docling 2.39.0__py3-none-any.whl → 2.41.0__py3-none-any.whl

docling/backend/docling_parse_v4_backend.py

@@ -187,7 +187,17 @@ class DoclingParseV4DocumentBackend(PdfDocumentBackend):
 
     def unload(self):
         super().unload()
-        self.dp_doc.unload()
-        with pypdfium2_lock:
-            self._pdoc.close()
-        self._pdoc = None
+        # Unload docling-parse document first
+        if self.dp_doc is not None:
+            self.dp_doc.unload()
+            self.dp_doc = None
+
+        # Then close pypdfium2 document with proper locking
+        if self._pdoc is not None:
+            with pypdfium2_lock:
+                try:
+                    self._pdoc.close()
+                except Exception:
+                    # Ignore cleanup errors
+                    pass
+            self._pdoc = None

docling/backend/msexcel_backend.py

@@ -337,10 +337,17 @@ class MsExcelDocumentBackend(DeclarativeDocumentBackend, PaginatedDocumentBacken
         # Collect the data within the bounds
         data = []
         visited_cells: set[tuple[int, int]] = set()
-        for ri in range(start_row, max_row + 1):
-            for rj in range(start_col, max_col + 1):
-                cell = sheet.cell(row=ri + 1, column=rj + 1)  # 1-based indexing
-
+        for ri, row in enumerate(
+            sheet.iter_rows(
+                min_row=start_row + 1,  # start_row is 0-based but iter_rows is 1-based
+                max_row=max_row + 1,
+                min_col=start_col + 1,
+                max_col=max_col + 1,
+                values_only=False,
+            ),
+            start_row,
+        ):
+            for rj, cell in enumerate(row, start_col):
                 # Check if the cell belongs to a merged range
                 row_span = 1
                 col_span = 1
@@ -397,10 +404,16 @@ class MsExcelDocumentBackend(DeclarativeDocumentBackend, PaginatedDocumentBacken
         """
         max_row: int = start_row
 
-        while max_row < sheet.max_row - 1:
-            # Get the cell value or check if it is part of a merged cell
-            cell = sheet.cell(row=max_row + 2, column=start_col + 1)
-
+        for ri, (cell,) in enumerate(
+            sheet.iter_rows(
+                min_row=start_row + 2,
+                max_row=sheet.max_row,
+                min_col=start_col + 1,
+                max_col=start_col + 1,
+                values_only=False,
+            ),
+            start_row + 1,
+        ):
             # Check if the cell is part of a merged range
             merged_range = next(
                 (mr for mr in sheet.merged_cells.ranges if cell.coordinate in mr),
@@ -414,7 +427,7 @@ class MsExcelDocumentBackend(DeclarativeDocumentBackend, PaginatedDocumentBacken
             if merged_range:
                 max_row = max(max_row, merged_range.max_row - 1)
             else:
-                max_row += 1
+                max_row = ri
 
         return max_row
 
@@ -433,10 +446,16 @@ class MsExcelDocumentBackend(DeclarativeDocumentBackend, PaginatedDocumentBacken
         """
         max_col: int = start_col
 
-        while max_col < sheet.max_column - 1:
-            # Get the cell value or check if it is part of a merged cell
-            cell = sheet.cell(row=start_row + 1, column=max_col + 2)
-
+        for rj, (cell,) in enumerate(
+            sheet.iter_cols(
+                min_row=start_row + 1,
+                max_row=start_row + 1,
+                min_col=start_col + 2,
+                max_col=sheet.max_column,
+                values_only=False,
+            ),
+            start_col + 1,
+        ):
             # Check if the cell is part of a merged range
             merged_range = next(
                 (mr for mr in sheet.merged_cells.ranges if cell.coordinate in mr),
@@ -450,7 +469,7 @@ class MsExcelDocumentBackend(DeclarativeDocumentBackend, PaginatedDocumentBacken
             if merged_range:
                 max_col = max(max_col, merged_range.max_col - 1)
             else:
-                max_col += 1
+                max_col = rj
 
         return max_col
 

docling/datamodel/asr_model_specs.py

@@ -22,7 +22,7 @@ WHISPER_TINY = InlineAsrNativeWhisperOptions(
     verbose=True,
     timestamps=True,
     word_timestamps=True,
-    temperatue=0.0,
+    temperature=0.0,
     max_new_tokens=256,
     max_time_chunk=30.0,
 )
@@ -33,7 +33,7 @@ WHISPER_SMALL = InlineAsrNativeWhisperOptions(
     verbose=True,
     timestamps=True,
     word_timestamps=True,
-    temperatue=0.0,
+    temperature=0.0,
     max_new_tokens=256,
     max_time_chunk=30.0,
 )
@@ -44,7 +44,7 @@ WHISPER_MEDIUM = InlineAsrNativeWhisperOptions(
     verbose=True,
     timestamps=True,
     word_timestamps=True,
-    temperatue=0.0,
+    temperature=0.0,
     max_new_tokens=256,
     max_time_chunk=30.0,
 )
@@ -55,7 +55,7 @@ WHISPER_BASE = InlineAsrNativeWhisperOptions(
     verbose=True,
     timestamps=True,
     word_timestamps=True,
-    temperatue=0.0,
+    temperature=0.0,
     max_new_tokens=256,
     max_time_chunk=30.0,
 )
@@ -66,7 +66,7 @@ WHISPER_LARGE = InlineAsrNativeWhisperOptions(
     verbose=True,
     timestamps=True,
     word_timestamps=True,
-    temperatue=0.0,
+    temperature=0.0,
     max_new_tokens=256,
     max_time_chunk=30.0,
 )
@@ -77,7 +77,7 @@ WHISPER_TURBO = InlineAsrNativeWhisperOptions(
     verbose=True,
     timestamps=True,
     word_timestamps=True,
-    temperatue=0.0,
+    temperature=0.0,
     max_new_tokens=256,
     max_time_chunk=30.0,
 )

docling/datamodel/base_models.py

@@ -12,6 +12,7 @@ from docling_core.types.doc import (
     Size,
     TableCell,
 )
+from docling_core.types.doc.base import PydanticSerCtxKey, round_pydantic_float
 from docling_core.types.doc.page import SegmentedPdfPage, TextCell
 from docling_core.types.io import (
     DocumentStream,
@@ -19,7 +20,14 @@ from docling_core.types.io import (
 
 # DO NOT REMOVE; explicitly exposed from this location
 from PIL.Image import Image
-from pydantic import BaseModel, ConfigDict, Field, computed_field
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    FieldSerializationInfo,
+    computed_field,
+    field_serializer,
+)
 
 if TYPE_CHECKING:
     from docling.backend.pdf_backend import PdfPageBackend
@@ -142,6 +150,10 @@ class Cluster(BaseModel):
     cells: List[TextCell] = []
     children: List["Cluster"] = []  # Add child cluster support
 
+    @field_serializer("confidence")
+    def _serialize(self, value: float, info: FieldSerializationInfo) -> float:
+        return round_pydantic_float(value, info.context, PydanticSerCtxKey.CONFID_PREC)
+
 
 class BasePageElement(BaseModel):
     label: DocItemLabel
@@ -194,6 +206,16 @@ class FigureElement(BasePageElement):
     predicted_class: Optional[str] = None
     confidence: Optional[float] = None
 
+    @field_serializer("confidence")
+    def _serialize(
+        self, value: Optional[float], info: FieldSerializationInfo
+    ) -> Optional[float]:
+        return (
+            round_pydantic_float(value, info.context, PydanticSerCtxKey.CONFID_PREC)
+            if value is not None
+            else None
+        )
+
 
 class FigureClassificationPrediction(BaseModel):
     figure_count: int = 0

docling/datamodel/layout_model_specs.py

@@ -0,0 +1,90 @@
+import logging
+from enum import Enum
+from pathlib import Path
+from typing import Optional
+
+from pydantic import BaseModel
+
+from docling.datamodel.accelerator_options import AcceleratorDevice
+
+_log = logging.getLogger(__name__)
+
+
+class LayoutModelConfig(BaseModel):
+    name: str
+    repo_id: str
+    revision: str
+    model_path: str
+    supported_devices: list[AcceleratorDevice] = [
+        AcceleratorDevice.CPU,
+        AcceleratorDevice.CUDA,
+        AcceleratorDevice.MPS,
+    ]
+
+    @property
+    def model_repo_folder(self) -> str:
+        return self.repo_id.replace("/", "--")
+
+
+# HuggingFace Layout Models
+
+# Default Docling Layout Model
+DOCLING_LAYOUT_V2 = LayoutModelConfig(
+    name="docling_layout_v2",
+    repo_id="ds4sd/docling-layout-old",
+    revision="main",
+    model_path="",
+)
+
+DOCLING_LAYOUT_HERON = LayoutModelConfig(
+    name="docling_layout_heron",
+    repo_id="ds4sd/docling-layout-heron",
+    revision="main",
+    model_path="",
+)
+
+DOCLING_LAYOUT_HERON_101 = LayoutModelConfig(
+    name="docling_layout_heron_101",
+    repo_id="ds4sd/docling-layout-heron-101",
+    revision="main",
+    model_path="",
+)
+
+DOCLING_LAYOUT_EGRET_MEDIUM = LayoutModelConfig(
+    name="docling_layout_egret_medium",
+    repo_id="ds4sd/docling-layout-egret-medium",
+    revision="main",
+    model_path="",
+)
+
+DOCLING_LAYOUT_EGRET_LARGE = LayoutModelConfig(
+    name="docling_layout_egret_large",
+    repo_id="ds4sd/docling-layout-egret-large",
+    revision="main",
+    model_path="",
+)
+
+DOCLING_LAYOUT_EGRET_XLARGE = LayoutModelConfig(
+    name="docling_layout_egret_xlarge",
+    repo_id="ds4sd/docling-layout-egret-xlarge",
+    revision="main",
+    model_path="",
+)
+
+# Example for a hypothetical alternative model
+# ALTERNATIVE_LAYOUT = LayoutModelConfig(
+#     name="alternative_layout",
+#     repo_id="someorg/alternative-layout",
+#     revision="main",
+#     model_path="model_artifacts/layout_alt",
+# )
+
+
+class LayoutModelType(str, Enum):
+    DOCLING_LAYOUT_V2 = "docling_layout_v2"
+    DOCLING_LAYOUT_HERON = "docling_layout_heron"
+    DOCLING_LAYOUT_HERON_101 = "docling_layout_heron_101"
+    DOCLING_LAYOUT_EGRET_MEDIUM = "docling_layout_egret_medium"
+    DOCLING_LAYOUT_EGRET_LARGE = "docling_layout_egret_large"
+    DOCLING_LAYOUT_EGRET_XLARGE = "docling_layout_egret_xlarge"
+    # ALTERNATIVE_LAYOUT = "alternative_layout"

docling/datamodel/pipeline_options.py

@@ -1,4 +1,5 @@
 import logging
+from datetime import datetime
 from enum import Enum
 from pathlib import Path
 from typing import Any, ClassVar, Dict, List, Literal, Optional, Union
@@ -15,6 +16,15 @@ from docling.datamodel import asr_model_specs
 
 # Import the following for backwards compatibility
 from docling.datamodel.accelerator_options import AcceleratorDevice, AcceleratorOptions
+from docling.datamodel.layout_model_specs import (
+    DOCLING_LAYOUT_EGRET_LARGE,
+    DOCLING_LAYOUT_EGRET_MEDIUM,
+    DOCLING_LAYOUT_EGRET_XLARGE,
+    DOCLING_LAYOUT_HERON,
+    DOCLING_LAYOUT_HERON_101,
+    DOCLING_LAYOUT_V2,
+    LayoutModelConfig,
+)
 from docling.datamodel.pipeline_options_asr_model import (
     InlineAsrOptions,
 )
@@ -265,6 +275,13 @@ class VlmPipelineOptions(PaginatedPipelineOptions):
     )
 
 
+class LayoutOptions(BaseModel):
+    """Options for layout processing."""
+
+    create_orphan_clusters: bool = True  # Whether to create clusters for orphaned cells
+    model_spec: LayoutModelConfig = DOCLING_LAYOUT_V2
+
+
 class AsrPipelineOptions(PipelineOptions):
     asr_options: Union[InlineAsrOptions] = asr_model_specs.WHISPER_TINY
     artifacts_path: Optional[Union[Path, str]] = None
@@ -289,6 +306,7 @@ class PdfPipelineOptions(PaginatedPipelineOptions):
     picture_description_options: PictureDescriptionBaseOptions = (
         smolvlm_picture_description
     )
+    layout_options: LayoutOptions = LayoutOptions()
 
     images_scale: float = 1.0
     generate_page_images: bool = False

docling/datamodel/pipeline_options_vlm_model.py

@@ -1,6 +1,7 @@
 from enum import Enum
-from typing import Any, Dict, List, Literal, Optional, Union
+from typing import Any, Callable, Dict, List, Literal, Optional, Union
 
+from docling_core.types.doc.page import SegmentedPage
 from pydantic import AnyUrl, BaseModel
 from typing_extensions import deprecated
 
@@ -9,9 +10,10 @@ from docling.datamodel.accelerator_options import AcceleratorDevice
 
 class BaseVlmOptions(BaseModel):
     kind: str
-    prompt: str
+    prompt: Union[str, Callable[[Optional[SegmentedPage]], str]]
     scale: float = 2.0
     max_size: Optional[int] = None
+    temperature: float = 0.0
 
 
 class ResponseFormat(str, Enum):
@@ -29,6 +31,12 @@ class TransformersModelType(str, Enum):
     AUTOMODEL = "automodel"
     AUTOMODEL_VISION2SEQ = "automodel-vision2seq"
     AUTOMODEL_CAUSALLM = "automodel-causallm"
+    AUTOMODEL_IMAGETEXTTOTEXT = "automodel-imagetexttotext"
+
+
+class TransformersPromptStyle(str, Enum):
+    CHAT = "chat"
+    RAW = "raw"
 
 
 class InlineVlmOptions(BaseVlmOptions):
@@ -42,6 +50,7 @@ class InlineVlmOptions(BaseVlmOptions):
 
     inference_framework: InferenceFramework
     transformers_model_type: TransformersModelType = TransformersModelType.AUTOMODEL
+    transformers_prompt_style: TransformersPromptStyle = TransformersPromptStyle.CHAT
     response_format: ResponseFormat
 
     torch_dtype: Optional[str] = None
@@ -51,7 +60,6 @@ class InlineVlmOptions(BaseVlmOptions):
         AcceleratorDevice.MPS,
     ]
 
-    temperature: float = 0.0
     stop_strings: List[str] = []
     extra_generation_config: Dict[str, Any] = {}
 

docling/models/api_vlm_model.py

@@ -29,12 +29,9 @@ class ApiVlmModel(BasePageModel):
 
             self.timeout = self.vlm_options.timeout
             self.concurrency = self.vlm_options.concurrency
-            self.prompt_content = (
-                f"This is a page from a document.\n{self.vlm_options.prompt}"
-            )
             self.params = {
                 **self.vlm_options.params,
-                "temperature": 0,
+                "temperature": self.vlm_options.temperature,
             }
 
     def __call__(
@@ -56,9 +53,14 @@ class ApiVlmModel(BasePageModel):
                         if hi_res_image.mode != "RGB":
                             hi_res_image = hi_res_image.convert("RGB")
 
+                    if callable(self.vlm_options.prompt):
+                        prompt = self.vlm_options.prompt(page.parsed_page)
+                    else:
+                        prompt = self.vlm_options.prompt
+
                     page_tags = api_image_request(
                         image=hi_res_image,
-                        prompt=self.prompt_content,
+                        prompt=prompt,
                         url=self.vlm_options.url,
                         timeout=self.timeout,
                         headers=self.vlm_options.headers,

docling/models/base_ocr_model.py

@@ -3,14 +3,13 @@ import logging
 from abc import abstractmethod
 from collections.abc import Iterable
 from pathlib import Path
-from typing import List, Optional, Type
+from typing import TYPE_CHECKING, List, Optional, Type
 
 import numpy as np
 from docling_core.types.doc import BoundingBox, CoordOrigin
 from docling_core.types.doc.page import TextCell
 from PIL import Image, ImageDraw
 from rtree import index
-from scipy.ndimage import binary_dilation, find_objects, label
 
 from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import Page
@@ -31,11 +30,16 @@ class BaseOcrModel(BasePageModel, BaseModelWithOptions):
         options: OcrOptions,
         accelerator_options: AcceleratorOptions,
     ):
+        # Make sure any delay/error from import occurs on ocr model init and not first use
+        from scipy.ndimage import binary_dilation, find_objects, label
+
         self.enabled = enabled
         self.options = options
 
     # Computes the optimum amount and coordinates of rectangles to OCR on a given page
     def get_ocr_rects(self, page: Page) -> List[BoundingBox]:
+        from scipy.ndimage import binary_dilation, find_objects, label
+
         BITMAP_COVERAGE_TRESHOLD = 0.75
         assert page.size is not None
 

docling/models/document_picture_classifier.py

@@ -14,7 +14,8 @@ from PIL import Image
 from pydantic import BaseModel
 
 from docling.datamodel.accelerator_options import AcceleratorOptions
-from docling.models.base_model import BaseEnrichmentModel
+from docling.datamodel.base_models import ItemAndImageEnrichmentElement
+from docling.models.base_model import BaseItemAndImageEnrichmentModel
 from docling.models.utils.hf_model_download import download_hf_model
 from docling.utils.accelerator_utils import decide_device
 
@@ -32,7 +33,7 @@ class DocumentPictureClassifierOptions(BaseModel):
     kind: Literal["document_picture_classifier"] = "document_picture_classifier"
 
 
-class DocumentPictureClassifier(BaseEnrichmentModel):
+class DocumentPictureClassifier(BaseItemAndImageEnrichmentModel):
     """
     A model for classifying pictures in documents.
 
@@ -135,7 +136,7 @@ class DocumentPictureClassifier(BaseEnrichmentModel):
     def __call__(
         self,
         doc: DoclingDocument,
-        element_batch: Iterable[NodeItem],
+        element_batch: Iterable[ItemAndImageEnrichmentElement],
     ) -> Iterable[NodeItem]:
         """
         Processes a batch of elements and enriches them with classification predictions.
@@ -144,7 +145,7 @@ class DocumentPictureClassifier(BaseEnrichmentModel):
         ----------
         doc : DoclingDocument
             The document containing the elements to be processed.
-        element_batch : Iterable[NodeItem]
+        element_batch : Iterable[ItemAndImageEnrichmentElement]
             A batch of pictures to classify.
 
         Returns
@@ -155,22 +156,20 @@ class DocumentPictureClassifier(BaseEnrichmentModel):
         """
         if not self.enabled:
             for element in element_batch:
-                yield element
+                yield element.item
             return
 
         images: List[Union[Image.Image, np.ndarray]] = []
         elements: List[PictureItem] = []
         for el in element_batch:
-            assert isinstance(el, PictureItem)
-            elements.append(el)
-            img = el.get_image(doc)
-            assert img is not None
-            images.append(img)
+            assert isinstance(el.item, PictureItem)
+            elements.append(el.item)
+            images.append(el.image)
 
         outputs = self.document_picture_classifier.predict(images)
 
-        for element, output in zip(elements, outputs):
-            element.annotations.append(
+        for item, output in zip(elements, outputs):
+            item.annotations.append(
                 PictureClassificationData(
                     provenance="DocumentPictureClassifier",
                     predicted_classes=[
@@ -183,4 +182,4 @@ class DocumentPictureClassifier(BaseEnrichmentModel):
                 )
             )
 
-            yield element
+            yield item

docling/models/layout_model.py

@@ -7,12 +7,13 @@ from typing import Optional
 
 import numpy as np
 from docling_core.types.doc import DocItemLabel
-from docling_ibm_models.layoutmodel.layout_predictor import LayoutPredictor
 from PIL import Image
 
 from docling.datamodel.accelerator_options import AcceleratorOptions
 from docling.datamodel.base_models import BoundingBox, Cluster, LayoutPrediction, Page
 from docling.datamodel.document import ConversionResult
+from docling.datamodel.layout_model_specs import DOCLING_LAYOUT_V2, LayoutModelConfig
+from docling.datamodel.pipeline_options import LayoutOptions
 from docling.datamodel.settings import settings
 from docling.models.base_model import BasePageModel
 from docling.models.utils.hf_model_download import download_hf_model
@@ -25,9 +26,6 @@ _log = logging.getLogger(__name__)
 
 
 class LayoutModel(BasePageModel):
-    _model_repo_folder = "ds4sd--docling-models"
-    _model_path = "model_artifacts/layout"
-
     TEXT_ELEM_LABELS = [
         DocItemLabel.TEXT,
         DocItemLabel.FOOTNOTE,
@@ -49,28 +47,38 @@ class LayoutModel(BasePageModel):
     CONTAINER_LABELS = [DocItemLabel.FORM, DocItemLabel.KEY_VALUE_REGION]
 
     def __init__(
-        self, artifacts_path: Optional[Path], accelerator_options: AcceleratorOptions
+        self,
+        artifacts_path: Optional[Path],
+        accelerator_options: AcceleratorOptions,
+        options: LayoutOptions,
     ):
+        from docling_ibm_models.layoutmodel.layout_predictor import LayoutPredictor
+
+        self.options = options
+
         device = decide_device(accelerator_options.device)
+        layout_model_config = options.model_spec
+        model_repo_folder = layout_model_config.model_repo_folder
+        model_path = layout_model_config.model_path
 
         if artifacts_path is None:
-            artifacts_path = self.download_models() / self._model_path
+            artifacts_path = (
+                self.download_models(layout_model_config=layout_model_config)
+                / model_path
+            )
         else:
-            # will become the default in the future
-            if (artifacts_path / self._model_repo_folder).exists():
-                artifacts_path = (
-                    artifacts_path / self._model_repo_folder / self._model_path
-                )
-            elif (artifacts_path / self._model_path).exists():
+            if (artifacts_path / model_repo_folder).exists():
+                artifacts_path = artifacts_path / model_repo_folder / model_path
+            elif (artifacts_path / model_path).exists():
                 warnings.warn(
                     "The usage of artifacts_path containing directly "
-                    f"{self._model_path} is deprecated. Please point "
+                    f"{model_path} is deprecated. Please point "
                     "the artifacts_path to the parent containing "
-                    f"the {self._model_repo_folder} folder.",
+                    f"the {model_repo_folder} folder.",
                     DeprecationWarning,
                     stacklevel=3,
                 )
-                artifacts_path = artifacts_path / self._model_path
+                artifacts_path = artifacts_path / model_path
 
         self.layout_predictor = LayoutPredictor(
             artifact_path=str(artifacts_path),
@@ -83,10 +91,11 @@ class LayoutModel(BasePageModel):
         local_dir: Optional[Path] = None,
         force: bool = False,
         progress: bool = False,
+        layout_model_config: LayoutModelConfig = DOCLING_LAYOUT_V2,
     ) -> Path:
         return download_hf_model(
-            repo_id="ds4sd/docling-models",
-            revision="v2.2.0",
+            repo_id=layout_model_config.repo_id,
+            revision=layout_model_config.revision,
             local_dir=local_dir,
             force=force,
             progress=progress,
@@ -176,7 +185,7 @@ class LayoutModel(BasePageModel):
                     # Apply postprocessing
 
                     processed_clusters, processed_cells = LayoutPostprocessor(
-                        page, clusters
+                        page, clusters, self.options
                     ).postprocess()
                     # Note: LayoutPostprocessor updates page.cells and page.parsed_page internally
 

docling/models/picture_description_vlm_model.py

@@ -1,3 +1,4 @@
+import threading
 from collections.abc import Iterable
 from pathlib import Path
 from typing import Optional, Type, Union
@@ -15,6 +16,9 @@ from docling.models.utils.hf_model_download import (
 )
 from docling.utils.accelerator_utils import decide_device
 
+# Global lock for model initialization to prevent threading issues
+_model_init_lock = threading.Lock()
+
 
 class PictureDescriptionVlmModel(
     PictureDescriptionBaseModel, HuggingFaceModelDownloadMixin
@@ -57,17 +61,18 @@ class PictureDescriptionVlmModel(
                 )
 
             # Initialize processor and model
-            self.processor = AutoProcessor.from_pretrained(artifacts_path)
-            self.model = AutoModelForVision2Seq.from_pretrained(
-                artifacts_path,
-                torch_dtype=torch.bfloat16,
-                _attn_implementation=(
-                    "flash_attention_2"
-                    if self.device.startswith("cuda")
-                    and accelerator_options.cuda_use_flash_attention2
-                    else "eager"
-                ),
-            ).to(self.device)
+            with _model_init_lock:
+                self.processor = AutoProcessor.from_pretrained(artifacts_path)
+                self.model = AutoModelForVision2Seq.from_pretrained(
+                    artifacts_path,
+                    torch_dtype=torch.bfloat16,
+                    _attn_implementation=(
+                        "flash_attention_2"
+                        if self.device.startswith("cuda")
+                        and accelerator_options.cuda_use_flash_attention2
+                        else "eager"
+                    ),
+                ).to(self.device)
 
             self.provenance = f"{self.options.repo_id}"
 

docling/models/plugins/defaults.py

@@ -1,13 +1,10 @@
-from docling.models.easyocr_model import EasyOcrModel
-from docling.models.ocr_mac_model import OcrMacModel
-from docling.models.picture_description_api_model import PictureDescriptionApiModel
-from docling.models.picture_description_vlm_model import PictureDescriptionVlmModel
-from docling.models.rapid_ocr_model import RapidOcrModel
-from docling.models.tesseract_ocr_cli_model import TesseractOcrCliModel
-from docling.models.tesseract_ocr_model import TesseractOcrModel
-
-
 def ocr_engines():
+    from docling.models.easyocr_model import EasyOcrModel
+    from docling.models.ocr_mac_model import OcrMacModel
+    from docling.models.rapid_ocr_model import RapidOcrModel
+    from docling.models.tesseract_ocr_cli_model import TesseractOcrCliModel
+    from docling.models.tesseract_ocr_model import TesseractOcrModel
+
     return {
         "ocr_engines": [
             EasyOcrModel,
@@ -20,6 +17,9 @@ def ocr_engines():
 
 
 def picture_description():
+    from docling.models.picture_description_api_model import PictureDescriptionApiModel
+    from docling.models.picture_description_vlm_model import PictureDescriptionVlmModel
+
     return {
         "picture_description": [
             PictureDescriptionVlmModel,

docling/models/readingorder_model.py

@@ -12,6 +12,9 @@ from docling_core.types.doc import (
     TableData,
 )
 from docling_core.types.doc.document import ContentLayer
+from docling_ibm_models.list_item_normalizer.list_marker_processor import (
+    ListItemMarkerProcessor,
+)
 from docling_ibm_models.reading_order.reading_order_rb import (
     PageElement as ReadingOrderPageElement,
     ReadingOrderPredictor,
@@ -40,6 +43,7 @@ class ReadingOrderModel:
     def __init__(self, options: ReadingOrderOptions):
         self.options = options
         self.ro_model = ReadingOrderPredictor()
+        self.list_item_processor = ListItemMarkerProcessor()
 
     def _assembled_to_readingorder_elements(
         self, conv_res: ConversionResult
@@ -92,7 +96,8 @@ class ReadingOrderModel:
             )
             if c_label == DocItemLabel.LIST_ITEM:
                 # TODO: Infer if this is a numbered or a bullet list item
-                doc.add_list_item(parent=doc_item, text=c_text, prov=c_prov)
+                l_item = doc.add_list_item(parent=doc_item, text=c_text, prov=c_prov)
+                self.list_item_processor.process_list_item(l_item)
             elif c_label == DocItemLabel.SECTION_HEADER:
                 doc.add_heading(parent=doc_item, text=c_text, prov=c_prov)
             else:
@@ -301,6 +306,8 @@ class ReadingOrderModel:
             new_item = out_doc.add_list_item(
                 text=cap_text, enumerated=False, prov=prov, parent=current_list
             )
+            self.list_item_processor.process_list_item(new_item)
+
         elif label == DocItemLabel.SECTION_HEADER:
             current_list = None
 

docling/models/table_structure_model.py

@@ -10,7 +10,6 @@ from docling_core.types.doc.page import (
     BoundingRectangle,
     TextCellUnit,
 )
-from docling_ibm_models.tableformer.data_management.tf_predictor import TFPredictor
 from PIL import ImageDraw
 
 from docling.datamodel.accelerator_options import AcceleratorDevice, AcceleratorOptions
@@ -70,6 +69,9 @@ class TableStructureModel(BasePageModel):
 
             # Third Party
             import docling_ibm_models.tableformer.common as c
+            from docling_ibm_models.tableformer.data_management.tf_predictor import (
+                TFPredictor,
+            )
 
             device = decide_device(accelerator_options.device)
 

docling/models/tesseract_ocr_model.py

@@ -144,7 +144,10 @@ class TesseractOcrModel(BaseOcrModel):
 
                         local_reader = self.reader
                         self.osd_reader.SetImage(high_res_image)
+
+                        doc_orientation = 0
                         osd = self.osd_reader.DetectOrientationScript()
+
                         # No text, or Orientation and Script detection failure
                         if osd is None:
                             _log.error(
@@ -158,11 +161,14 @@ class TesseractOcrModel(BaseOcrModel):
                             # to OCR in the hope OCR will succeed while OSD failed
                             if self._is_auto:
                                 continue
-                        doc_orientation = parse_tesseract_orientation(osd["orient_deg"])
-                        if doc_orientation != 0:
-                            high_res_image = high_res_image.rotate(
-                                -doc_orientation, expand=True
+                        else:
+                            doc_orientation = parse_tesseract_orientation(
+                                osd["orient_deg"]
                             )
+                            if doc_orientation != 0:
+                                high_res_image = high_res_image.rotate(
+                                    -doc_orientation, expand=True
+                                )
                         if self._is_auto:
                             script = osd["script_name"]
                             script = map_tesseract_script(script)

docling/models/vlm_models_inline/hf_transformers_model.py

@@ -13,6 +13,7 @@ from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options_vlm_model import (
     InlineVlmOptions,
     TransformersModelType,
+    TransformersPromptStyle,
 )
 from docling.models.base_model import BasePageModel
 from docling.models.utils.hf_model_download import (
@@ -41,6 +42,7 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
             from transformers import (
                 AutoModel,
                 AutoModelForCausalLM,
+                AutoModelForImageTextToText,
                 AutoModelForVision2Seq,
                 AutoProcessor,
                 BitsAndBytesConfig,
@@ -91,6 +93,11 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
                 == TransformersModelType.AUTOMODEL_VISION2SEQ
             ):
                 model_cls = AutoModelForVision2Seq
+            elif (
+                self.vlm_options.transformers_model_type
+                == TransformersModelType.AUTOMODEL_IMAGETEXTTOTEXT
+            ):
+                model_cls = AutoModelForImageTextToText
 
             self.processor = AutoProcessor.from_pretrained(
                 artifacts_path,
@@ -128,7 +135,11 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
                     )
 
                     # Define prompt structure
-                    prompt = self.formulate_prompt()
+                    if callable(self.vlm_options.prompt):
+                        user_prompt = self.vlm_options.prompt(page.parsed_page)
+                    else:
+                        user_prompt = self.vlm_options.prompt
+                    prompt = self.formulate_prompt(user_prompt)
 
                     inputs = self.processor(
                         text=prompt, images=[hi_res_image], return_tensors="pt"
@@ -162,10 +173,13 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
 
                 yield page
 
-    def formulate_prompt(self) -> str:
+    def formulate_prompt(self, user_prompt: str) -> str:
         """Formulate a prompt for the VLM."""
 
-        if self.vlm_options.repo_id == "microsoft/Phi-4-multimodal-instruct":
+        if self.vlm_options.transformers_prompt_style == TransformersPromptStyle.RAW:
+            return user_prompt
+
+        elif self.vlm_options.repo_id == "microsoft/Phi-4-multimodal-instruct":
             _log.debug("Using specialized prompt for Phi-4")
             # more info here: https://huggingface.co/microsoft/Phi-4-multimodal-instruct#loading-the-model-locally
 
@@ -173,25 +187,30 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
             assistant_prompt = "<|assistant|>"
             prompt_suffix = "<|end|>"
 
-            prompt = f"{user_prompt}<|image_1|>{self.vlm_options.prompt}{prompt_suffix}{assistant_prompt}"
+            prompt = f"{user_prompt}<|image_1|>{user_prompt}{prompt_suffix}{assistant_prompt}"
             _log.debug(f"prompt for {self.vlm_options.repo_id}: {prompt}")
 
             return prompt
 
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {
-                        "type": "text",
-                        "text": "This is a page from a document.",
-                    },
-                    {"type": "image"},
-                    {"type": "text", "text": self.vlm_options.prompt},
-                ],
-            }
-        ]
-        prompt = self.processor.apply_chat_template(
-            messages, add_generation_prompt=False
+        elif self.vlm_options.transformers_prompt_style == TransformersPromptStyle.CHAT:
+            messages = [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "text",
+                            "text": "This is a page from a document.",
+                        },
+                        {"type": "image"},
+                        {"type": "text", "text": user_prompt},
+                    ],
+                }
+            ]
+            prompt = self.processor.apply_chat_template(
+                messages, add_generation_prompt=False
+            )
+            return prompt
+
+        raise RuntimeError(
+            f"Uknown prompt style `{self.vlm_options.transformers_prompt_style}`. Valid values are {', '.join(s.value for s in TransformersPromptStyle)}."
         )
-        return prompt

docling/models/vlm_models_inline/mlx_model.py

@@ -56,8 +56,6 @@ class HuggingFaceMlxModel(BasePageModel, HuggingFaceModelDownloadMixin):
             elif (artifacts_path / repo_cache_folder).exists():
                 artifacts_path = artifacts_path / repo_cache_folder
 
-            self.param_question = vlm_options.prompt
-
             ## Load the model
             self.vlm_model, self.processor = load(artifacts_path)
             self.config = load_config(artifacts_path)
@@ -86,8 +84,12 @@ class HuggingFaceMlxModel(BasePageModel, HuggingFaceModelDownloadMixin):
                         if hi_res_image.mode != "RGB":
                             hi_res_image = hi_res_image.convert("RGB")
 
+                    if callable(self.vlm_options.prompt):
+                        user_prompt = self.vlm_options.prompt(page.parsed_page)
+                    else:
+                        user_prompt = self.vlm_options.prompt
                     prompt = self.apply_chat_template(
-                        self.processor, self.config, self.param_question, num_images=1
+                        self.processor, self.config, user_prompt, num_images=1
                     )
 
                     start_time = time.time()

docling/pipeline/standard_pdf_pipeline.py

@@ -10,6 +10,7 @@ from docling.backend.abstract_backend import AbstractDocumentBackend
 from docling.backend.pdf_backend import PdfDocumentBackend
 from docling.datamodel.base_models import AssembledUnit, Page
 from docling.datamodel.document import ConversionResult
+from docling.datamodel.layout_model_specs import LayoutModelConfig
 from docling.datamodel.pipeline_options import PdfPipelineOptions
 from docling.datamodel.settings import settings
 from docling.models.base_ocr_model import BaseOcrModel
@@ -36,9 +37,6 @@ _log = logging.getLogger(__name__)
 
 
 class StandardPdfPipeline(PaginatedPipeline):
-    _layout_model_path = LayoutModel._model_path
-    _table_model_path = TableStructureModel._model_path
-
     def __init__(self, pipeline_options: PdfPipelineOptions):
         super().__init__(pipeline_options)
         self.pipeline_options: PdfPipelineOptions
@@ -80,6 +78,7 @@ class StandardPdfPipeline(PaginatedPipeline):
             LayoutModel(
                 artifacts_path=artifacts_path,
                 accelerator_options=pipeline_options.accelerator_options,
+                options=pipeline_options.layout_options,
             ),
             # Table structure model
             TableStructureModel(
@@ -128,6 +127,7 @@ class StandardPdfPipeline(PaginatedPipeline):
         if (
             self.pipeline_options.do_formula_enrichment
             or self.pipeline_options.do_code_enrichment
+            or self.pipeline_options.do_picture_classification
             or self.pipeline_options.do_picture_description
         ):
             self.keep_backend = True

docling/pipeline/vlm_pipeline.py

@@ -117,6 +117,7 @@ class VlmPipeline(PaginatedPipeline):
             page._backend = conv_res.input._backend.load_page(page.page_no)  # type: ignore
             if page._backend is not None and page._backend.is_valid():
                 page.size = page._backend.get_size()
+                page.parsed_page = page._backend.get_segmented_page()
 
         return page
 

docling/utils/accelerator_utils.py

@@ -1,8 +1,6 @@
 import logging
 from typing import List, Optional
 
-import torch
-
 from docling.datamodel.accelerator_options import AcceleratorDevice
 
 _log = logging.getLogger(__name__)
@@ -18,6 +16,8 @@ def decide_device(
     1. AUTO: Check for the best available device on the system.
     2. User-defined: Check if the device actually exists, otherwise fall-back to CPU
     """
+    import torch
+
     device = "cpu"
 
     has_cuda = torch.backends.cuda.is_built() and torch.cuda.is_available()

docling/utils/layout_postprocessor.py

@@ -9,6 +9,7 @@ from docling_core.types.doc.page import TextCell
 from rtree import index
 
 from docling.datamodel.base_models import BoundingBox, Cluster, Page
+from docling.datamodel.pipeline_options import LayoutOptions
 
 _log = logging.getLogger(__name__)
 
@@ -194,12 +195,16 @@ class LayoutPostprocessor:
         DocItemLabel.TITLE: DocItemLabel.SECTION_HEADER,
     }
 
-    def __init__(self, page: Page, clusters: List[Cluster]) -> None:
+    def __init__(
+        self, page: Page, clusters: List[Cluster], options: LayoutOptions
+    ) -> None:
         """Initialize processor with page and clusters."""
+
         self.cells = page.cells
         self.page = page
         self.page_size = page.size
         self.all_clusters = clusters
+        self.options = options
         self.regular_clusters = [
             c for c in clusters if c.label not in self.SPECIAL_TYPES
         ]
@@ -267,7 +272,7 @@ class LayoutPostprocessor:
 
         # Handle orphaned cells
         unassigned = self._find_unassigned_cells(clusters)
-        if unassigned:
+        if unassigned and self.options.create_orphan_clusters:
             next_id = max((c.id for c in self.all_clusters), default=0) + 1
             orphan_clusters = []
             for i, cell in enumerate(unassigned):

docling/utils/model_downloader.py

@@ -2,6 +2,7 @@ import logging
 from pathlib import Path
 from typing import Optional
 
+from docling.datamodel.layout_model_specs import DOCLING_LAYOUT_V2
 from docling.datamodel.pipeline_options import (
     granite_picture_description,
     smolvlm_picture_description,
@@ -46,7 +47,7 @@ def download_models(
     if with_layout:
         _log.info("Downloading layout model...")
         LayoutModel.download_models(
-            local_dir=output_dir / LayoutModel._model_repo_folder,
+            local_dir=output_dir / DOCLING_LAYOUT_V2.model_repo_folder,
             force=force,
             progress=progress,
         )

docling/utils/ocr_utils.py

@@ -41,7 +41,7 @@ def tesseract_box_to_bounding_rectangle(
     im_size: Tuple[int, int],
 ) -> BoundingRectangle:
     # box is in the top, left, height, width format, top left coordinates
-    rect = rotate_bounding_box(bbox, angle=-orientation, im_size=im_size)
+    rect = rotate_bounding_box(bbox, angle=orientation, im_size=im_size)
     rect = BoundingRectangle(
         r_x0=rect.r_x0 / scale,
         r_y0=rect.r_y0 / scale,

docling/utils/orientation.py

@@ -14,43 +14,36 @@ def rotate_bounding_box(
     # coordinate system. Then other corners are found rotating counterclockwise
     bbox = bbox.to_top_left_origin(im_size[1])
     left, top, width, height = bbox.l, bbox.t, bbox.width, bbox.height
-    im_h, im_w = im_size
+    im_w, im_h = im_size
     angle = angle % 360
     if angle == 0:
-        r_x0 = left
-        r_y0 = top + height
-        r_x1 = r_x0 + width
-        r_y1 = r_y0
-        r_x2 = r_x0 + width
-        r_y2 = r_y0 - height
-        r_x3 = r_x0
-        r_y3 = r_y0 - height
+        return BoundingRectangle.from_bounding_box(bbox)
     elif angle == 90:
-        r_x0 = im_w - (top + height)
-        r_y0 = left
+        r_x0 = top + height
+        r_y0 = im_w - left
         r_x1 = r_x0
-        r_y1 = r_y0 + width
-        r_x2 = r_x0 + height
-        r_y2 = r_y0 + width
-        r_x3 = r_x0
-        r_y3 = r_y0 + width
+        r_y1 = r_y0 - width
+        r_x2 = r_x1 - height
+        r_y2 = r_y1
+        r_x3 = r_x2
+        r_y3 = r_y0
     elif angle == 180:
-        r_x0 = im_h - left
-        r_y0 = im_w - (top + height)
+        r_x0 = im_w - left
+        r_y0 = im_h - (top + height)
         r_x1 = r_x0 - width
         r_y1 = r_y0
-        r_x2 = r_x0 - width
-        r_y2 = r_y0 + height
+        r_x2 = r_x1
+        r_y2 = r_y1 + height
         r_x3 = r_x0
-        r_y3 = r_y0 + height
+        r_y3 = r_y2
     elif angle == 270:
-        r_x0 = top + height
-        r_y0 = im_h - left
+        r_x0 = im_h - (top + height)
+        r_y0 = left
         r_x1 = r_x0
-        r_y1 = r_y0 - width
-        r_x2 = r_x0 - height
-        r_y2 = r_y0 - width
-        r_x3 = r_x0 - height
+        r_y1 = r_y0 + width
+        r_x2 = r_x1 + height
+        r_y2 = r_y1
+        r_x3 = r_x2
         r_y3 = r_y0
     else:
         msg = (
@@ -58,7 +51,7 @@ def rotate_bounding_box(
             f" {sorted(CLIPPED_ORIENTATIONS)}"
         )
         raise ValueError(msg)
-    return BoundingRectangle(
+    rectangle = BoundingRectangle(
         r_x0=r_x0,
         r_y0=r_y0,
         r_x1=r_x1,
@@ -69,3 +62,4 @@ def rotate_bounding_box(
         r_y3=r_y3,
         coord_origin=CoordOrigin.TOPLEFT,
     )
+    return rectangle

{docling-2.39.0.dist-info → docling-2.41.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docling
-Version: 2.39.0
+Version: 2.41.0
 Summary: SDK and CLI for parsing PDF, DOCX, HTML, and more, to a unified document representation for powering downstream workflows such as gen AI applications.
 Author-email: Christoph Auer <cau@zurich.ibm.com>, Michele Dolfi <dol@zurich.ibm.com>, Maxim Lysak <mly@zurich.ibm.com>, Nikos Livathinos <nli@zurich.ibm.com>, Ahmed Nassar <ahn@zurich.ibm.com>, Panos Vagenas <pva@zurich.ibm.com>, Peter Staar <taa@zurich.ibm.com>
 License-Expression: MIT
@@ -26,9 +26,9 @@ Requires-Python: <4.0,>=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: pydantic<3.0.0,>=2.0.0
-Requires-Dist: docling-core[chunking]<3.0.0,>=2.39.0
-Requires-Dist: docling-ibm-models<4.0.0,>=3.4.4
+Requires-Dist: docling-core[chunking]<3.0.0,>=2.42.0
 Requires-Dist: docling-parse<5.0.0,>=4.0.0
+Requires-Dist: docling-ibm-models<4,>=3.6.0
 Requires-Dist: filetype<2.0.0,>=1.2.0
 Requires-Dist: pypdfium2<5.0.0,>=4.30.0
 Requires-Dist: pydantic-settings<3.0.0,>=2.3.0
@@ -57,12 +57,12 @@ Requires-Dist: ocrmac<2.0.0,>=1.0.0; sys_platform == "darwin" and extra == "ocrm
 Provides-Extra: vlm
 Requires-Dist: transformers<5.0.0,>=4.46.0; extra == "vlm"
 Requires-Dist: accelerate<2.0.0,>=1.2.1; extra == "vlm"
-Requires-Dist: mlx-vlm>=0.1.22; (python_version >= "3.10" and sys_platform == "darwin" and platform_machine == "arm64") and extra == "vlm"
+Requires-Dist: mlx-vlm<0.2,>=0.1.22; (python_version >= "3.10" and sys_platform == "darwin" and platform_machine == "arm64") and extra == "vlm"
 Provides-Extra: rapidocr
 Requires-Dist: rapidocr-onnxruntime<2.0.0,>=1.4.0; python_version < "3.13" and extra == "rapidocr"
 Requires-Dist: onnxruntime<2.0.0,>=1.7.0; extra == "rapidocr"
 Provides-Extra: asr
-Requires-Dist: openai-whisper>=20240930; extra == "asr"
+Requires-Dist: openai-whisper>=20250625; extra == "asr"
 Dynamic: license-file
 
 <p align="center">

{docling-2.39.0.dist-info → docling-2.41.0.dist-info}/RECORD

@@ -8,10 +8,10 @@ docling/backend/asciidoc_backend.py,sha256=RDNLrPJHxROiM7-NQdZn3DdvAyiPAndbSWcZo
 docling/backend/csv_backend.py,sha256=2g9famYG2W-ID9jEdZPxc6O8QGv1vWQfjN8pL-QMBE0,4536
 docling/backend/docling_parse_backend.py,sha256=9rUo1vPxX6QLzGqF-2B2iEYglZg6YQ3Uea00XrLluTg,7918
 docling/backend/docling_parse_v2_backend.py,sha256=3ckTfke8IICjaImlIzc3TRhG7KDuxDDba0AuCEcjA-M,9500
-docling/backend/docling_parse_v4_backend.py,sha256=7tQvpCwpYoq98PNszDkrXaFhy5eWmQqMP4RjWWPLPgw,6197
+docling/backend/docling_parse_v4_backend.py,sha256=qR_WRVq9JGtRioWCw6MnLWgbvXbC6Y1yds7Ol1-E6UQ,6550
 docling/backend/html_backend.py,sha256=Z959dzqYQO2pPE4xgPRxC5MR9j3nFGtiD6_F_osQ2iI,20670
 docling/backend/md_backend.py,sha256=mfwGj8g2hGC-Q_HREtl_Web65uMVXD-Ie1nRqWTXzF0,21013
-docling/backend/msexcel_backend.py,sha256=3j0WQfqDpgPXdPMCguefdv7arcNVDedPD6gl54cmLn8,18110
+docling/backend/msexcel_backend.py,sha256=cq8MQ2RSh6pqCiVrldjOerSww7dOPTWmCQoCBI57i6w,18579
 docling/backend/mspowerpoint_backend.py,sha256=wJgB2JStEPfD7MPpWQlpPN7bffPxaHFUnKD4wj8SLxU,15114
 docling/backend/msword_backend.py,sha256=7mzPCF4bGWZPst5ntoV3aSxH5WUu2nBP-l8lgQT3tdw,44544
 docling/backend/noop_backend.py,sha256=EOPbD86FzZPX-K_DpNrJh0_lC0bZz--4DpG-OagDNGY,1688
@@ -33,66 +33,67 @@ docling/cli/models.py,sha256=9yLGp6QRJGpR86U3SjmWAXDt3MvBaJLLY4xDVdsu3O8,4160
 docling/cli/tools.py,sha256=QhtRxQG0TVrfsMqdv5i7J0_qQy1ZZyWYnHPwJl7b5oY,322
 docling/datamodel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling/datamodel/accelerator_options.py,sha256=wv6dOFTVAwr9onkE-0pfUqX_fDb6gX53iPPE6o8nKjI,2511
-docling/datamodel/asr_model_specs.py,sha256=L7ETXsUKVbPsVcPLhEIMxQjd4UzMGZBVsy74CLsZBkU,2181
-docling/datamodel/base_models.py,sha256=67o1ptOTT8tW7i-g6gM2JKEX_1CDbmKEMQ_B9ZYM2z0,11156
+docling/datamodel/asr_model_specs.py,sha256=Wg7z3zm_wXIWu122iPVy0RMECsA_JCFHrlFF-xxHoVQ,2187
+docling/datamodel/base_models.py,sha256=9FslHkGUNmBp264LpLL_2JTfDAdaikldYs3SiQOHb5A,11828
 docling/datamodel/document.py,sha256=CA_dgt4V_phze5HXpfgfKNBKd1cPC1o3WE_IENX63EM,16252
-docling/datamodel/pipeline_options.py,sha256=7mKv1IThXYpu3osggp_Y2h7E5C8nbxJLQXS7JJPMvYQ,9479
+docling/datamodel/layout_model_specs.py,sha256=GSkJ-Z_0PVgwWGi7C7TsxbzRjlrWS9ZrHJjHumv-Z5U,2339
+docling/datamodel/pipeline_options.py,sha256=aMwpbyEMbAC-xGJnjQp8iw2ocpSU4eiD8D73gHf7T4U,10033
 docling/datamodel/pipeline_options_asr_model.py,sha256=7X068xl-qpbyPxC7-TwX7Q6tLyZXGT5h1osZ_xLNLM0,1454
-docling/datamodel/pipeline_options_vlm_model.py,sha256=rtDMVtKFZbgQD269w8FvHMXEhdRBrsA4rVYk6A-M-b4,2063
+docling/datamodel/pipeline_options_vlm_model.py,sha256=z-pUqwRA8nJp6C3SEXZLem2zvSYdgavaAVYa8wkAIZY,2400
 docling/datamodel/settings.py,sha256=ajMz7Ao2m0ZGYkfArqTDDbiF89O408mtgeh06PUi0MA,1900
 docling/datamodel/vlm_model_specs.py,sha256=--jZexGeu-s_lWp7y_WwWEf6CD1J4XqADrS1-OY_pWM,4737
 docling/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-docling/models/api_vlm_model.py,sha256=GDDJGAia4SJjK7JFxsZy5oEU-D8yQo8Kb3NvvPbTvT0,2820
+docling/models/api_vlm_model.py,sha256=foBvzaWeHFH1t-VdvRWLdiXiiofhvhjvHqRI0eNA_3w,2923
 docling/models/base_model.py,sha256=NNjIapqCruAEAWR-CCdsNgXc2QkwiPYAcaQ_ZYe1W28,2978
-docling/models/base_ocr_model.py,sha256=HtrefTq9Zy4UnUInMchPv0tbobiA7CQU5VUauKJD7IU,8006
+docling/models/base_ocr_model.py,sha256=kT8TylASOpPlY60rIG6VL6_eLVsfg5KvEVnZHzDWtR0,8193
 docling/models/code_formula_model.py,sha256=5uWh-eI-Ejmv3DujKJoKKgJBuvPLokt7AJ_ybt8VHEw,11373
-docling/models/document_picture_classifier.py,sha256=fkJLV7pMy3v6iNwOzVb6zdBU1dGtBM1ARHLIRPfoAG4,6124
+docling/models/document_picture_classifier.py,sha256=9JvoWeH5uQBC7levjM8zptk7UT-b8EQnD-2EnxTjTT4,6202
 docling/models/easyocr_model.py,sha256=ECPBd-48cCw5s935NsPJO_C_1QuK_yAUGloMM77WqIM,7387
-docling/models/layout_model.py,sha256=EJuRXW0rFdnNPS5AifdEsr812EATUqAioeMCVjw8PL0,8460
+docling/models/layout_model.py,sha256=8bfLVKCS2A-ePTQK-T4M2K_Ah-jUVj71YOtwZvZ9rsU,8825
 docling/models/ocr_mac_model.py,sha256=y-1DSFDbACHpEwNTfQwzN9ab8r5j5rBFNPtQ48BzsrA,5396
 docling/models/page_assemble_model.py,sha256=TvN1naez7dUodLxpUUBzpuMCpqZBTf6YSpewxgjzmrg,6323
 docling/models/page_preprocessing_model.py,sha256=x8MI4mvjizqEqAb5511dtrNRCJSb-lSmwHw0tmHPFiI,5103
 docling/models/picture_description_api_model.py,sha256=o3EkV5aHW_6WzE_fdj_VRnNCrS_btclO_ZCLAUqrfl0,2377
 docling/models/picture_description_base_model.py,sha256=kLthLhdlgwhootQ4_xhhcAk6A-vso5-qcsFJ3TcYfO0,2991
-docling/models/picture_description_vlm_model.py,sha256=7LeCx9ZdPxsmWJ468OtxCdAkH48A1HD0iwH9cs_7-1Q,3800
+docling/models/picture_description_vlm_model.py,sha256=nAUt-eZOX2GvaCiV2BJO7VppxUbP7udVIF4oe_sEYXo,4000
 docling/models/rapid_ocr_model.py,sha256=AMdc66s_iWO4p6nQ0LNjQMUYVxrDSxMyLNPpjPYt6N8,5916
-docling/models/readingorder_model.py,sha256=QHb5fyiqmxU8lg4W5IzdukqHPh6V7rNw_57O4-z-Az4,14615
-docling/models/table_structure_model.py,sha256=dQf6u_zn5fHCkHzmTwYfCbRtZCBddsyAM0WNVBUUQzk,12473
+docling/models/readingorder_model.py,sha256=bZoXHaSwUsa8niSmJrbCuy784ixCeBXT-RQBUfgHJ4A,14925
+docling/models/table_structure_model.py,sha256=RFXo73f2q4XuKyaSqbxpznh7JVtlLcT0FsOWl9oZbSg,12518
 docling/models/tesseract_ocr_cli_model.py,sha256=qcM3-n7Z_dm1CGBhVUcNr2XT41iXnU32zk4RqKHBl9I,12775
-docling/models/tesseract_ocr_model.py,sha256=9DPAE7XP7smej7HYhr7mdwpuxSjAcv_GPrYZG3bb1RA,10587
+docling/models/tesseract_ocr_model.py,sha256=GdI5Cjfi87qcehVbM3wdKRvKkl_F9A4bwTUbjXZCJYA,10745
 docling/models/factories/__init__.py,sha256=x_EM5dDg_A3HBcBYzOoqwmA2AFLtJ1IzYDPX-R1A-Sg,868
 docling/models/factories/base_factory.py,sha256=MfWIljMETi5aaVR-6qLTelW8u1gwDAQsOwg3fu7O4Qc,4028
 docling/models/factories/ocr_factory.py,sha256=G5RkmkKvkl-ihpo6qSj8WC77VdlVSQ1s0ekwUX2ILts,316
 docling/models/factories/picture_description_factory.py,sha256=Ru3-TnVVEKf5O07C_UpGf2HCOHc7j20AJzfficw3agM,385
 docling/models/plugins/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-docling/models/plugins/defaults.py,sha256=qslXGnRX07Z3GGttNriqaox0v0vXp4zs4KLurHCZjp4,858
+docling/models/plugins/defaults.py,sha256=OAHWW2tCcUXSyDMFxV_lXVRjSBJ1n6z-Eb3R8cDucU4,886
 docling/models/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling/models/utils/hf_model_download.py,sha256=scBEfsM4yl7xPzqe7UtPvDh9RfQZQnuOhqQKilYBHls,984
 docling/models/vlm_models_inline/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-docling/models/vlm_models_inline/hf_transformers_model.py,sha256=w9_N4ccjmYYK5yYQou0LSMGaj6gs8l0hULvXbkfYXSQ,7425
-docling/models/vlm_models_inline/mlx_model.py,sha256=qpyi6fGHm0vPqW2yeTsRBKOTTshNJ1LAPbH1SBDp8Y8,5784
+docling/models/vlm_models_inline/hf_transformers_model.py,sha256=LAnWFIHGblWln6DQMLtCQQW3-YUPDMbgeD2tjfM8vLM,8415
+docling/models/vlm_models_inline/mlx_model.py,sha256=p-H6wG31iVRoOjsqYaCVa4pEzxMP3vzLcsUatMjDJDQ,5948
 docling/pipeline/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling/pipeline/asr_pipeline.py,sha256=tQkhu9fXdkSuYIL22xzV2YRUlQh-9qktHBbs2qeXhJI,9070
 docling/pipeline/base_pipeline.py,sha256=14yQrDjsojl4RgbBjKFSEfVBYR_sULZfBI1uDzFLi8Y,9331
 docling/pipeline/simple_pipeline.py,sha256=TXZOwR7hZRji462ZTIpte0VJjzbxvNVE8dbLFANDhSU,2253
-docling/pipeline/standard_pdf_pipeline.py,sha256=2Hqg2wnAXfbZbLUOQrRus8PMEuZ549jR1mfR86-CAB4,12659
-docling/pipeline/vlm_pipeline.py,sha256=IrjDbajCPmUPep_jATKNiABST4tQ8mvpkQz9mtBQ8qQ,15279
+docling/pipeline/standard_pdf_pipeline.py,sha256=yFishq4Cu01BiBGHk3Irr7ogcTQKeSC0QZImQVAhIaY,12740
+docling/pipeline/vlm_pipeline.py,sha256=0lj8tbXNpYF8OLBoLqP2BZfFpTHi40RoHVfvO_Nah4Q,15349
 docling/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-docling/utils/accelerator_utils.py,sha256=Fww4UiTiuIB91iuPgUZTy-DYpCGRMI8YuCYKhFb0gjA,2905
+docling/utils/accelerator_utils.py,sha256=DSajLxVx1JEVT0zt5de26llciLNlVfIDfSa2zYCFJzQ,2909
 docling/utils/api_image_request.py,sha256=_CgdzmPqdsyXmyYUFGLZcXcoH586qC6A1p5vsNbj1Q0,1416
 docling/utils/export.py,sha256=VwVUnYDk3mhGmISDbVm306fwpGNnoojouStBD4UajXI,4673
 docling/utils/glm_utils.py,sha256=TKOWQqWAHsX_w4fvoAA7_2xCi_urhnp1DsmjY8_sk5w,12274
-docling/utils/layout_postprocessor.py,sha256=laTPGGj-hv16Zh1TRcn8NK0POKs7d3jeaV1pRR_TjIU,24228
+docling/utils/layout_postprocessor.py,sha256=QuTZZq4LNs1eM_n_2gubVfAuLBMkJiozfs3hp-jUpK4,24399
 docling/utils/locks.py,sha256=RzqQtD5UispgV71pGN_nU6GYfeN11BN0Sh_Dq9ycqGo,52
-docling/utils/model_downloader.py,sha256=6TDxFOvMRYT8JyYyaQS_wXMJzNga61ImY3sFdks66qM,4004
-docling/utils/ocr_utils.py,sha256=AOaDAHr5S74d-IRVR_LKhKynUTIurAwLJ3wNeY58gPA,2326
-docling/utils/orientation.py,sha256=xXlOfowL54FKwjsTFrM7y3ogk1wChLNn_-u74tYIf1s,2011
+docling/utils/model_downloader.py,sha256=3vijCsAIVwWqehGBDRxRq7mJ3yRb9-zBsG00iqjqegU,4076
+docling/utils/ocr_utils.py,sha256=nmresYyfin0raanpQc_GGeU3WoLsfExf6SEXNIQ7Djg,2325
+docling/utils/orientation.py,sha256=jTyLxyT31FlOodZoBMlADHNQK2lAWKYVs5z7pXd_6Cg,1842
 docling/utils/profiling.py,sha256=YaMGoB9MMZpagF9mb5ndoHj8Lpb9aIdb7El-Pl7IcFs,1753
 docling/utils/utils.py,sha256=kJtIYuzXeOyJHYlxmLAo7dGM5rEsDa1i84qEsUj1nio,1908
 docling/utils/visualization.py,sha256=tY2ylE2aiQKkmzlSLnFW-HTfFyqUUMguW18ldd1PLfo,2868
-docling-2.39.0.dist-info/licenses/LICENSE,sha256=mBb7ErEcM8VS9OhiGHnQ2kk75HwPhr54W1Oiz3965MY,1088
-docling-2.39.0.dist-info/METADATA,sha256=RhW8SMq3CPy0AIX67Fvkuk8CaF1slhq22Z-J78qGhHI,10273
-docling-2.39.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-docling-2.39.0.dist-info/entry_points.txt,sha256=hzVlbeE0aMSTQ9S0-NTYN0Hmgsn6qL_EA2qX4UbkAuY,149
-docling-2.39.0.dist-info/top_level.txt,sha256=vkIywP-USjFyYo1AIRQbWQQaL3xB5jf8vkCYdTIfNic,8
-docling-2.39.0.dist-info/RECORD,,
+docling-2.41.0.dist-info/licenses/LICENSE,sha256=mBb7ErEcM8VS9OhiGHnQ2kk75HwPhr54W1Oiz3965MY,1088
+docling-2.41.0.dist-info/METADATA,sha256=KYqB0miKX2x2ESNy8tNHdAlyTCONqhwGLR2iag2PcQ0,10274
+docling-2.41.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+docling-2.41.0.dist-info/entry_points.txt,sha256=hzVlbeE0aMSTQ9S0-NTYN0Hmgsn6qL_EA2qX4UbkAuY,149
+docling-2.41.0.dist-info/top_level.txt,sha256=vkIywP-USjFyYo1AIRQbWQQaL3xB5jf8vkCYdTIfNic,8
+docling-2.41.0.dist-info/RECORD,,