # Low confidence
+
+    def test_coverage_by_center_point(self, service: GapFillingService):
+        """Test coverage detection via center point."""
+        region = TextRegion(text="Test", bbox=[150, 150, 250, 200], confidence=0.9, page=1)
+
+        element = DocumentElement(
+            element_id="pp_1",
+            type=ElementType.TEXT,
+            content="Container",
+            bbox=BoundingBox(x0=100, y0=100, x1=300, y1=250),  # Contains region's center
+            confidence=0.95
+        )
+
+        is_covered = service._is_region_covered(region, [element])
+        assert is_covered is True
+
+    def test_coverage_by_iou(self, service: GapFillingService):
+        """Test coverage detection via IoU threshold."""
+        region = TextRegion(text="Test", bbox=[100, 100, 200, 150], confidence=0.9, page=1)
+
+        element = DocumentElement(
+            element_id="pp_1",
+            type=ElementType.TEXT,
+            content="Overlap",
+            bbox=BoundingBox(x0=150, y0=100, x1=250, y1=150),  # Partial overlap
+            confidence=0.95
+        )
+
+        # Calculate expected IoU
+        # Intersection: (150-200) x (100-150) = 50 x 50 = 2500
+        # Union: 100x50 + 100x50 - 2500 = 7500
+        # IoU = 2500/7500 = 0.33 > 0.15 threshold
+
+        is_covered = service._is_region_covered(region, [element])
+        assert is_covered is True
+
+    def test_deduplication(
+        self,
+        service: GapFillingService,
+        sample_pp_elements: List[DocumentElement]
+    ):
+        """Test deduplication removes high-overlap regions."""
+        uncovered = [
+            # High overlap with pp_2 (First paragraph)
+            TextRegion(text="First paragraph variant", bbox=[100, 200, 500, 250], confidence=0.9, page=1),
+            # No overlap
+            TextRegion(text="Unique region", bbox=[100, 500, 300, 550], confidence=0.9, page=1),
+        ]
+
+        deduplicated = service.deduplicate_regions(uncovered, sample_pp_elements)
+
+        assert len(deduplicated) == 1
+        assert deduplicated[0].text == "Unique region"
+
+    def test_convert_regions_to_elements(self, service: GapFillingService):
+        """Test conversion of TextRegions to DocumentElements."""
+        regions = [
+            TextRegion(text="Test text 1", bbox=[100, 100, 200, 150], confidence=0.85, page=1),
+            TextRegion(text="Test text 2", bbox=[100, 200, 200, 250], confidence=0.90, page=1),
+        ]
+
+        elements = service.convert_regions_to_elements(regions, page_number=1, start_element_id=0)
+
+        assert len(elements) == 2
+        assert elements[0].element_id == "gap_fill_1_0"
+        assert elements[0].type == ElementType.TEXT
+        assert elements[0].content == "Test text 1"
+        assert elements[0].confidence == 0.85
+        assert elements[0].metadata.get('source') == 'gap_filling'
+
+        assert elements[1].element_id == "gap_fill_1_1"
+        assert elements[1].content == "Test text 2"
+
+    def test_recalculate_reading_order(self, service: GapFillingService):
+        """Test reading order recalculation."""
+        elements = [
+            DocumentElement(
+                element_id="e3",
+                type=ElementType.TEXT,
+                content="Bottom",
+                bbox=BoundingBox(x0=100, y0=300, x1=200, y1=350),
+                confidence=0.9
+            ),
+            DocumentElement(
+                element_id="e1",
+                type=ElementType.TEXT,
+                content="Top",
+                bbox=BoundingBox(x0=100, y0=100, x1=200, y1=150),
+                confidence=0.9
+            ),
+            DocumentElement(
+                element_id="e2",
+                type=ElementType.TEXT,
+                content="Middle",
+                bbox=BoundingBox(x0=100, y0=200, x1=200, y1=250),
+                confidence=0.9
+            ),
+        ]
+
+        reading_order = service.recalculate_reading_order(elements)
+
+        # Should be sorted by y0: Top (100), Middle (200), Bottom (300)
+        assert reading_order == [1, 2, 0]  # Indices of elements in reading order
+
+    def test_fill_gaps_integration(
+        self,
+        service: GapFillingService,
+    ):
+        """Integration test for fill_gaps method."""
+        # Raw OCR regions (dict format as received from OCR service)
+        raw_regions = [
+            {'text': 'Header', 'bbox': [100, 50, 300, 80], 'confidence': 0.95, 'page': 1},
+            {'text': 'Title', 'bbox': [100, 100, 500, 150], 'confidence': 0.92, 'page': 1},
+            {'text': 'Paragraph 1', 'bbox': [100, 200, 500, 250], 'confidence': 0.90, 'page': 1},
+            {'text': 'Paragraph 2', 'bbox': [100, 300, 500, 350], 'confidence': 0.88, 'page': 1},
+            {'text': 'Paragraph 3', 'bbox': [100, 400, 500, 450], 'confidence': 0.86, 'page': 1},
+            {'text': 'Footer', 'bbox': [100, 900, 300, 930], 'confidence': 0.85, 'page': 1},
+        ]
+
+        # PP-StructureV3 only detected Title (missing 5 out of 6 regions = 16.7% coverage)
+        pp_elements = [
+            DocumentElement(
+                element_id="pp_1",
+                type=ElementType.TITLE,
+                content="Title",
+                bbox=BoundingBox(x0=100, y0=100, x1=500, y1=150),
+                confidence=0.95
+            ),
+        ]
+
+        supplemented, stats = service.fill_gaps(
+            raw_ocr_regions=raw_regions,
+            pp_structure_elements=pp_elements,
+            page_number=1
+        )
+
+        # Should have activated and supplemented missing regions
+        assert stats['activated'] is True
+        assert stats['coverage_ratio'] < 0.7
+        assert len(supplemented) == 5  # Header, Paragraph 1, 2, 3, Footer
+
+    def test_fill_gaps_no_activation_when_coverage_high(self, service: GapFillingService):
+        """Test fill_gaps does not activate when coverage is high."""
+        raw_regions = [
+            {'text': 'Text 1', 'bbox': [100, 100, 200, 150], 'confidence': 0.9, 'page': 1},
+        ]
+
+        pp_elements = [
+            DocumentElement(
+                element_id="pp_1",
+                type=ElementType.TEXT,
+                content="Text 1",
+                bbox=BoundingBox(x0=50, y0=50, x1=250, y1=200),  # Fully covers
+                confidence=0.95
+            ),
+        ]
+
+        supplemented, stats = service.fill_gaps(
+            raw_ocr_regions=raw_regions,
+            pp_structure_elements=pp_elements,
+            page_number=1
+        )
+
+        assert stats['activated'] is False
+        assert len(supplemented) == 0
+
+    def test_skip_element_types_not_supplemented(self, service: GapFillingService):
+        """Test that TABLE/IMAGE/etc. elements are not supplemented over."""
+        raw_regions = [
+            {'text': 'Table cell text', 'bbox': [100, 100, 200, 150], 'confidence': 0.9, 'page': 1},
+        ]
+
+        # PP-StructureV3 has a table covering this region
+        pp_elements = [
+            DocumentElement(
+                element_id="pp_1",
+                type=ElementType.TABLE,
+                content="<table>...</table>",
+                bbox=BoundingBox(x0=50, y0=50, x1=250, y1=200),
+                confidence=0.95
+            ),
+        ]
+
+        # The region should be considered covered by the table
+        supplemented, stats = service.fill_gaps(
+            raw_ocr_regions=raw_regions,
+            pp_structure_elements=pp_elements,
+            page_number=1
+        )
+
+        # Should not supplement because the table covers it
+        assert len(supplemented) == 0
+
+    def test_coordinate_scaling(self, service: GapFillingService):
+        """Test coordinate alignment with different dimensions."""
+        # OCR was done at 2000x3000, PP-Structure at 1000x1500
+        ocr_dimensions = {'width': 2000, 'height': 3000}
+        pp_dimensions = Dimensions(width=1000, height=1500)
+
+        raw_regions = [
+            # At OCR scale: (200, 300) to (400, 450) -> at PP scale: (100, 150) to (200, 225)
+            {'text': 'Scaled text', 'bbox': [200, 300, 400, 450], 'confidence': 0.9, 'page': 1},
+        ]
+
+        pp_elements = [
+            DocumentElement(
+                element_id="pp_1",
+                type=ElementType.TEXT,
+                content="Scaled text",
+                bbox=BoundingBox(x0=100, y0=150, x1=200, y1=225),  # Should cover after scaling
+                confidence=0.95
+            ),
+        ]
+
+        supplemented, stats = service.fill_gaps(
+            raw_ocr_regions=raw_regions,
+            pp_structure_elements=pp_elements,
+            page_number=1,
+            ocr_dimensions=ocr_dimensions,
+            pp_dimensions=pp_dimensions
+        )
+
+        # After scaling, the region should be covered
+        assert stats['coverage_ratio'] >= 0.7 or len(supplemented) == 0
+
+    def test_iou_calculation(self, service: GapFillingService):
+        """Test IoU calculation accuracy."""
+        # Two identical boxes
+        bbox1 = (0, 0, 100, 100)
+        bbox2 = (0, 0, 100, 100)
+        assert service._calculate_iou(bbox1, bbox2) == 1.0
+
+        # No overlap
+        bbox1 = (0, 0, 100, 100)
+        bbox2 = (200, 200, 300, 300)
+        assert service._calculate_iou(bbox1, bbox2) == 0.0
+
+        # 50% overlap
+        bbox1 = (0, 0, 100, 100)
+        bbox2 = (50, 0, 150, 100)  # Shifted right by 50
+        # Intersection: 50x100 = 5000
+        # Union: 10000 + 10000 - 5000 = 15000
+        # IoU = 5000/15000 = 0.333...
+        iou = service._calculate_iou(bbox1, bbox2)
+        assert abs(iou - 1/3) < 0.01
+
+    def test_point_in_bbox(self, service: GapFillingService):
+        """Test point-in-bbox check."""
+        bbox = (100, 100, 200, 200)
+
+        # Inside
+        assert service._point_in_bbox(150, 150, bbox) is True
+
+        # On edge
+        assert service._point_in_bbox(100, 100, bbox) is True
+        assert service._point_in_bbox(200, 200, bbox) is True
+
+        # Outside
+        assert service._point_in_bbox(50, 150, bbox) is False
+        assert service._point_in_bbox(250, 150, bbox) is False
+
+    def test_merge_adjacent_regions(self, service: GapFillingService):
+        """Test merging of adjacent text regions."""
+        regions = [
+            TextRegion(text="Hello", bbox=[100, 100, 150, 130], confidence=0.9, page=1),
+            TextRegion(text="World", bbox=[160, 100, 210, 130], confidence=0.85, page=1),  # Adjacent
+            TextRegion(text="Far away", bbox=[100, 300, 200, 330], confidence=0.9, page=1),  # Not adjacent
+        ]
+
+        merged = service.merge_adjacent_regions(regions, max_horizontal_gap=20, max_vertical_gap=10)
+
+        assert len(merged) == 2
+        # First two should be merged
+        assert "Hello" in merged[0].text and "World" in merged[0].text
+        assert merged[1].text == "Far away"
+
+
+class TestTextRegion:
+    """Tests for TextRegion dataclass."""
+
+    def test_normalized_bbox_4_values(self):
+        """Test bbox normalization with 4 values."""
+        region = TextRegion(text="Test", bbox=[100, 200, 300, 400], confidence=0.9, page=1)
+        assert region.normalized_bbox == (100, 200, 300, 400)
+
+    def test_normalized_bbox_polygon_flat(self):
+        """Test bbox normalization with flat polygon format (8 values)."""
+        # Polygon: 4 points as flat list [x1, y1, x2, y2, x3, y3, x4, y4]
+        region = TextRegion(
+            text="Test",
+            bbox=[100, 200, 300, 200, 300, 400, 100, 400],
+            confidence=0.9,
+            page=1
+        )
+        assert region.normalized_bbox == (100, 200, 300, 400)
+
+    def test_normalized_bbox_polygon_nested(self):
+        """Test bbox normalization with nested polygon format (PaddleOCR format)."""
+        # PaddleOCR format: [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
+        region = TextRegion(
+            text="Test",
+            bbox=[[100, 200], [300, 200], [300, 400], [100, 400]],
+            confidence=0.9,
+            page=1
+        )
+        assert region.normalized_bbox == (100, 200, 300, 400)
+
+    def test_normalized_bbox_numpy_polygon(self):
+        """Test bbox normalization with numpy-like nested format."""
+        # Sometimes PaddleOCR returns numpy arrays converted to lists
+        region = TextRegion(
+            text="Test",
+            bbox=[[100.5, 200.5], [300.5, 200.5], [300.5, 400.5], [100.5, 400.5]],
+            confidence=0.9,
+            page=1
+        )
+        bbox = region.normalized_bbox
+        assert bbox[0] == 100.5
+        assert bbox[1] == 200.5
+        assert bbox[2] == 300.5
+        assert bbox[3] == 400.5
+
+    def test_center_calculation(self):
+        """Test center point calculation."""
+        region = TextRegion(text="Test", bbox=[100, 200, 300, 400], confidence=0.9, page=1)
+        assert region.center == (200, 300)
+
+    def test_center_calculation_nested_bbox(self):
+        """Test center point calculation with nested bbox format."""
+        region = TextRegion(
+            text="Test",
+            bbox=[[100, 200], [300, 200], [300, 400], [100, 400]],
+            confidence=0.9,
+            page=1
+        )
+        assert region.center == (200, 300)
+
+
+class TestOCRToUnifiedConverterIntegration:
+    """Integration tests for OCRToUnifiedConverter with gap filling."""
+
+    def test_converter_with_gap_filling_enabled(self):
+        """Test converter initializes with gap filling enabled."""
+        from app.services.ocr_to_unified_converter import OCRToUnifiedConverter
+
+        converter = OCRToUnifiedConverter(enable_gap_filling=True)
+        assert converter.gap_filling_service is not None
+
+    def test_converter_with_gap_filling_disabled(self):
+        """Test converter initializes without gap filling."""
+        from app.services.ocr_to_unified_converter import OCRToUnifiedConverter
+
+        converter = OCRToUnifiedConverter(enable_gap_filling=False)
+        assert converter.gap_filling_service is None

docs/commit-history-report.md

@@ -0,0 +1,31 @@
+# Tool_OCR Commit History Review (2025-11-12 ~ 2025-11-26)
+
+本報告依 `git log` 全量 97 筆提交整理，涵蓋開發脈絡、里程碑、測試/品質信號與後續風險。提交類型統計：35 `feat` / 37 `fix` / 9 `chore` / 5 `test` / 4 `docs` / 2 `refactor`，主要集中於 2025-11-18、11-19、11-20 與 11-24 的密集開發。
+
+## 時間軸與里程碑
+- **前期基礎與前端現代化 (11-12~11-13)**：`21bc2f9`, `57cf912` 將前端改為 Tailwind v4 +專業 UI，`0f81d5e` 單容器 Docker 化、`d7e6473` WSL Ubuntu 開發環境。
+- **GPU 加速與相容性 (11-14)**：`6452797` 提案 + `7536f43` 實作 GPU OCR，`d80d60f`/`3694411`/`80c091b` 修正 Paddle 3.x API 與安裝來源，`b048f2d` 暫停圖表識別以避免 API 缺口。
+- **外部 Auth V2 與管理後台 (11-14~11-16)**：`28e419f`~`fd98018` 完成外部認證 V2、資料表前綴與架構移除 V1；`8f94191` 新增後台/稽核/Token 檢查；`90fca50`/`6bb5b76` 讓 18/18 測試全過。
+- **V2 UI 串接與初版版面保持 PDF (11-16~11-18)**：前端/後端全面切換 V2 API (`ad5c8be` 之後)，`fa1abcd` 版面保持 PDF + 多次座標/重疊修正 (`d33f605`~`0edc56b`)，強化 logging (`d99d37d`)。
+- **雙軌處理架構 (11-18~11-20)**：`2d50c12` + `82139c8` 導入 OCR/Direct 雙軌與 UnifiedDocument；`a3a6fbe`/`ab89a40`/`ecdce96` 完成轉換、JSON 匯出與 PDF 支援；`1d0b638` 後端 API，`c2288ba` 前端支援，`c50a5e9` 單元/整合測試；`0974fc3` E2E 修復，`ef335cf` Office 直抽，`b997f93`/`9f449e8` GPU 記憶體管理與文件化，`2ecd022` E2E 測試完成。
+- **PDF 版面復原計畫 (11-20 提案，11-24 實作高峰)**：`cf894b0` 提案後，`0aff468` Phase1 圖片/表格修復，`3fc32bc` Phase2 風格保存，`77fe4cc`/`ad879d4`/`75c194f` 等完成 Alignment、List、Span 級渲染與多欄位；一系列 `93bd9f5`~`3358d97` 針對位置/重疊/缺圖修正，`4325d02` 專案清理並封存提案。
+- **PP-Structure V3 調校 (11-25)**：`a659e7a` 改善複雜圖示結構保留，`2312b4c` 前端可調 `pp_structure` 參數 + 測試，`0999898` 多頁 PDF 座標校正。
+- **記憶體管理與混合抽圖 (11-25~11-26)**：`ba8ddf2` 提案，`1afdb82` 混合圖片抽取+記憶體管理落地，`b997f93` 系列 GPU 釋放/可選 torch，引入 ModelManager、ServicePool、MemoryGuard（詳見 `openspec/changes/archive/2025-11-26-enhance-memory-management`）；`a227311` 封存提案但僅完成 75/80 任務（剩餘文件化）；隨後多筆修復（`79cffe6`~`fa9b542`）處理 PDF 回歸與文字渲染，`6e050eb` 為最新 OCR 軌表格格式/裁剪修正。
+
+## 品質與測試信號
+- 11-16 完成 V2 API 測試 18/18 (`6bb5b76`)，建立初步信心。
+- 雙軌導入時新增單元/整合/E2E 測試 (`0fcb249`, `c50a5e9`, `2ecd022`)，但後續 PDF 版面復原大量依賴人工驗證，Phase 4 測試仍未完成（見下）。
+- 記憶體管理變更伴隨 57+18+10 測試檔（任務 8.1 完成），但文件化缺失可能影響交接與調參。
+- 11-24 大量 PDF 修復連續提交顯示迭代式修 bug，建議增加回歸測試覆蓋（特別是表格/多欄/列表與跨軌道 PDF）。
+
+## 未盡事項與風險
+- **記憶體管理文件化缺口**：`openspec/changes/archive/2025-11-26-enhance-memory-management/tasks.md` 未完成 Section 8.2（架構說明、調校指南、疑難排解、監控、遷移指南），可能影響部署可操作性。
+- **PDF 版面復原驗證不足**：同一變更的 Phase 4 測試/效能/文件與多類文件驗證均未勾選，現階段品質依賴手測。
+- **近期修正集中於 PDF 與表格**（`79cffe6`, `5c561f4`, `19bd5fd`, `fa9b542`, `6e050eb`），顯示 Direct/OCR 軌 PDF 路徑仍脆弱；缺乏自動化回歸易再度回歸。
+- **主分支狀態**：`main` 比 `origin/main` 超前 1 提交（`6e050eb`），請推送前確認 CI/測試。
+
+## 建議後續行動
+1) 完成記憶體管理文件（架構、調參、故障排除、Prometheus 監控指南）並加入 sanity check。  
+2) 為 PDF 版面復原建立最小回歸集：多欄文檔、含圖表/表格的 Direct/OCR 軌、列表與 span 混排。  
+3) 圍繞 `processing_track` 分流與 UnifiedDocument/PDF 生成的邊界條件增加測試（LOGO/未知元素、跨頁表格、OCR/Direct 混合圖片）。  
+4) 推送前跑現有單元/整合/E2E 測試，補上近兩週新增場景的腳本以降低回歸風險。

docs/project-risk-assessment.md

@@ -0,0 +1,24 @@
+# Project Risk & Issue Outlook
+
+本文件整理當前專案的可預見問題、潛在問題與建議修復方向（依風險與可行性排序）。依據來源：`git log`（97 commits, 2025-11-12~11-26）、`docs/architecture-overview.md`、`openspec/changes/archive/2025-11-26-enhance-memory-management/tasks.md` 等。
+
+## 可預見的問題項目
+- **記憶體管理文件缺口**：`openspec/changes/archive/2025-11-26-enhance-memory-management/tasks.md` 的 8.2 文檔未完成，ModelManager/ServicePool/MemoryGuard 的調參與故障處置缺乏 runbook，部署或擴容時易踩坑。方向：補完架構說明、調參指南、故障排解與監控落地範例（Prometheus 指標與警戒值）。
+- **PDF 生成回歸風險高**：版面保持與表格/圖片渲染在 `fa1abcd` 之後多次修正（例如 `d33f605`→`92e326b`、`108784a`→`3358d97`、`6e050eb`），顯示缺少自動回歸。方向：建立最小回歸集（多欄文本、含圖表/表格、列表/Span 混排）與 golden PDF/JSON 比對，覆蓋 Direct/OCR 雙軌。
+- **最新 OCR 表格格式修復未經回歸**：`6e050eb` 修正 OCR 軌表格資料格式與裁剪，無對應測試。方向：為 OCR 軌加表格解析/PDF 出圖的整合測試，確保與前端下載/展示一致。
+- **PP-Structure 參數調校可能影響資源**：`frontend` 支援前端可調 `pp_structure_params`（`2312b4c`），若缺乏 guard，可能放大 GPU/記憶體壓力。方向：在後端對超參做白名單與上限檢查，並納入 MemoryGuard 預估。
+- **Chart 能力啟停策略缺少驗證**：`b048f2d` 禁用 → `7e12f16` 重新啟用；缺少覆蓋率與性能數據。方向：為 chart 模型啟用/關閉建立健康檢查與 A/B 測試數據收集。
+
+## 潛在的問題項目
+- **UnifiedDocument 結構漂移風險**：雙軌共用輸出，近期多次調整（列表、Span、多欄、LOGO 元素），缺少結構驗證或 schema 鎖定。可能導致前端/匯出器/PDF 生成不一致。方向：定義 JSON Schema 或 pydantic 驗證，建立 contract 測試。
+- **服務池與記憶體守護的長時間行為未驗證**：雖有單元/整合測試，缺乏長時間 soak/stress（GPU 記憶碎片、模型 unload/reload、信號處理）。方向：加入 24h soak 測試與記憶體走勢告警，驗證 SIGTERM/SIGINT 清理。
+- **LibreOffice 轉檔鏈低觀測性**：Office 直抽與轉 PDF (`ef335cf`) 依賴系統 LibreOffice，缺少失敗監控與重試策略。方向：為轉檔階段增加 metrics/告警，並提供 fallback/重試。
+- **前端/後端 API 契約缺少檢查**：多次 V1→V2 遷移與新增參數（`pp_structure_params` 等），目前僅靠 E2E，缺少型別/契約檢查。方向：加入 OpenAPI 契約測試或生成型別校驗（ts-sdk 對齊 FastAPI schema）。
+- **混合抽圖/圖片保存路徑邊界**：Direct/OCR 混合抽圖與 `_save_image` 實作曾多次修復，仍缺少對 None/缺檔路徑的防禦。方向：為缺檔/無圖的 PDF 生成加強斷言與 fallback。
+
+## 建議修復與方向
+1) **完成記憶體管理文檔與樣板設定**：在 `docs/` 新增 MemoryGuard/ServicePool 調參與故障排除指南，附 `.env` 範例與 Prometheus 規則，對應 tasks 8.2 清單。
+2) **建立 PDF/UnifiedDocument 回歸套件**：收集代表性樣本（多欄、表格、列表、含圖/LOGO、OCR/Direct 雙軌），產生 golden JSON/PDF，加入 CI 比對，並為 `6e050eb` 相關表格路徑新增測試。
+3) **加入 UnifiedDocument Schema 驗證**：定義 schema（pydantic/JSON Schema），在匯出/PDF 生成前驗證；同時讓前端型別由 OpenAPI 生成以防 drift。
+4) **PP-Structure 參數防護與資源估算**：後端實作白名單/上限與 MemoryGuard 預估，避免前端自由調參造成 GPU OOM；增加拒絕/降級回饋。
+5) **長時間穩定性與轉檔可觀測性**：增加 soak/stress pipeline，追蹤 GPU/CPU/記憶碎片；為 LibreOffice/轉檔階段加 metrics、重試與錯誤分類告警。

frontend/src/components/LayoutModelSelector.tsx

@@ -0,0 +1,110 @@
+import { cn } from '@/lib/utils'
+import { Check, FileText, Globe, BookOpen } from 'lucide-react'
+import { useTranslation } from 'react-i18next'
+import type { LayoutModel } from '@/types/apiV2'
+
+interface LayoutModelSelectorProps {
+  value: LayoutModel
+  onChange: (model: LayoutModel) => void
+  disabled?: boolean
+  className?: string
+}
+
+const MODEL_ICONS: Record<LayoutModel, React.ReactNode> = {
+  chinese: <FileText className="w-5 h-5" />,
+  default: <Globe className="w-5 h-5" />,
+  cdla: <BookOpen className="w-5 h-5" />,
+}
+
+export default function LayoutModelSelector({
+  value,
+  onChange,
+  disabled = false,
+  className,
+}: LayoutModelSelectorProps) {
+  const { t } = useTranslation()
+  const models: LayoutModel[] = ['chinese', 'default', 'cdla']
+
+  const getModelInfo = (model: LayoutModel) => ({
+    label: t(`processing.layoutModel.${model}`),
+    description: t(`processing.layoutModel.${model}Desc`),
+  })
+
+  return (
+    <div className={cn('border rounded-lg p-4 bg-white', className)}>
+      {/* Header */}
+      <div className="flex items-center gap-2 mb-4">
+        <FileText className="w-5 h-5 text-gray-600" />
+        <h3 className="text-lg font-semibold text-gray-900">{t('processing.layoutModel.title')}</h3>
+      </div>
+
+      {/* Model Options */}
+      <div className="space-y-3">
+        {models.map((model) => {
+          const info = getModelInfo(model)
+          const isSelected = value === model
+
+          return (
+            <button
+              key={model}
+              type="button"
+              disabled={disabled}
+              onClick={() => onChange(model)}
+              className={cn(
+                'w-full flex items-start gap-4 p-4 rounded-lg border-2 transition-all text-left',
+                isSelected
+                  ? 'border-blue-500 bg-blue-50'
+                  : 'border-gray-200 hover:border-gray-300 hover:bg-gray-50',
+                disabled && 'opacity-50 cursor-not-allowed'
+              )}
+            >
+              {/* Icon */}
+              <div
+                className={cn(
+                  'p-2 rounded-lg flex-shrink-0',
+                  isSelected ? 'bg-blue-100 text-blue-600' : 'bg-gray-100 text-gray-500'
+                )}
+              >
+                {MODEL_ICONS[model]}
+              </div>
+
+              {/* Content */}
+              <div className="flex-1 min-w-0">
+                <div className="flex items-center gap-2">
+                  <span
+                    className={cn(
+                      'font-medium',
+                      isSelected ? 'text-blue-700' : 'text-gray-900'
+                    )}
+                  >
+                    {info.label}
+                  </span>
+                  {model === 'chinese' && (
+                    <span className="text-xs bg-green-100 text-green-700 px-2 py-0.5 rounded-full">
+                      {t('processing.layoutModel.recommended')}
+                    </span>
+                  )}
+                </div>
+                <p className="text-sm text-gray-500 mt-1">{info.description}</p>
+              </div>
+
+              {/* Check mark */}
+              {isSelected && (
+                <div className="flex-shrink-0">
+                  <Check className="w-5 h-5 text-blue-600" />
+                </div>
+              )}
+            </button>
+          )
+        })}
+      </div>
+
+      {/* Info Note */}
+      <div className="mt-4 p-3 bg-blue-50 border border-blue-200 rounded-md">
+        <p className="text-sm text-blue-800">
+          {t('processing.layoutModel.note')}
+        </p>
+      </div>
+    </div>
+  )
+}

frontend/src/components/PPStructureParams.tsx

@@ -1,408 +0,0 @@
-import { useState, useEffect } from 'react'
-import { Settings, RotateCcw, HelpCircle, Save, Upload, Download, Check, AlertCircle } from 'lucide-react'
-import { cn } from '@/lib/utils'
-import type { PPStructureV3Params } from '@/types/apiV2'
-
-const STORAGE_KEY = 'pp_structure_params_presets'
-const LAST_USED_KEY = 'pp_structure_params_last_used'
-
-interface PPStructureParamsProps {
-  value: PPStructureV3Params
-  onChange: (params: PPStructureV3Params) => void
-  disabled?: boolean
-  className?: string
-}
-
-interface ParamConfig {
-  key: keyof PPStructureV3Params
-  label: string
-  description: string
-  min: number
-  max: number
-  step: number
-  default: number
-  type: 'slider'
-}
-
-interface SelectParamConfig {
-  key: keyof PPStructureV3Params
-  label: string
-  description: string
-  options: Array<{ value: string; label: string }>
-  default: string
-  type: 'select'
-}
-
-// Preset configurations
-const PRESETS = {
-  default: {} as PPStructureV3Params,
-  'high-quality': {
-    layout_detection_threshold: 0.1,
-    layout_nms_threshold: 0.15,
-    text_det_thresh: 0.1,
-    text_det_box_thresh: 0.2,
-    layout_merge_bboxes_mode: 'small' as const,
-  } as PPStructureV3Params,
-  fast: {
-    layout_detection_threshold: 0.3,
-    layout_nms_threshold: 0.3,
-    text_det_thresh: 0.3,
-    text_det_box_thresh: 0.4,
-    layout_merge_bboxes_mode: 'large' as const,
-  } as PPStructureV3Params,
-}
-
-const PARAM_CONFIGS: Array<ParamConfig | SelectParamConfig> = [
-  {
-    key: 'layout_detection_threshold',
-    label: 'Layout Detection Threshold',
-    description: 'Lower = detect more blocks (including weak signals), Higher = only high-confidence blocks',
-    min: 0,
-    max: 1,
-    step: 0.05,
-    default: 0.2,
-    type: 'slider' as const,
-  },
-  {
-    key: 'layout_nms_threshold',
-    label: 'Layout NMS Threshold',
-    description: 'Lower = aggressive overlap removal, Higher = allow more overlapping boxes',
-    min: 0,
-    max: 1,
-    step: 0.05,
-    default: 0.2,
-    type: 'slider' as const,
-  },
-  {
-    key: 'layout_merge_bboxes_mode',
-    label: 'Layout Merge Mode',
-    description: 'Bounding box merging strategy',
-    options: [
-      { value: 'small', label: 'Small (Conservative)' },
-      { value: 'union', label: 'Union (Balanced)' },
-      { value: 'large', label: 'Large (Aggressive)' },
-    ],
-    default: 'small',
-    type: 'select' as const,
-  },
-  {
-    key: 'layout_unclip_ratio',
-    label: 'Layout Unclip Ratio',
-    description: 'Larger = looser bounding boxes, Smaller = tighter bounding boxes',
-    min: 0.5,
-    max: 3.0,
-    step: 0.1,
-    default: 1.2,
-    type: 'slider' as const,
-  },
-  {
-    key: 'text_det_thresh',
-    label: 'Text Detection Threshold',
-    description: 'Lower = detect more small/low-contrast text, Higher = cleaner but may miss text',
-    min: 0,
-    max: 1,
-    step: 0.05,
-    default: 0.2,
-    type: 'slider' as const,
-  },
-  {
-    key: 'text_det_box_thresh',
-    label: 'Text Box Threshold',
-    description: 'Lower = more text boxes retained, Higher = fewer false positives',
-    min: 0,
-    max: 1,
-    step: 0.05,
-    default: 0.3,
-    type: 'slider' as const,
-  },
-  {
-    key: 'text_det_unclip_ratio',
-    label: 'Text Unclip Ratio',
-    description: 'Larger = looser text boxes, Smaller = tighter text boxes',
-    min: 0.5,
-    max: 3.0,
-    step: 0.1,
-    default: 1.2,
-    type: 'slider' as const,
-  },
-]
-
-export default function PPStructureParams({
-  value,
-  onChange,
-  disabled = false,
-  className,
-}: PPStructureParamsProps) {
-  const [showTooltip, setShowTooltip] = useState<string | null>(null)
-  const [isExpanded, setIsExpanded] = useState(false)
-  const [selectedPreset, setSelectedPreset] = useState<string>('custom')
-  const [showSaveSuccess, setShowSaveSuccess] = useState(false)
-
-  // Load last used parameters on mount
-  useEffect(() => {
-    try {
-      const lastUsed = localStorage.getItem(LAST_USED_KEY)
-      if (lastUsed && Object.keys(value).length === 0) {
-        const params = JSON.parse(lastUsed)
-        onChange(params)
-      }
-    } catch (error) {
-      console.error('Failed to load last used parameters:', error)
-    }
-  }, [])
-
-  // Save to localStorage when parameters change
-  useEffect(() => {
-    if (Object.keys(value).length > 0) {
-      try {
-        localStorage.setItem(LAST_USED_KEY, JSON.stringify(value))
-      } catch (error) {
-        console.error('Failed to save parameters:', error)
-      }
-    }
-  }, [value])
-
-  const handleReset = () => {
-    onChange({})
-    setSelectedPreset('default')
-    setShowSaveSuccess(false)
-  }
-
-  const handlePresetChange = (presetKey: string) => {
-    setSelectedPreset(presetKey)
-    if (presetKey === 'custom') return
-
-    const preset = PRESETS[presetKey as keyof typeof PRESETS]
-    if (preset) {
-      onChange(preset)
-      setShowSaveSuccess(false)
-    }
-  }
-
-  const handleChange = (key: keyof PPStructureV3Params, newValue: any) => {
-    const newParams = {
-      ...value,
-      [key]: newValue,
-    }
-    onChange(newParams)
-    setSelectedPreset('custom')
-  }
-
-  const handleExport = () => {
-    const dataStr = JSON.stringify(value, null, 2)
-    const dataUri = 'data:application/json;charset=utf-8,' + encodeURIComponent(dataStr)
-    const exportFileDefaultName = 'pp_structure_params.json'
-
-    const linkElement = document.createElement('a')
-    linkElement.setAttribute('href', dataUri)
-    linkElement.setAttribute('download', exportFileDefaultName)
-    linkElement.click()
-  }
-
-  const handleImport = () => {
-    const input = document.createElement('input')
-    input.type = 'file'
-    input.accept = 'application/json'
-    input.onchange = (e) => {
-      const file = (e.target as HTMLInputElement).files?.[0]
-      if (file) {
-        const reader = new FileReader()
-        reader.onload = (event) => {
-          try {
-            const params = JSON.parse(event.target?.result as string)
-            onChange(params)
-            setSelectedPreset('custom')
-            setShowSaveSuccess(true)
-            setTimeout(() => setShowSaveSuccess(false), 3000)
-          } catch (error) {
-            console.error('Failed to import parameters:', error)
-          }
-        }
-        reader.readAsText(file)
-      }
-    }
-    input.click()
-  }
-
-  const hasCustomValues = Object.keys(value).length > 0
-
-  return (
-    <div className={cn('border rounded-lg p-4 bg-white', className)}>
-      {/* Header */}
-      <div className="flex items-center justify-between mb-4">
-        <div className="flex items-center gap-2">
-          <Settings className="w-5 h-5 text-gray-600" />
-          <h3 className="text-lg font-semibold text-gray-900">PP-StructureV3 Parameters</h3>
-          {hasCustomValues && (
-            <span className="text-xs bg-blue-100 text-blue-700 px-2 py-1 rounded">Custom</span>
-          )}
-          {showSaveSuccess && (
-            <span className="flex items-center gap-1 text-xs bg-green-100 text-green-700 px-2 py-1 rounded animate-in fade-in">
-              <Check className="w-3 h-3" />
-              Saved
-            </span>
-          )}
-        </div>
-        <div className="flex items-center gap-2">
-          <button
-            type="button"
-            onClick={() => setIsExpanded(!isExpanded)}
-            className="text-sm text-blue-600 hover:text-blue-700 px-3 py-1.5 rounded-md hover:bg-blue-50"
-          >
-            {isExpanded ? 'Hide' : 'Show'} Parameters
-          </button>
-        </div>
-      </div>
-
-      {/* Preset Selector & Actions */}
-      {isExpanded && (
-        <div className="mb-4 p-3 bg-gray-50 rounded-md space-y-3">
-          <div className="flex items-center gap-3">
-            <label className="text-sm font-medium text-gray-700">Preset:</label>
-            <select
-              value={selectedPreset}
-              onChange={(e) => handlePresetChange(e.target.value)}
-              disabled={disabled}
-              className="flex-1 px-3 py-1.5 text-sm border border-gray-300 rounded-md focus:outline-none focus:ring-2 focus:ring-blue-500 disabled:bg-gray-100"
-            >
-              <option value="default">Default (Backend Settings)</option>
-              <option value="high-quality">High Quality (Lower Thresholds)</option>
-              <option value="fast">Fast (Higher Thresholds)</option>
-              <option value="custom">Custom</option>
-            </select>
-          </div>
-
-          <div className="flex items-center gap-2">
-            <button
-              type="button"
-              onClick={handleReset}
-              disabled={disabled || !hasCustomValues}
-              className={cn(
-                'flex items-center gap-1 px-3 py-1.5 text-sm rounded-md transition-colors',
-                disabled || !hasCustomValues
-                  ? 'bg-gray-200 text-gray-400 cursor-not-allowed'
-                  : 'bg-white border border-gray-300 text-gray-700 hover:bg-gray-50'
-              )}
-            >
-              <RotateCcw className="w-4 h-4" />
-              Reset
-            </button>
-            <button
-              type="button"
-              onClick={handleExport}
-              disabled={disabled || !hasCustomValues}
-              className={cn(
-                'flex items-center gap-1 px-3 py-1.5 text-sm rounded-md transition-colors',
-                disabled || !hasCustomValues
-                  ? 'bg-gray-200 text-gray-400 cursor-not-allowed'
-                  : 'bg-white border border-gray-300 text-gray-700 hover:bg-gray-50'
-              )}
-            >
-              <Download className="w-4 h-4" />
-              Export
-            </button>
-            <button
-              type="button"
-              onClick={handleImport}
-              disabled={disabled}
-              className={cn(
-                'flex items-center gap-1 px-3 py-1.5 text-sm rounded-md transition-colors',
-                disabled
-                  ? 'bg-gray-200 text-gray-400 cursor-not-allowed'
-                  : 'bg-white border border-gray-300 text-gray-700 hover:bg-gray-50'
-              )}
-            >
-              <Upload className="w-4 h-4" />
-              Import
-            </button>
-          </div>
-        </div>
-      )}
-
-      {/* Expanded Parameters */}
-      {isExpanded && (
-        <div className="space-y-6 pt-4 border-t">
-          {PARAM_CONFIGS.map((config) => (
-            <div key={config.key} className="space-y-2">
-              <div className="flex items-center justify-between">
-                <div className="flex items-center gap-2">
-                  <label htmlFor={config.key} className="text-sm font-medium text-gray-700">
-                    {config.label}
-                  </label>
-                  <button
-                    type="button"
-                    onMouseEnter={() => setShowTooltip(config.key)}
-                    onMouseLeave={() => setShowTooltip(null)}
-                    className="text-gray-400 hover:text-gray-600 relative"
-                  >
-                    <HelpCircle className="w-4 h-4" />
-                    {showTooltip === config.key && (
-                      <div className="absolute left-6 top-0 w-64 p-2 bg-gray-900 text-white text-xs rounded shadow-lg z-10">
-                        {config.description}
-                      </div>
-                    )}
-                  </button>
-                </div>
-                {config.type === 'slider' && (
-                  <div className="flex items-center gap-2">
-                    <span className="text-sm font-semibold text-blue-600">
-                      {value[config.key] ?? config.default}
-                    </span>
-                    {value[config.key] !== undefined && value[config.key] !== config.default && (
-                      <span className="text-xs text-gray-500">
-                        (default: {config.default})
-                      </span>
-                    )}
-                  </div>
-                )}
-              </div>
-
-              {config.type === 'slider' ? (
-                <input
-                  type="range"
-                  id={config.key}
-                  min={config.min}
-                  max={config.max}
-                  step={config.step}
-                  value={value[config.key] ?? config.default}
-                  onChange={(e) => handleChange(config.key, parseFloat(e.target.value))}
-                  disabled={disabled}
-                  className="w-full h-2 bg-gray-200 rounded-lg appearance-none cursor-pointer disabled:cursor-not-allowed disabled:opacity-50"
-                />
-              ) : (
-                <select
-                  id={config.key}
-                  value={(value[config.key] as string) ?? config.default}
-                  onChange={(e) => handleChange(config.key, e.target.value)}
-                  disabled={disabled}
-                  className="w-full px-3 py-2 border border-gray-300 rounded-md focus:outline-none focus:ring-2 focus:ring-blue-500 disabled:bg-gray-100 disabled:cursor-not-allowed"
-                >
-                  {config.options.map((option) => (
-                    <option key={option.value} value={option.value}>
-                      {option.label}
-                    </option>
-                  ))}
-                </select>
-              )}
-            </div>
-          ))}
-
-          {/* Info Note */}
-          <div className="mt-4 p-3 bg-blue-50 border border-blue-200 rounded-md">
-            <p className="text-sm text-blue-800">
-              <strong>Note:</strong> These parameters only apply when using the OCR track. Adjusting them
-              can help improve accuracy for specific document types.
-            </p>
-          </div>
-        </div>
-      )}
-
-      {/* Collapsed Summary */}
-      {!isExpanded && hasCustomValues && (
-        <div className="text-sm text-gray-600">
-          {Object.keys(value).length} parameter(s) customized
-        </div>
-      )}
-    </div>
-  )
-}

frontend/src/i18n/locales/zh-TW.json

@@ -52,6 +52,17 @@
       "language": "識別語言",
       "threshold": "信心度閾值",
       "layoutDetection": "版面偵測"
+    },
+    "layoutModel": {
+      "title": "版面偵測模型",
+      "chinese": "中文文件模型",
+      "chineseDesc": "PP-DocLayout-S - 適用於中文表單、合約、發票（推薦）",
+      "default": "標準模型",
+      "defaultDesc": "PubLayNet 模型 - 適用於英文學術論文、報告",
+      "cdla": "CDLA 模型",
+      "cdlaDesc": "專用中文版面分析模型 - 適用於複雜中文版面",
+      "recommended": "推薦",
+      "note": "版面模型會影響文件結構（表格、文字區塊、圖片）的偵測效果。請根據您的文件類型選擇適合的模型。"
     }
   },
   "results": {

frontend/src/pages/ProcessingPage.tsx

@@ -9,10 +9,10 @@ import { Badge } from '@/components/ui/badge'
 import { useToast } from '@/components/ui/toast'
 import { apiClientV2 } from '@/services/apiV2'
 import { Play, CheckCircle, FileText, AlertCircle, Clock, Activity, Loader2 } from 'lucide-react'
-import PPStructureParams from '@/components/PPStructureParams'
+import LayoutModelSelector from '@/components/LayoutModelSelector'
 import TaskNotFound from '@/components/TaskNotFound'
 import { useTaskValidation } from '@/hooks/useTaskValidation'
-import type { PPStructureV3Params, ProcessingOptions } from '@/types/apiV2'
+import type { LayoutModel, ProcessingOptions } from '@/types/apiV2'
 
 export default function ProcessingPage() {
   const { t } = useTranslation()
@@ -31,8 +31,8 @@ export default function ProcessingPage() {
     },
   })
 
-  // PP-StructureV3 parameters state
-  const [ppStructureParams, setPpStructureParams] = useState<PPStructureV3Params>({})
+  // Layout model state (default to 'chinese' for best Chinese document support)
+  const [layoutModel, setLayoutModel] = useState<LayoutModel>('chinese')
 
   // Start OCR processing
   const processOCRMutation = useMutation({
@@ -40,11 +40,7 @@ export default function ProcessingPage() {
       const options: ProcessingOptions = {
         use_dual_track: true,
         language: 'ch',
-      }
-
-      // Only include pp_structure_params if user has customized them
-      if (Object.keys(ppStructureParams).length > 0) {
-        options.pp_structure_params = ppStructureParams
+        layout_model: layoutModel,
       }
 
       return apiClientV2.startTask(taskId!, options)
@@ -346,11 +342,11 @@ export default function ProcessingPage() {
         </Card>
       )}
 
-      {/* PP-StructureV3 Parameters (only show when task is pending) */}
+      {/* Layout Model Selection (only show when task is pending) */}
       {isPending && (
-        <PPStructureParams
-          value={ppStructureParams}
-          onChange={setPpStructureParams}
+        <LayoutModelSelector
+          value={layoutModel}
+          onChange={setLayoutModel}
           disabled={processOCRMutation.isPending}
         />
       )}

frontend/src/types/apiV2.ts

@@ -73,15 +73,14 @@ export interface DocumentAnalysisResponse {
   page_count: number | null
 }
 
-export interface PPStructureV3Params {
-  layout_detection_threshold?: number  // 0-1: Lower=more blocks, Higher=high confidence only
-  layout_nms_threshold?: number       // 0-1: Lower=aggressive overlap removal, Higher=allow more overlap
-  layout_merge_bboxes_mode?: 'union' | 'large' | 'small'  // small=conservative, large=aggressive, union=middle
-  layout_unclip_ratio?: number        // >0: Larger=looser boxes, Smaller=tighter boxes
-  text_det_thresh?: number            // 0-1: Lower=detect more small/low-contrast text, Higher=cleaner
-  text_det_box_thresh?: number        // 0-1: Lower=more text boxes, Higher=fewer false positives
-  text_det_unclip_ratio?: number      // >0: Larger=looser text boxes, Smaller=tighter boxes
-}
+/**
+ * Layout detection model selection for OCR track.
+ * Different models are optimized for different document types:
+ * - chinese: PP-DocLayout-S - Best for Chinese forms, contracts, invoices
+ * - default: PubLayNet-based - Best for English academic papers
+ * - cdla: Specialized for Chinese document layout analysis
+ */
+export type LayoutModel = 'chinese' | 'default' | 'cdla'
 
 export interface ProcessingOptions {
   use_dual_track?: boolean
@@ -89,7 +88,7 @@ export interface ProcessingOptions {
   language?: string
   include_layout?: boolean
   include_images?: boolean
-  pp_structure_params?: PPStructureV3Params  // Fine-tuning parameters for PP-StructureV3 (OCR track only)
+  layout_model?: LayoutModel  // Layout detection model selection (OCR track only)
 }
 
 export interface TaskCreate {

openspec/changes/archive/2025-11-26-fix-ocr-table-empty-columns/proposal.md

@@ -0,0 +1,28 @@
+# Change: Fix OCR Track Table Empty Columns and Alignment
+
+## Why
+
+PP-Structure 生成的表格經常包含空白欄位（所有 row 該欄皆為空/空白），導致轉換後的 UnifiedDocument 表格出現空欄與欄位錯位。目前 OCR Track 直接使用原始資料，未進行清理，影響 PDF/JSON/Markdown 輸出品質。
+
+## What Changes
+
+- 新增 `trim_empty_columns()` 函數，清理 OCR Track 表格的空欄
+- 在 `_convert_table_data` 入口調用清洗邏輯，確保 TableData 乾淨
+- 處理 col_span 重算：若 span 跨過被移除欄位，縮小 span
+- 更新 columns/cols 數值、調整各 cell 的 col 索引
+- 可選：依 bbox x0 進行欄對齊排序
+
+## Impact
+
+- Affected specs: `ocr-processing`
+- Affected code:
+  - `backend/app/services/ocr_to_unified_converter.py` (主要修改)
+- 不影響 Direct/HYBRID 路徑
+- PDF/JSON/Markdown 輸出將更乾淨
+
+## Constraints
+
+- 保持表格 bbox、頁面座標不變
+- 不修改 Direct/HYBRID 路徑
+- 只移除「所有行皆空」的欄；若表頭空但數據有值，不應移除
+- 保留原 bbox，避免 PDF 版面漂移

openspec/changes/archive/2025-11-26-fix-ocr-table-empty-columns/specs/ocr-processing/spec.md

@@ -0,0 +1,61 @@
+## ADDED Requirements
+
+### Requirement: OCR Table Empty Column Cleanup
+
+The OCR Track converter SHALL clean up PP-Structure generated tables by removing columns where all rows have empty or whitespace-only content.
+
+The system SHALL:
+1. Identify columns where every cell's content is empty or contains only whitespace (using `.strip()` to determine emptiness)
+2. Remove identified empty columns from the table structure
+3. Update the `columns`/`cols` value to reflect the new column count
+4. Recalculate each cell's `col` index to maintain continuity
+5. Adjust `col_span` values when spans cross removed columns (shrink span size)
+6. Remove cells entirely when their complete span falls within removed columns
+7. Preserve original bbox and page coordinates (no layout drift)
+8. If `columns` is 0 or missing after cleanup, fill with the calculated column count
+
+The cleanup SHALL NOT:
+- Remove columns where the header is empty but data rows contain values
+- Modify tables in Direct or HYBRID track
+- Alter the original bbox coordinates
+
+#### Scenario: All rows in column are empty
+- **WHEN** a table has a column where all cells contain only empty or whitespace content
+- **THEN** that column is removed
+- **AND** remaining cells have their `col` indices decremented appropriately
+- **AND** `cols` count is reduced by 1
+
+#### Scenario: Column has empty header but data has values
+- **WHEN** a table has a column where the header cell is empty
+- **AND** at least one data row cell in that column contains non-whitespace content
+- **THEN** that column is NOT removed
+
+#### Scenario: Cell span crosses removed column
+- **WHEN** a cell has `col_span > 1`
+- **AND** one or more columns within the span are removed
+- **THEN** the `col_span` is reduced by the number of removed columns within the span
+
+#### Scenario: Cell span entirely within removed columns
+- **WHEN** a cell's entire span falls within columns that are all removed
+- **THEN** that cell is removed from the table
+
+#### Scenario: Missing columns metadata
+- **WHEN** the table dict has `columns` set to 0 or missing
+- **AFTER** cleanup is performed
+- **THEN** `columns` is set to the calculated number of remaining columns
+
+### Requirement: OCR Table Column Alignment by Bbox
+
+(Optional Enhancement) When bbox coordinates are available for table cells, the OCR Track converter SHALL use cell bbox x0 coordinates to improve column alignment accuracy.
+
+The system SHALL:
+1. Sort cells by bbox `x0` coordinate before assigning column indices
+2. Reassign `col` indices based on spatial position rather than HTML order
+
+This requirement is optional and implementation MAY be deferred if bbox data is not reliably available.
+
+#### Scenario: Cells reordered by bbox position
+- **WHEN** bbox coordinates are available for table cells
+- **AND** the original HTML order does not match spatial order
+- **THEN** cells are reordered by `x0` coordinate
+- **AND** `col` indices are reassigned to reflect spatial positioning

openspec/changes/archive/2025-11-26-fix-ocr-table-empty-columns/tasks.md

@@ -0,0 +1,43 @@
+# Tasks: Fix OCR Track Table Empty Columns
+
+## 1. Core Implementation
+
+- [x] 1.1 在 `ocr_to_unified_converter.py` 實作 `trim_empty_columns(table_dict: Dict[str, Any]) -> Dict[str, Any]`
+  - 依據 cells 陣列計算每一欄是否「所有 row 的內容皆為空/空白」
+  - 使用 `.strip()` 判斷空白字元
+- [x] 1.2 實作欄位移除邏輯
+  - 更新 columns/cols 數值
+  - 調整各 cell 的 col 索引
+- [x] 1.3 實作 col_span 重算邏輯
+  - 若 span 跨過被移除欄位，縮小 span
+  - 若整個 span 落在被刪欄位上，移除該 cell
+- [x] 1.4 在 `_convert_table_data` 入口呼叫 `trim_empty_columns`
+  - 在建 TableData 之前執行清洗
+  - 同時也在 `_extract_table_data` (HTML 表格解析) 中加入清洗
+- [ ] 1.5 (可選) 依 bbox x0/x1 進行欄對齊排序
+  - 若可取得 bbox 網格，先依 x0 排序再重排 col index
+  - 此功能延後實作，待 bbox 資料確認可用性後進行
+
+## 2. Testing & Validation
+
+- [x] 2.1 單元測試通過
+  - 測試基本空欄移除
+  - 測試表頭空但數據有值（不移除）
+  - 測試 col_span 跨越被移除欄位（縮小 span）
+  - 測試 cell 完全落在被移除欄位（移除 cell）
+  - 測試無空欄情況（不變更）
+- [x] 2.2 檢查現有 OCR 結果
+  - 現有結果中無「整欄為空」的表格
+  - 實作已就緒，遇到空欄時會正確清理
+- [x] 2.3 確認 Direct/HYBRID 表格不變
+  - `OCRToUnifiedConverter` 僅在 `ocr_service.py` 中使用
+  - Direct 軌使用 `DirectExtractionEngine`，不受影響
+
+## 3. Edge Cases & Validation
+
+- [x] 3.1 處理 columns 欄位為 0/缺失的情況
+  - 以計算後的欄數回填，避免 downstream 依賴出錯
+- [x] 3.2 處理表頭為空但數據有值的情況
+  - 只移除「所有行皆空」的欄
+- [x] 3.3 確保不直接修改 `backend/storage/results/...`
+  - 修改 converter，需重新跑任務驗證

openspec/changes/archive/2025-11-27-add-ocr-track-gap-filling/design.md

@@ -0,0 +1,183 @@
+# Design: OCR Track Gap Filling
+
+## Context
+
+PP-StructureV3 版面分析模型在處理某些掃描文件時會嚴重漏檢。實測顯示 Raw PaddleOCR 能偵測 56 個文字區域，但 PP-StructureV3 僅輸出 9 個元素（遺失 84%）。
+
+問題發生在 PP-StructureV3 內部的 Layout Detection Model，這是 PaddleOCR 函式庫的限制，無法從外部修復。但 Raw OCR 的 `text_regions` 資料仍然完整可用。
+
+### Stakeholders
+- **End users**: 需要完整的 OCR 輸出，不能有大量文字遺失
+- **OCR track**: 需要整合 Raw OCR 與 PP-StructureV3 結果
+- **Direct/Hybrid track**: 不應受此變更影響
+
+## Goals / Non-Goals
+
+### Goals
+- 偵測 PP-StructureV3 漏檢區域並以 Raw OCR 結果補回
+- 確保補回的文字不會與現有元素重複
+- 維持正確的閱讀順序
+- 僅影響 OCR track，不改變其他 track 的行為
+
+### Non-Goals
+- 不修改 PP-StructureV3 或 PaddleOCR 內部邏輯
+- 不處理圖片/表格/圖表等非文字元素的補漏
+- 不實作複雜的版面分析（僅做 gap filling）
+
+## Decisions
+
+### Decision 1: 覆蓋判定策略
+**選擇**: 優先使用「中心點落入」判定，輔以 IoU 閾值
+
+**理由**:
+- 中心點判定計算簡單，效能好
+- IoU 閾值作為補充，處理邊界情況
+- 建議 IoU 閾值 0.1~0.2，避免低 IoU 被誤判為未覆蓋
+
+**替代方案**:
+- 純 IoU 判定：計算量較大，且對部分重疊的處理較複雜
+- 面積比例判定：對不同大小的區域不夠公平
+
+### Decision 2: 補漏觸發條件
+**選擇**: 當 PP-Structure 覆蓋率 < 70% 或元素數顯著低於 Raw OCR
+
+**理由**:
+- 避免正常文件出現重複文字
+- 70% 閾值經驗值，可透過設定調整
+- 元素數比較作為快速判斷條件
+
+### Decision 3: 補漏元素類型
+**選擇**: 僅補 TEXT 類型，跳過 TABLE/IMAGE/FIGURE/FLOWCHART/HEADER/FOOTER
+
+**理由**:
+- PP-StructureV3 對結構化元素（表格、圖片）的識別通常較準確
+- 補回原始 OCR 文字可能破壞表格結構
+- 這些元素需要保持結構完整性
+
+### Decision 4: 重複判定與去重
+**選擇**: IoU > 0.5 的 Raw OCR 區域視為與 PP-Structure TEXT 重複，跳過
+
+**理由**:
+- 0.5 是常見的重疊閾值
+- 避免同一文字出現兩次
+- 對細碎的 Raw OCR 框可考慮輕量合併
+
+### Decision 5: 座標對齊
+**選擇**: 使用 `ocr_dimensions` 進行 bbox 換算
+
+**理由**:
+- OCR 可能有 resize 處理
+- 確保 Raw OCR 與 PP-Structure 的座標在同一空間
+- 避免因尺寸不一致導致覆蓋誤判
+
+## Data Flow
+
+```
+┌─────────────────┐     ┌──────────────────────┐
+│  Raw OCR Result │     │ PP-StructureV3 Result│
+│  (56 regions)   │     │    (9 elements)      │
+└────────┬────────┘     └──────────┬───────────┘
+         │                         │
+         └────────────┬────────────┘
+                      │
+              ┌───────▼───────┐
+              │ GapFillingService │
+              │ 1. Calculate coverage
+              │ 2. Find uncovered regions
+              │ 3. Filter by confidence
+              │ 4. Deduplicate
+              │ 5. Merge if needed
+              └───────┬───────┘
+                      │
+              ┌───────▼───────┐
+              │ OCRToUnifiedConverter │
+              │ - Combine elements
+              │ - Recalculate reading order
+              └───────┬───────┘
+                      │
+              ┌───────▼───────┐
+              │ UnifiedDocument │
+              │ (complete content)
+              └───────────────┘
+```
+
+## Algorithm: Gap Detection
+
+```python
+def find_uncovered_regions(
+    raw_ocr_regions: List[TextRegion],
+    pp_structure_elements: List[Element],
+    iou_threshold: float = 0.15
+) -> List[TextRegion]:
+    """
+    Find Raw OCR regions not covered by PP-Structure elements.
+
+    Coverage criteria (either one):
+    1. Center point of raw region falls inside any PP-Structure bbox
+    2. IoU with any PP-Structure bbox > iou_threshold
+    """
+    uncovered = []
+
+    # Filter PP-Structure elements: only consider TEXT, skip TABLE/IMAGE/etc.
+    text_elements = [e for e in pp_structure_elements
+                     if e.type not in SKIP_TYPES]
+
+    for region in raw_ocr_regions:
+        center = get_center(region.bbox)
+        is_covered = False
+
+        for element in text_elements:
+            # Check center point
+            if point_in_bbox(center, element.bbox):
+                is_covered = True
+                break
+
+            # Check IoU
+            if calculate_iou(region.bbox, element.bbox) > iou_threshold:
+                is_covered = True
+                break
+
+        if not is_covered:
+            uncovered.append(region)
+
+    return uncovered
+```
+
+## Configuration Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `gap_filling_enabled` | bool | True | 是否啟用 gap filling |
+| `gap_filling_coverage_threshold` | float | 0.7 | 覆蓋率低於此值時啟用 |
+| `gap_filling_iou_threshold` | float | 0.15 | 覆蓋判定 IoU 閾值 |
+| `gap_filling_confidence_threshold` | float | 0.3 | Raw OCR 信心度門檻 |
+| `gap_filling_dedup_iou_threshold` | float | 0.5 | 去重 IoU 閾值 |
+
+## Risks / Trade-offs
+
+### Risk 1: 補漏造成文字重複
+**Mitigation**: 設定 dedup_iou_threshold，對高重疊區域進行去重
+
+### Risk 2: 閱讀順序錯亂
+**Mitigation**: 補回元素後重新計算整頁的 reading_order（依 y0, x0 排序）
+
+### Risk 3: 效能影響
+**Mitigation**:
+- 先做快速的覆蓋率檢查，若 > 70% 則跳過 gap filling
+- 使用 R-tree 或 interval tree 加速 bbox 查詢（若效能成為瓶頸）
+
+### Risk 4: 座標不對齊
+**Mitigation**: 使用 `ocr_dimensions` 確保座標空間一致
+
+## Migration Plan
+
+1. 新增功能為可選（預設啟用）
+2. 可透過設定關閉 gap filling
+3. 不影響現有 API 介面
+4. 向後相容：不傳參數時使用預設行為
+
+## Open Questions
+
+1. 是否需要 UI 開關讓使用者選擇啟用/停用 gap filling？
+2. 對於細碎的 Raw OCR 框，是否需要實作合併邏輯？（同行、相鄰且間距很小）
+3. 是否需要在輸出中標記哪些元素是補漏來的？（debug 用途）

openspec/changes/archive/2025-11-27-add-ocr-track-gap-filling/proposal.md

@@ -0,0 +1,30 @@
+# Change: Add OCR Track Gap Filling with Raw OCR Text Regions
+
+## Why
+
+PP-StructureV3 的版面分析模型在處理某些掃描文件時會嚴重漏檢，導致大量文字內容遺失。實測 scan.pdf 顯示：
+- Raw PaddleOCR 文字識別：偵測到 **56 個文字區域**
+- PP-StructureV3 版面分析：僅輸出 **9 個元素**
+- 遺失比例：約 **84%** 的內容未被 PP-StructureV3 識別
+
+問題根源在於 PP-StructureV3 內部的 Layout Detection Model 對掃描文件類型支援不足，而非我們的程式碼問題。Raw OCR 能正確偵測所有文字區域，但這些資訊在 PP-StructureV3 的結構化處理過程中被遺失。
+
+## What Changes
+
+實作「混合式處理」(Hybrid Approach)：使用 Raw OCR 的文字區域來補充 PP-StructureV3 遺失的內容。
+
+- **新增** `GapFillingService` 類別，負責偵測並補回 PP-StructureV3 遺漏的文字區域
+- **新增** 覆蓋率計算邏輯（中心點落入或 IoU 閾值判斷）
+- **新增** 自動啟用條件：當 PP-Structure 覆蓋率 < 70% 或元素數顯著低於 Raw OCR 框數
+- **修改** `OCRToUnifiedConverter` 整合 gap filling 邏輯
+- **新增** 重新計算 reading_order 邏輯（依 y0, x0 排序）
+- **新增** 測試案例：PP-Structure 嚴重漏檢案例、無漏檢正常文件驗證
+
+## Impact
+
+- **Affected specs**: `ocr-processing`
+- **Affected code**:
+  - `backend/app/services/ocr_to_unified_converter.py` - 整合 gap filling
+  - `backend/app/services/gap_filling_service.py` - 新增 (核心邏輯)
+  - `backend/tests/test_gap_filling.py` - 新增 (測試)
+- **Track isolation**: 僅作用於 OCR track；Direct/Hybrid track 不受影響

openspec/changes/archive/2025-11-27-add-ocr-track-gap-filling/specs/ocr-processing/spec.md

@@ -0,0 +1,111 @@
+## ADDED Requirements
+
+### Requirement: OCR Track Gap Filling with Raw OCR Regions
+
+The system SHALL detect and fill gaps in PP-StructureV3 output by supplementing with Raw OCR text regions when significant content loss is detected.
+
+#### Scenario: Gap filling activates when coverage is low
+- **GIVEN** an OCR track processing task
+- **WHEN** PP-StructureV3 outputs elements that cover less than 70% of Raw OCR text regions
+- **THEN** the system SHALL activate gap filling
+- **AND** identify Raw OCR regions not covered by any PP-StructureV3 element
+- **AND** supplement these regions as TEXT elements in the output
+
+#### Scenario: Coverage is determined by center-point and IoU
+- **GIVEN** a Raw OCR text region with bounding box
+- **WHEN** checking if the region is covered by PP-StructureV3
+- **THEN** the region SHALL be considered covered if its center point falls inside any PP-StructureV3 element bbox
+- **OR** if IoU with any PP-StructureV3 element exceeds 0.15 threshold
+- **AND** regions not meeting either criterion SHALL be marked as uncovered
+
+#### Scenario: Only TEXT elements are supplemented
+- **GIVEN** uncovered Raw OCR regions identified for supplementation
+- **WHEN** PP-StructureV3 has detected TABLE, IMAGE, FIGURE, FLOWCHART, HEADER, or FOOTER elements
+- **THEN** the system SHALL NOT supplement regions that overlap with these structural elements
+- **AND** only supplement regions as TEXT type to preserve structural integrity
+
+#### Scenario: Supplemented regions meet confidence threshold
+- **GIVEN** Raw OCR regions to be supplemented
+- **WHEN** a region has confidence score below 0.3
+- **THEN** the system SHALL skip that region
+- **AND** only supplement regions with confidence >= 0.3
+
+#### Scenario: Deduplication prevents repeated text
+- **GIVEN** a Raw OCR region being considered for supplementation
+- **WHEN** the region has IoU > 0.5 with any existing PP-StructureV3 TEXT element
+- **THEN** the system SHALL skip that region to prevent duplicate text
+- **AND** the original PP-StructureV3 element SHALL be preserved
+
+#### Scenario: Reading order is recalculated after gap filling
+- **GIVEN** supplemented elements have been added to the page
+- **WHEN** assembling the final element list
+- **THEN** the system SHALL recalculate reading order for the entire page
+- **AND** sort elements by y0 coordinate (top to bottom) then x0 (left to right)
+- **AND** ensure logical document flow is maintained
+
+#### Scenario: Coordinate alignment with ocr_dimensions
+- **GIVEN** Raw OCR processing may involve image resizing
+- **WHEN** comparing Raw OCR bbox with PP-StructureV3 bbox
+- **THEN** the system SHALL use ocr_dimensions to normalize coordinates
+- **AND** ensure both sources reference the same coordinate space
+- **AND** prevent coverage misdetection due to scale differences
+
+#### Scenario: Supplemented elements have complete metadata
+- **GIVEN** a Raw OCR region being added as supplemented element
+- **WHEN** creating the DocumentElement
+- **THEN** the element SHALL include page_number
+- **AND** include confidence score from Raw OCR
+- **AND** include original bbox coordinates
+- **AND** optionally include source indicator for debugging
+
+### Requirement: Gap Filling Track Isolation
+
+The gap filling feature SHALL only apply to OCR track processing and SHALL NOT affect Direct or Hybrid track outputs.
+
+#### Scenario: Gap filling only activates for OCR track
+- **GIVEN** a document processing task
+- **WHEN** the processing track is OCR
+- **THEN** the system SHALL evaluate and apply gap filling as needed
+- **AND** produce enhanced output with supplemented content
+
+#### Scenario: Direct track is unaffected
+- **GIVEN** a document processing task with Direct track
+- **WHEN** the task is processed
+- **THEN** the system SHALL NOT invoke any gap filling logic
+- **AND** produce output identical to current Direct track behavior
+
+#### Scenario: Hybrid track is unaffected
+- **GIVEN** a document processing task with Hybrid track
+- **WHEN** the task is processed
+- **THEN** the system SHALL NOT invoke gap filling logic
+- **AND** use existing Hybrid track processing pipeline
+
+### Requirement: Gap Filling Configuration
+
+The system SHALL provide configurable parameters for gap filling behavior.
+
+#### Scenario: Gap filling can be disabled via configuration
+- **GIVEN** gap_filling_enabled is set to false in configuration
+- **WHEN** OCR track processing runs
+- **THEN** the system SHALL skip all gap filling logic
+- **AND** output only PP-StructureV3 results as before
+
+#### Scenario: Coverage threshold is configurable
+- **GIVEN** gap_filling_coverage_threshold is set to 0.8
+- **WHEN** PP-StructureV3 coverage is 75%
+- **THEN** the system SHALL activate gap filling
+- **AND** supplement uncovered regions
+
+#### Scenario: IoU thresholds are configurable
+- **GIVEN** custom IoU thresholds configured:
+  - gap_filling_iou_threshold: 0.2
+  - gap_filling_dedup_iou_threshold: 0.6
+- **WHEN** evaluating coverage and deduplication
+- **THEN** the system SHALL use the configured values
+- **AND** apply them consistently throughout gap filling process
+
+#### Scenario: Confidence threshold is configurable
+- **GIVEN** gap_filling_confidence_threshold is set to 0.5
+- **WHEN** supplementing Raw OCR regions
+- **THEN** the system SHALL only include regions with confidence >= 0.5
+- **AND** filter out lower confidence regions

openspec/changes/archive/2025-11-27-add-ocr-track-gap-filling/tasks.md

@@ -0,0 +1,44 @@
+# Tasks: Add OCR Track Gap Filling
+
+## 1. Core Implementation
+
+- [x] 1.1 Create `gap_filling_service.py` with `GapFillingService` class
+- [x] 1.2 Implement bbox coverage calculation (center-point and IoU methods)
+- [x] 1.3 Implement gap detection logic (find uncovered raw OCR regions)
+- [x] 1.4 Implement confidence threshold filtering for supplemented regions
+- [x] 1.5 Implement element type filtering (only supplement TEXT, skip TABLE/IMAGE/FIGURE/etc.)
+- [x] 1.6 Implement reading order recalculation (sort by y0, x0)
+- [x] 1.7 Implement deduplication logic (skip high IoU overlaps with PP-Structure TEXT)
+- [x] 1.8 Implement optional text merging for fragmented adjacent regions
+
+## 2. Integration
+
+- [x] 2.1 Modify `OCRToUnifiedConverter` to accept raw OCR text_regions
+- [x] 2.2 Add gap filling activation condition check (coverage < 70% or element count disparity)
+- [x] 2.3 Ensure coordinate alignment between raw OCR and PP-Structure (ocr_dimensions handling)
+- [x] 2.4 Add page metadata (page_number, confidence, bbox) to supplemented elements
+- [x] 2.5 Ensure track isolation (only OCR track, not Direct/Hybrid)
+
+## 3. Configuration
+
+- [x] 3.1 Add configurable parameters to settings:
+  - `gap_filling_enabled`: bool (default: True)
+  - `gap_filling_coverage_threshold`: float (default: 0.7)
+  - `gap_filling_iou_threshold`: float (default: 0.15)
+  - `gap_filling_confidence_threshold`: float (default: 0.3)
+  - `gap_filling_dedup_iou_threshold`: float (default: 0.5)
+
+## 4. Testing(with env)
+
+- [x] 4.1 Create test fixtures with PP-Structure severe miss-detection case(with scan.pdf / scan2.pdf)
+- [x] 4.2 Test gap detection correctly identifies uncovered regions
+- [x] 4.3 Test supplemented elements have correct metadata
+- [x] 4.4 Test reading order is correctly recalculated
+- [x] 4.5 Test deduplication prevents duplicate text
+- [x] 4.6 Test normal document without miss-detection has no duplicate/inflation
+- [x] 4.7 Test track isolation (Direct track unaffected)
+
+## 5. Documentation
+
+- [x] 5.1 Add inline documentation to GapFillingService
+- [x] 5.2 Update configuration documentation with new settings

openspec/changes/archive/2025-11-27-simplify-ppstructure-model-selection/proposal.md

@@ -0,0 +1,40 @@
+# Change: Simplify PP-StructureV3 Configuration with Layout Model Selection
+
+## Why
+
+Current PP-StructureV3 parameter adjustment UI exposes 7 technical ML parameters (thresholds, ratios, merge modes) that are difficult for end users to understand. Meanwhile, switching to a different layout detection model (e.g., CDLA-trained models for Chinese documents) would have a much greater impact on OCR quality than fine-tuning these parameters.
+
+**Problems with current approach:**
+- Users don't understand what `layout_detection_threshold` or `text_det_unclip_ratio` mean
+- Wrong parameter values can make OCR results worse
+- The default model (PubLayNet-based) is optimized for English academic papers, not Chinese business documents
+- Model selection is far more impactful than parameter tuning
+
+## What Changes
+
+### Backend Changes
+- **REMOVED**: API parameter `pp_structure_params` from task start endpoint
+- **ADDED**: New API parameter `layout_model` with predefined options:
+  - `"default"` - Standard model (PubLayNet-based, for English documents)
+  - `"chinese"` - PP-DocLayout-S model (for Chinese documents, forms, contracts)
+  - `"cdla"` - CDLA model (alternative Chinese document layout model)
+- **MODIFIED**: PP-StructureV3 initialization uses `layout_detection_model_name` based on selection
+- Keep fine-tuning parameters in backend `config.py` with optimized defaults
+
+### Frontend Changes
+- **REMOVED**: `PPStructureParams.tsx` component (slider/dropdown UI for 7 parameters)
+- **ADDED**: Simple radio button/dropdown for layout model selection with clear descriptions
+- **MODIFIED**: Task start request body to send `layout_model` instead of `pp_structure_params`
+
+### API Changes
+- **BREAKING**: Remove `pp_structure_params` from `POST /api/v2/tasks/{task_id}/start`
+- **ADDED**: New optional parameter `layout_model: "default" | "chinese" | "cdla"`
+
+## Impact
+
+- Affected specs: `ocr-processing`
+- Affected code:
+  - Backend: `app/routers/tasks.py`, `app/services/ocr_service.py`, `app/core/config.py`
+  - Frontend: `src/components/PPStructureParams.tsx` (remove), `src/types/apiV2.ts`, task start form
+- Breaking change: Clients using `pp_structure_params` will need to migrate to `layout_model`
+- User impact: Simpler UI, better default OCR quality for Chinese documents

openspec/changes/archive/2025-11-27-simplify-ppstructure-model-selection/specs/ocr-processing/spec.md

@@ -0,0 +1,86 @@
+# ocr-processing Specification Delta
+
+## REMOVED Requirements
+
+### Requirement: Frontend-Adjustable PP-StructureV3 Parameters
+**Reason**: Complex ML parameters are difficult for end users to understand and tune. Model selection provides better UX and more significant quality improvements.
+**Migration**: Replace `pp_structure_params` API parameter with `layout_model` parameter.
+
+### Requirement: PP-StructureV3 Parameter UI Controls
+**Reason**: Slider/dropdown UI for 7 technical parameters adds complexity without proportional benefit. Simple model selection is more user-friendly.
+**Migration**: Remove `PPStructureParams.tsx` component, add `LayoutModelSelector.tsx` component.
+
+## ADDED Requirements
+
+### Requirement: Layout Model Selection
+The system SHALL allow users to select a layout detection model optimized for their document type, providing a simple choice between pre-configured models instead of manual parameter tuning.
+
+#### Scenario: User selects Chinese document model
+- **GIVEN** a user is processing Chinese business documents (forms, contracts, invoices)
+- **WHEN** the user selects "Chinese Document Model" (PP-DocLayout-S)
+- **THEN** the OCR engine SHALL use the PP-DocLayout-S layout detection model
+- **AND** the model SHALL be optimized for 23 Chinese document element types
+- **AND** table and form detection accuracy SHALL be improved over the default model
+
+#### Scenario: User selects standard model for English documents
+- **GIVEN** a user is processing English academic papers or reports
+- **WHEN** the user selects "Standard Model" (PubLayNet-based)
+- **THEN** the OCR engine SHALL use the default PubLayNet-based layout detection model
+- **AND** the model SHALL be optimized for English document layouts
+
+#### Scenario: User selects CDLA model for specialized Chinese layout
+- **GIVEN** a user is processing Chinese documents with complex layouts
+- **WHEN** the user selects "CDLA Model"
+- **THEN** the OCR engine SHALL use the picodet_lcnet_x1_0_fgd_layout_cdla model
+- **AND** the model SHALL provide specialized Chinese document layout analysis
+
+#### Scenario: Layout model is sent via API request
+- **GIVEN** a frontend application with model selection UI
+- **WHEN** the user starts task processing with a selected model
+- **THEN** the frontend SHALL send the model choice in the request body:
+  ```json
+  POST /api/v2/tasks/{task_id}/start
+  {
+    "use_dual_track": true,
+    "force_track": "ocr",
+    "language": "ch",
+    "layout_model": "chinese"
+  }
+  ```
+- **AND** the backend SHALL configure PP-StructureV3 with the corresponding model
+
+#### Scenario: Default model when not specified
+- **GIVEN** an API request without `layout_model` parameter
+- **WHEN** the task is started
+- **THEN** the system SHALL use "chinese" (PP-DocLayout-S) as the default model
+- **AND** processing SHALL work correctly without requiring model selection
+
+#### Scenario: Invalid model name is rejected
+- **GIVEN** a request with an invalid `layout_model` value
+- **WHEN** the user sends `layout_model: "invalid_model"`
+- **THEN** the API SHALL return 422 Validation Error
+- **AND** provide a clear error message listing valid model options
+
+### Requirement: Layout Model Selection UI
+The frontend SHALL provide a simple, user-friendly interface for selecting layout detection models with clear descriptions of each option.
+
+#### Scenario: Model options are displayed with descriptions
+- **GIVEN** the model selection UI is displayed
+- **WHEN** the user views the available options
+- **THEN** the UI SHALL show the following options:
+  - "Chinese Document Model (Recommended)" - for Chinese forms, contracts, invoices
+  - "Standard Model" - for English academic papers, reports
+  - "CDLA Model" - for specialized Chinese layout analysis
+- **AND** each option SHALL have a brief description of its use case
+
+#### Scenario: Chinese model is selected by default
+- **GIVEN** the user opens the task processing interface
+- **WHEN** the model selection is displayed
+- **THEN** "Chinese Document Model" SHALL be pre-selected as the default
+- **AND** the user MAY change the selection before starting processing
+
+#### Scenario: Model selection is visible only for OCR track
+- **GIVEN** a document processing interface
+- **WHEN** the user selects processing track
+- **THEN** layout model selection SHALL be shown ONLY when OCR track is selected or auto-detected
+- **AND** SHALL be hidden for Direct track (which does not use PP-StructureV3)

openspec/changes/archive/2025-11-27-simplify-ppstructure-model-selection/tasks.md

@@ -0,0 +1,56 @@
+# Implementation Tasks
+
+## 1. Backend API Changes
+
+- [x] 1.1 Update `app/schemas/task.py` to add `layout_model` enum type
+- [x] 1.2 Update `app/routers/tasks.py` to replace `pp_structure_params` with `layout_model` parameter
+- [x] 1.3 Update `app/services/ocr_service.py` to map `layout_model` to `layout_detection_model_name`
+- [x] 1.4 Remove custom PP-Structure engine creation logic (use model selection instead)
+- [x] 1.5 Add backward compatibility: default to "chinese" if no model specified
+
+## 2. Backend Configuration
+
+- [x] 2.1 Keep `layout_detection_model_name` in `config.py` as fallback default
+- [x] 2.2 Keep fine-tuning parameters in `config.py` (not exposed to API)
+- [x] 2.3 Document available layout models in config comments
+
+## 3. Frontend Changes
+
+- [x] 3.1 Remove `PPStructureParams.tsx` component
+- [x] 3.2 Update `src/types/apiV2.ts`:
+  - Remove `PPStructureV3Params` interface
+  - Add `LayoutModel` type: `"default" | "chinese" | "cdla"`
+  - Update `ProcessingOptions` to use `layout_model` instead of `pp_structure_params`
+- [x] 3.3 Create `LayoutModelSelector.tsx` component with:
+  - Radio buttons or dropdown for model selection
+  - Clear descriptions for each model option
+  - Default selection: "chinese"
+- [x] 3.4 Update task start form to use new `LayoutModelSelector`
+- [x] 3.5 Update API calls to send `layout_model` instead of `pp_structure_params`
+
+## 4. Internationalization
+
+- [x] 4.1 Add i18n strings for layout model options:
+  - `layoutModel.default`: "Standard Model (English documents)"
+  - `layoutModel.chinese`: "Chinese Document Model (Recommended)"
+  - `layoutModel.cdla`: "CDLA Model (Chinese layout analysis)"
+- [x] 4.2 Add i18n strings for model descriptions
+
+## 5. Testing
+
+- [x] 5.1 Create new tests for `layout_model` parameter (`test_layout_model_api.py`, `test_layout_model.py`)
+- [x] 5.2 Archive tests for `pp_structure_params` validation (moved to `tests/archived/`)
+- [x] 5.3 Add tests for layout model selection (19 tests passing)
+- [x] 5.4 Test backward compatibility (no model specified → use chinese default)
+
+## 6. Documentation
+
+- [ ] 6.1 Update API documentation for task start endpoint
+- [ ] 6.2 Remove PP-Structure parameter documentation
+- [ ] 6.3 Add layout model selection documentation
+
+## 7. Cleanup
+
+- [x] 7.1 Remove localStorage keys for PP-Structure params (`pp_structure_params_presets`, `pp_structure_params_last_used`)
+- [x] 7.2 Remove any unused imports/types related to PP-Structure params
+- [x] 7.3 Archive old PP-Structure params test files

openspec/specs/ocr-processing/spec.md

@@ -3,100 +3,186 @@
 ## Purpose
 TBD - created by archiving change frontend-adjustable-ppstructure-params. Update Purpose after archive.
 ## Requirements
-### Requirement: Frontend-Adjustable PP-StructureV3 Parameters
-The system SHALL allow frontend users to dynamically adjust PP-StructureV3 OCR parameters for fine-tuning document processing without backend configuration changes.
+### Requirement: OCR Track Gap Filling with Raw OCR Regions
 
-#### Scenario: User adjusts layout detection threshold
-- **GIVEN** a user is processing a document with OCR track
-- **WHEN** the user sets `layout_detection_threshold` to 0.1 (lower than default 0.2)
-- **THEN** the OCR engine SHALL detect more layout blocks including weak signals
-- **AND** the processing SHALL use the custom parameter instead of backend defaults
-- **AND** the custom parameter SHALL NOT be cached for reuse
+The system SHALL detect and fill gaps in PP-StructureV3 output by supplementing with Raw OCR text regions when significant content loss is detected.
 
-#### Scenario: User selects high-quality preset configuration
-- **GIVEN** a user wants to process a complex document with many small text elements
-- **WHEN** the user selects "High Quality" preset mode
-- **THEN** the system SHALL automatically set:
-  - `layout_detection_threshold` to 0.1
-  - `layout_nms_threshold` to 0.15
-  - `text_det_thresh` to 0.1
-  - `text_det_box_thresh` to 0.2
-- **AND** process the document with these optimized parameters
+#### Scenario: Gap filling activates when coverage is low
+- **GIVEN** an OCR track processing task
+- **WHEN** PP-StructureV3 outputs elements that cover less than 70% of Raw OCR text regions
+- **THEN** the system SHALL activate gap filling
+- **AND** identify Raw OCR regions not covered by any PP-StructureV3 element
+- **AND** supplement these regions as TEXT elements in the output
 
-#### Scenario: User adjusts text detection parameters
-- **GIVEN** a document with low-contrast text
-- **WHEN** the user sets:
-  - `text_det_thresh` to 0.05 (very low)
-  - `text_det_unclip_ratio` to 1.5 (larger boxes)
-- **THEN** the OCR SHALL detect more small and low-contrast text
-- **AND** text bounding boxes SHALL be expanded by the specified ratio
+#### Scenario: Coverage is determined by center-point and IoU
+- **GIVEN** a Raw OCR text region with bounding box
+- **WHEN** checking if the region is covered by PP-StructureV3
+- **THEN** the region SHALL be considered covered if its center point falls inside any PP-StructureV3 element bbox
+- **OR** if IoU with any PP-StructureV3 element exceeds 0.15 threshold
+- **AND** regions not meeting either criterion SHALL be marked as uncovered
 
-#### Scenario: Parameters are sent via API request body
-- **GIVEN** a frontend application with parameter adjustment UI
-- **WHEN** the user starts task processing with custom parameters
-- **THEN** the frontend SHALL send parameters in the request body (not query params):
+#### Scenario: Only TEXT elements are supplemented
+- **GIVEN** uncovered Raw OCR regions identified for supplementation
+- **WHEN** PP-StructureV3 has detected TABLE, IMAGE, FIGURE, FLOWCHART, HEADER, or FOOTER elements
+- **THEN** the system SHALL NOT supplement regions that overlap with these structural elements
+- **AND** only supplement regions as TEXT type to preserve structural integrity
+
+#### Scenario: Supplemented regions meet confidence threshold
+- **GIVEN** Raw OCR regions to be supplemented
+- **WHEN** a region has confidence score below 0.3
+- **THEN** the system SHALL skip that region
+- **AND** only supplement regions with confidence >= 0.3
+
+#### Scenario: Deduplication prevents repeated text
+- **GIVEN** a Raw OCR region being considered for supplementation
+- **WHEN** the region has IoU > 0.5 with any existing PP-StructureV3 TEXT element
+- **THEN** the system SHALL skip that region to prevent duplicate text
+- **AND** the original PP-StructureV3 element SHALL be preserved
+
+#### Scenario: Reading order is recalculated after gap filling
+- **GIVEN** supplemented elements have been added to the page
+- **WHEN** assembling the final element list
+- **THEN** the system SHALL recalculate reading order for the entire page
+- **AND** sort elements by y0 coordinate (top to bottom) then x0 (left to right)
+- **AND** ensure logical document flow is maintained
+
+#### Scenario: Coordinate alignment with ocr_dimensions
+- **GIVEN** Raw OCR processing may involve image resizing
+- **WHEN** comparing Raw OCR bbox with PP-StructureV3 bbox
+- **THEN** the system SHALL use ocr_dimensions to normalize coordinates
+- **AND** ensure both sources reference the same coordinate space
+- **AND** prevent coverage misdetection due to scale differences
+
+#### Scenario: Supplemented elements have complete metadata
+- **GIVEN** a Raw OCR region being added as supplemented element
+- **WHEN** creating the DocumentElement
+- **THEN** the element SHALL include page_number
+- **AND** include confidence score from Raw OCR
+- **AND** include original bbox coordinates
+- **AND** optionally include source indicator for debugging
+
+### Requirement: Gap Filling Track Isolation
+
+The gap filling feature SHALL only apply to OCR track processing and SHALL NOT affect Direct or Hybrid track outputs.
+
+#### Scenario: Gap filling only activates for OCR track
+- **GIVEN** a document processing task
+- **WHEN** the processing track is OCR
+- **THEN** the system SHALL evaluate and apply gap filling as needed
+- **AND** produce enhanced output with supplemented content
+
+#### Scenario: Direct track is unaffected
+- **GIVEN** a document processing task with Direct track
+- **WHEN** the task is processed
+- **THEN** the system SHALL NOT invoke any gap filling logic
+- **AND** produce output identical to current Direct track behavior
+
+#### Scenario: Hybrid track is unaffected
+- **GIVEN** a document processing task with Hybrid track
+- **WHEN** the task is processed
+- **THEN** the system SHALL NOT invoke gap filling logic
+- **AND** use existing Hybrid track processing pipeline
+
+### Requirement: Gap Filling Configuration
+
+The system SHALL provide configurable parameters for gap filling behavior.
+
+#### Scenario: Gap filling can be disabled via configuration
+- **GIVEN** gap_filling_enabled is set to false in configuration
+- **WHEN** OCR track processing runs
+- **THEN** the system SHALL skip all gap filling logic
+- **AND** output only PP-StructureV3 results as before
+
+#### Scenario: Coverage threshold is configurable
+- **GIVEN** gap_filling_coverage_threshold is set to 0.8
+- **WHEN** PP-StructureV3 coverage is 75%
+- **THEN** the system SHALL activate gap filling
+- **AND** supplement uncovered regions
+
+#### Scenario: IoU thresholds are configurable
+- **GIVEN** custom IoU thresholds configured:
+  - gap_filling_iou_threshold: 0.2
+  - gap_filling_dedup_iou_threshold: 0.6
+- **WHEN** evaluating coverage and deduplication
+- **THEN** the system SHALL use the configured values
+- **AND** apply them consistently throughout gap filling process
+
+#### Scenario: Confidence threshold is configurable
+- **GIVEN** gap_filling_confidence_threshold is set to 0.5
+- **WHEN** supplementing Raw OCR regions
+- **THEN** the system SHALL only include regions with confidence >= 0.5
+- **AND** filter out lower confidence regions
+
+### Requirement: Layout Model Selection
+The system SHALL allow users to select a layout detection model optimized for their document type, providing a simple choice between pre-configured models instead of manual parameter tuning.
+
+#### Scenario: User selects Chinese document model
+- **GIVEN** a user is processing Chinese business documents (forms, contracts, invoices)
+- **WHEN** the user selects "Chinese Document Model" (PP-DocLayout-S)
+- **THEN** the OCR engine SHALL use the PP-DocLayout-S layout detection model
+- **AND** the model SHALL be optimized for 23 Chinese document element types
+- **AND** table and form detection accuracy SHALL be improved over the default model
+
+#### Scenario: User selects standard model for English documents
+- **GIVEN** a user is processing English academic papers or reports
+- **WHEN** the user selects "Standard Model" (PubLayNet-based)
+- **THEN** the OCR engine SHALL use the default PubLayNet-based layout detection model
+- **AND** the model SHALL be optimized for English document layouts
+
+#### Scenario: User selects CDLA model for specialized Chinese layout
+- **GIVEN** a user is processing Chinese documents with complex layouts
+- **WHEN** the user selects "CDLA Model"
+- **THEN** the OCR engine SHALL use the picodet_lcnet_x1_0_fgd_layout_cdla model
+- **AND** the model SHALL provide specialized Chinese document layout analysis
+
+#### Scenario: Layout model is sent via API request
+- **GIVEN** a frontend application with model selection UI
+- **WHEN** the user starts task processing with a selected model
+- **THEN** the frontend SHALL send the model choice in the request body:
   ```json
   POST /api/v2/tasks/{task_id}/start
   {
     "use_dual_track": true,
     "force_track": "ocr",
     "language": "ch",
-    "pp_structure_params": {
-      "layout_detection_threshold": 0.15,
-      "layout_merge_bboxes_mode": "small",
-      "text_det_thresh": 0.1
-    }
+    "layout_model": "chinese"
   }
   ```
-- **AND** the backend SHALL parse and apply these parameters
+- **AND** the backend SHALL configure PP-StructureV3 with the corresponding model
 
-#### Scenario: Backward compatibility is maintained
-- **GIVEN** existing API clients without PP-StructureV3 parameter support
-- **WHEN** a task is started without `pp_structure_params`
-- **THEN** the system SHALL use backend default settings
-- **AND** processing SHALL work exactly as before
-- **AND** no errors SHALL occur
+#### Scenario: Default model when not specified
+- **GIVEN** an API request without `layout_model` parameter
+- **WHEN** the task is started
+- **THEN** the system SHALL use "chinese" (PP-DocLayout-S) as the default model
+- **AND** processing SHALL work correctly without requiring model selection
 
-#### Scenario: Invalid parameters are rejected
-- **GIVEN** a request with invalid parameter values
-- **WHEN** the user sends:
-  - `layout_detection_threshold` = 1.5 (exceeds max 1.0)
-  - `layout_merge_bboxes_mode` = "invalid" (not in allowed values)
+#### Scenario: Invalid model name is rejected
+- **GIVEN** a request with an invalid `layout_model` value
+- **WHEN** the user sends `layout_model: "invalid_model"`
 - **THEN** the API SHALL return 422 Validation Error
-- **AND** provide clear error messages about invalid parameters
+- **AND** provide a clear error message listing valid model options
 
-#### Scenario: Custom parameters affect only current processing
-- **GIVEN** multiple concurrent OCR processing tasks
-- **WHEN** Task A uses custom parameters and Task B uses defaults
-- **THEN** Task A SHALL process with its custom parameters
-- **AND** Task B SHALL process with default parameters
-- **AND** no parameter interference SHALL occur between tasks
+### Requirement: Layout Model Selection UI
+The frontend SHALL provide a simple, user-friendly interface for selecting layout detection models with clear descriptions of each option.
 
-### Requirement: PP-StructureV3 Parameter UI Controls
-The frontend SHALL provide intuitive UI controls for adjusting PP-StructureV3 parameters with appropriate constraints and help text.
+#### Scenario: Model options are displayed with descriptions
+- **GIVEN** the model selection UI is displayed
+- **WHEN** the user views the available options
+- **THEN** the UI SHALL show the following options:
+  - "Chinese Document Model (Recommended)" - for Chinese forms, contracts, invoices
+  - "Standard Model" - for English academic papers, reports
+  - "CDLA Model" - for specialized Chinese layout analysis
+- **AND** each option SHALL have a brief description of its use case
 
-#### Scenario: Slider controls for numeric parameters
-- **GIVEN** the parameter adjustment UI is displayed
-- **WHEN** the user adjusts a numeric parameter slider
-- **THEN** the slider SHALL enforce min/max constraints:
-  - Threshold parameters: 0.0 to 1.0
-  - Ratio parameters: > 0 (typically 0.5 to 3.0)
-- **AND** display current value in real-time
-- **AND** show help text explaining the parameter effect
+#### Scenario: Chinese model is selected by default
+- **GIVEN** the user opens the task processing interface
+- **WHEN** the model selection is displayed
+- **THEN** "Chinese Document Model" SHALL be pre-selected as the default
+- **AND** the user MAY change the selection before starting processing
 
-#### Scenario: Dropdown for merge mode selection
-- **GIVEN** the layout merge mode parameter
-- **WHEN** the user clicks the dropdown
-- **THEN** the UI SHALL show exactly three options:
-  - "small" (conservative merging)
-  - "large" (aggressive merging)
-  - "union" (middle ground)
-- **AND** display description for each option
-
-#### Scenario: Parameters shown only for OCR track
+#### Scenario: Model selection is visible only for OCR track
 - **GIVEN** a document processing interface
 - **WHEN** the user selects processing track
-- **THEN** PP-StructureV3 parameters SHALL be shown ONLY when OCR track is selected
-- **AND** SHALL be hidden for Direct track
-- **AND** SHALL be disabled for Auto track until track is determined
+- **THEN** layout model selection SHALL be shown ONLY when OCR track is selected or auto-detected
+- **AND** SHALL be hidden for Direct track (which does not use PP-StructureV3)