docling 1.6.2__py3-none-any.whl → 1.19.1__py3-none-any.whl

docling/datamodel/document.py

@@ -4,14 +4,16 @@ from pathlib import Path, PurePath
 from typing import ClassVar, Dict, Iterable, List, Optional, Tuple, Type, Union
 
 from docling_core.types import BaseCell, BaseText
-from docling_core.types import BoundingBox as DsBoundingBox
 from docling_core.types import Document as DsDocument
 from docling_core.types import DocumentDescription as DsDocumentDescription
 from docling_core.types import FileInfoObject as DsFileInfoObject
 from docling_core.types import PageDimensions, PageReference, Prov, Ref
 from docling_core.types import Table as DsSchemaTable
 from docling_core.types import TableCell
+from docling_core.types.doc.base import BoundingBox as DsBoundingBox
+from docling_core.types.doc.base import Figure
 from pydantic import BaseModel
+from typing_extensions import deprecated
 
 from docling.backend.abstract_backend import PdfDocumentBackend
 from docling.backend.docling_parse_backend import DoclingParseDocumentBackend
@@ -19,6 +21,7 @@ from docling.datamodel.base_models import (
     AssembledUnit,
     ConversionStatus,
     DocumentStream,
+    ErrorItem,
     FigureElement,
     Page,
     PageElement,
@@ -48,6 +51,15 @@ layout_label_to_ds_type = {
     "Text": "paragraph",
 }
 
+_EMPTY_DOC = DsDocument(
+    _name="",
+    description=DsDocumentDescription(logs=[]),
+    file_info=DsFileInfoObject(
+        filename="",
+        document_hash="",
+    ),
+)
+
 
 class InputDocument(BaseModel):
     file: PurePath = None
@@ -79,7 +91,9 @@ class InputDocument(BaseModel):
                     self.valid = False
                 else:
                     self.document_hash = create_file_hash(path_or_stream)
-                    self._backend = pdf_backend(path_or_stream=path_or_stream)
+                    self._backend = pdf_backend(
+                        path_or_stream=path_or_stream, document_hash=self.document_hash
+                    )
 
             elif isinstance(path_or_stream, BytesIO):
                 self.file = PurePath(filename)
@@ -89,7 +103,9 @@ class InputDocument(BaseModel):
                     self.valid = False
                 else:
                     self.document_hash = create_file_hash(path_or_stream)
-                    self._backend = pdf_backend(path_or_stream=path_or_stream)
+                    self._backend = pdf_backend(
+                        path_or_stream=path_or_stream, document_hash=self.document_hash
+                    )
 
             if self.document_hash and self._backend.page_count() > 0:
                 self.page_count = self._backend.page_count()
@@ -110,18 +126,19 @@ class InputDocument(BaseModel):
             # raise
 
 
+@deprecated("Use `ConversionResult` instead.")
 class ConvertedDocument(BaseModel):
     input: InputDocument
 
     status: ConversionStatus = ConversionStatus.PENDING  # failure, success
-    errors: List[Dict] = []  # structure to keep errors
+    errors: List[ErrorItem] = []  # structure to keep errors
 
     pages: List[Page] = []
-    assembled: Optional[AssembledUnit] = None
+    assembled: AssembledUnit = AssembledUnit()
 
-    output: Optional[DsDocument] = None
+    output: DsDocument = _EMPTY_DOC
 
-    def to_ds_document(self) -> DsDocument:
+    def _to_ds_document(self) -> DsDocument:
         title = ""
         desc = DsDocumentDescription(logs=[])
 
@@ -206,6 +223,8 @@ class ConvertedDocument(BaseModel):
                                 celltype = "col_header"
                             elif cell.row_header:
                                 celltype = "row_header"
+                            elif cell.row_section:
+                                celltype = "row_section"
 
                             def make_spans(cell):
                                 for rspan in range(
@@ -261,7 +280,7 @@ class ConvertedDocument(BaseModel):
                     ),
                 )
                 figures.append(
-                    BaseCell(
+                    Figure(
                         prov=[
                             Prov(
                                 bbox=target_bbox,
@@ -292,16 +311,91 @@ class ConvertedDocument(BaseModel):
         return ds_doc
 
     def render_as_dict(self):
-        if self.output:
-            return self.output.model_dump(by_alias=True, exclude_none=True)
-        else:
-            return {}
+        return self.output.model_dump(by_alias=True, exclude_none=True)
+
+    def render_as_markdown(
+        self,
+        delim: str = "\n\n",
+        main_text_start: int = 0,
+        main_text_stop: Optional[int] = None,
+        main_text_labels: list[str] = [
+            "title",
+            "subtitle-level-1",
+            "paragraph",
+            "caption",
+            "table",
+            "figure",
+        ],
+        strict_text: bool = False,
+        image_placeholder: str = "<!-- image -->",
+    ):
+        return self.output.export_to_markdown(
+            delim=delim,
+            main_text_start=main_text_start,
+            main_text_stop=main_text_stop,
+            main_text_labels=main_text_labels,
+            strict_text=strict_text,
+            image_placeholder=image_placeholder,
+        )
 
-    def render_as_markdown(self):
-        if self.output:
-            return self.output.export_to_markdown()
-        else:
-            return ""
+    def render_as_text(
+        self,
+        delim: str = "\n\n",
+        main_text_start: int = 0,
+        main_text_stop: Optional[int] = None,
+        main_text_labels: list[str] = [
+            "title",
+            "subtitle-level-1",
+            "paragraph",
+            "caption",
+        ],
+    ):
+        return self.output.export_to_markdown(
+            delim=delim,
+            main_text_start=main_text_start,
+            main_text_stop=main_text_stop,
+            main_text_labels=main_text_labels,
+            strict_text=True,
+        )
+
+    def render_as_doctags(
+        self,
+        delim: str = "\n\n",
+        main_text_start: int = 0,
+        main_text_stop: Optional[int] = None,
+        main_text_labels: list[str] = [
+            "title",
+            "subtitle-level-1",
+            "paragraph",
+            "caption",
+            "table",
+            "figure",
+        ],
+        xsize: int = 100,
+        ysize: int = 100,
+        add_location: bool = True,
+        add_content: bool = True,
+        add_page_index: bool = True,
+        # table specific flags
+        add_table_cell_location: bool = False,
+        add_table_cell_label: bool = True,
+        add_table_cell_text: bool = True,
+    ) -> str:
+        return self.output.export_to_document_tokens(
+            delim=delim,
+            main_text_start=main_text_start,
+            main_text_stop=main_text_stop,
+            main_text_labels=main_text_labels,
+            xsize=xsize,
+            ysize=ysize,
+            add_location=add_location,
+            add_content=add_content,
+            add_page_index=add_page_index,
+            # table specific flags
+            add_table_cell_location=add_table_cell_location,
+            add_table_cell_label=add_table_cell_label,
+            add_table_cell_text=add_table_cell_text,
+        )
 
     def render_element_images(
         self, element_types: Tuple[PageElement] = (FigureElement,)
@@ -318,6 +412,10 @@ class ConvertedDocument(BaseModel):
                 yield element, cropped_im
 
 
+class ConversionResult(ConvertedDocument):
+    pass
+
+
 class DocumentConversionInput(BaseModel):
 
     _path_or_stream_iterator: Iterable[Union[Path, DocumentStream]] = None

docling/datamodel/pipeline_options.py

@@ -0,0 +1,67 @@
+from enum import Enum, auto
+from typing import List, Literal, Optional, Union
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class TableFormerMode(str, Enum):
+    FAST = auto()
+    ACCURATE = auto()
+
+
+class TableStructureOptions(BaseModel):
+    do_cell_matching: bool = (
+        True
+        # True:  Matches predictions back to PDF cells. Can break table output if PDF cells
+        #        are merged across table columns.
+        # False: Let table structure model define the text cells, ignore PDF cells.
+    )
+    mode: TableFormerMode = TableFormerMode.FAST
+
+
+class OcrOptions(BaseModel):
+    kind: str
+
+
+class EasyOcrOptions(OcrOptions):
+    kind: Literal["easyocr"] = "easyocr"
+    lang: List[str] = ["fr", "de", "es", "en"]
+    use_gpu: bool = True  # same default as easyocr.Reader
+    model_storage_directory: Optional[str] = None
+    download_enabled: bool = True  # same default as easyocr.Reader
+
+    model_config = ConfigDict(
+        extra="forbid",
+        protected_namespaces=(),
+    )
+
+
+class TesseractCliOcrOptions(OcrOptions):
+    kind: Literal["tesseract"] = "tesseract"
+    lang: List[str] = ["fra", "deu", "spa", "eng"]
+    tesseract_cmd: str = "tesseract"
+    path: Optional[str] = None
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+
+
+class TesseractOcrOptions(OcrOptions):
+    kind: Literal["tesserocr"] = "tesserocr"
+    lang: List[str] = ["fra", "deu", "spa", "eng"]
+    path: Optional[str] = None
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+
+
+class PipelineOptions(BaseModel):
+    do_table_structure: bool = True  # True: perform table structure extraction
+    do_ocr: bool = True  # True: perform OCR, replace programmatic PDF text
+
+    table_structure_options: TableStructureOptions = TableStructureOptions()
+    ocr_options: Union[EasyOcrOptions, TesseractCliOcrOptions, TesseractOcrOptions] = (
+        Field(EasyOcrOptions(), discriminator="kind")
+    )

docling/document_converter.py

@@ -7,7 +7,6 @@ from pathlib import Path
 from typing import Iterable, Optional, Type, Union
 
 import requests
-from docling_core.types import Document
 from PIL import ImageDraw
 from pydantic import AnyHttpUrl, TypeAdapter, ValidationError
 
@@ -16,14 +15,16 @@ from docling.datamodel.base_models import (
     AssembledUnit,
     AssembleOptions,
     ConversionStatus,
+    DoclingComponentType,
+    ErrorItem,
     Page,
-    PipelineOptions,
 )
 from docling.datamodel.document import (
-    ConvertedDocument,
+    ConversionResult,
     DocumentConversionInput,
     InputDocument,
 )
+from docling.datamodel.pipeline_options import PipelineOptions
 from docling.datamodel.settings import settings
 from docling.models.ds_glm_model import GlmModel
 from docling.models.page_assemble_model import PageAssembleModel
@@ -66,12 +67,15 @@ class DocumentConverter:
         from huggingface_hub import snapshot_download
 
         download_path = snapshot_download(
-            repo_id="ds4sd/docling-models", force_download=force, local_dir=local_dir
+            repo_id="ds4sd/docling-models",
+            force_download=force,
+            local_dir=local_dir,
+            revision="v2.0.0",
         )
 
         return Path(download_path)
 
-    def convert(self, input: DocumentConversionInput) -> Iterable[ConvertedDocument]:
+    def convert(self, input: DocumentConversionInput) -> Iterable[ConversionResult]:
 
         for input_batch in chunkify(
             input.docs(pdf_backend=self.pdf_backend), settings.perf.doc_batch_size
@@ -84,9 +88,9 @@ class DocumentConverter:
             #   yield from pool.map(self.process_document, input_batch)
 
             # Note: Pdfium backend is not thread-safe, thread pool usage was disabled.
-            yield from map(self.process_document, input_batch)
+            yield from map(self._process_document, input_batch)
 
-    def convert_single(self, source: Path | AnyHttpUrl | str) -> Document:
+    def convert_single(self, source: Path | AnyHttpUrl | str) -> ConversionResult:
         """Convert a single document.
 
         Args:
@@ -97,7 +101,7 @@ class DocumentConverter:
             RuntimeError: If conversion fails.
 
         Returns:
-            Document: The converted document object.
+            ConversionResult: The conversion result object.
         """
         with tempfile.TemporaryDirectory() as temp_dir:
             try:
@@ -127,51 +131,49 @@ class DocumentConverter:
                         f"Unexpected file path type encountered: {type(source)}"
                     )
             conv_inp = DocumentConversionInput.from_paths(paths=[local_path])
-            converted_docs_iter = self.convert(conv_inp)
-            converted_doc: ConvertedDocument = next(converted_docs_iter)
-        if converted_doc.status not in {
+            conv_res_iter = self.convert(conv_inp)
+            conv_res: ConversionResult = next(conv_res_iter)
+        if conv_res.status not in {
             ConversionStatus.SUCCESS,
-            ConversionStatus.SUCCESS_WITH_ERRORS,
+            ConversionStatus.PARTIAL_SUCCESS,
         }:
-            raise RuntimeError(f"Conversion failed with status: {converted_doc.status}")
-        doc = converted_doc.to_ds_document()
-        return doc
+            raise RuntimeError(f"Conversion failed with status: {conv_res.status}")
+        return conv_res
 
-    def process_document(self, in_doc: InputDocument) -> ConvertedDocument:
+    def _process_document(self, in_doc: InputDocument) -> ConversionResult:
         start_doc_time = time.time()
-        converted_doc = ConvertedDocument(input=in_doc)
+        conv_res = ConversionResult(input=in_doc)
+
+        _log.info(f"Processing document {in_doc.file.name}")
 
         if not in_doc.valid:
-            converted_doc.status = ConversionStatus.FAILURE
-            return converted_doc
+            conv_res.status = ConversionStatus.FAILURE
+            return conv_res
 
         for i in range(0, in_doc.page_count):
-            converted_doc.pages.append(Page(page_no=i))
+            conv_res.pages.append(Page(page_no=i))
 
         all_assembled_pages = []
 
         try:
             # Iterate batches of pages (page_batch_size) in the doc
-            for page_batch in chunkify(
-                converted_doc.pages, settings.perf.page_batch_size
-            ):
-
+            for page_batch in chunkify(conv_res.pages, settings.perf.page_batch_size):
                 start_pb_time = time.time()
                 # Pipeline
 
                 # 1. Initialise the page resources
                 init_pages = map(
-                    functools.partial(self.initialize_page, in_doc), page_batch
+                    functools.partial(self._initialize_page, in_doc), page_batch
                 )
 
                 # 2. Populate page image
                 pages_with_images = map(
-                    functools.partial(self.populate_page_images, in_doc), init_pages
+                    functools.partial(self._populate_page_images, in_doc), init_pages
                 )
 
                 # 3. Populate programmatic page cells
                 pages_with_cells = map(
-                    functools.partial(self.parse_page_cells, in_doc),
+                    functools.partial(self._parse_page_cells, in_doc),
                     pages_with_images,
                 )
 
@@ -197,28 +199,45 @@ class DocumentConverter:
                 end_pb_time = time.time() - start_pb_time
                 _log.info(f"Finished converting page batch time={end_pb_time:.3f}")
 
-            # Free up mem resources of PDF backend
-            in_doc._backend.unload()
-
-            converted_doc.pages = all_assembled_pages
-            self.assemble_doc(converted_doc)
+            conv_res.pages = all_assembled_pages
+            self._assemble_doc(conv_res)
+
+            status = ConversionStatus.SUCCESS
+            for page in conv_res.pages:
+                if not page._backend.is_valid():
+                    conv_res.errors.append(
+                        ErrorItem(
+                            component_type=DoclingComponentType.PDF_BACKEND,
+                            module_name=type(page._backend).__name__,
+                            error_message=f"Page {page.page_no} failed to parse.",
+                        )
+                    )
+                    status = ConversionStatus.PARTIAL_SUCCESS
 
-            converted_doc.status = ConversionStatus.SUCCESS
+            conv_res.status = status
 
         except Exception as e:
-            converted_doc.status = ConversionStatus.FAILURE
+            conv_res.status = ConversionStatus.FAILURE
             trace = "\n".join(traceback.format_exception(e))
-            _log.info(f"Encountered an error during conversion: {trace}")
+            _log.info(
+                f"Encountered an error during conversion of document {in_doc.document_hash}:\n"
+                f"{trace}"
+            )
+
+        finally:
+            # Always unload the PDF backend, even in case of failure
+            if in_doc._backend:
+                in_doc._backend.unload()
 
         end_doc_time = time.time() - start_doc_time
         _log.info(
             f"Finished converting document time-pages={end_doc_time:.2f}/{in_doc.page_count}"
         )
 
-        return converted_doc
+        return conv_res
 
     # Initialise and load resources for a page, before downstream steps (populate images, cells, ...)
-    def initialize_page(self, doc: InputDocument, page: Page) -> Page:
+    def _initialize_page(self, doc: InputDocument, page: Page) -> Page:
         page._backend = doc._backend.load_page(page.page_no)
         page.size = page._backend.get_size()
         page.page_hash = create_hash(doc.document_hash + ":" + str(page.page_no))
@@ -226,9 +245,11 @@ class DocumentConverter:
         return page
 
     # Generate the page image and store it in the page object
-    def populate_page_images(self, doc: InputDocument, page: Page) -> Page:
+    def _populate_page_images(self, doc: InputDocument, page: Page) -> Page:
         # default scale
-        page.get_image(scale=1.0)
+        page.get_image(
+            scale=1.0
+        )  # puts the page image on the image cache at default scale
 
         # user requested scales
         if self.assemble_options.images_scale is not None:
@@ -240,7 +261,7 @@ class DocumentConverter:
         return page
 
     # Extract and populate the page cells and store it in the page object
-    def parse_page_cells(self, doc: InputDocument, page: Page) -> Page:
+    def _parse_page_cells(self, doc: InputDocument, page: Page) -> Page:
         page.cells = page._backend.get_text_cells()
 
         # DEBUG code:
@@ -255,12 +276,12 @@ class DocumentConverter:
 
         return page
 
-    def assemble_doc(self, converted_doc: ConvertedDocument):
+    def _assemble_doc(self, conv_res: ConversionResult):
         all_elements = []
         all_headers = []
         all_body = []
 
-        for p in converted_doc.pages:
+        for p in conv_res.pages:
 
             for el in p.assembled.body:
                 all_body.append(el)
@@ -269,8 +290,8 @@ class DocumentConverter:
             for el in p.assembled.elements:
                 all_elements.append(el)
 
-        converted_doc.assembled = AssembledUnit(
+        conv_res.assembled = AssembledUnit(
             elements=all_elements, headers=all_headers, body=all_body
         )
 
-        converted_doc.output = self.glm_model(converted_doc)
+        conv_res.output = self.glm_model(conv_res)

docling/models/base_ocr_model.py

@@ -3,21 +3,21 @@ import logging
 from abc import abstractmethod
 from typing import Iterable, List, Tuple
 
-import numpy
 import numpy as np
 from PIL import Image, ImageDraw
 from rtree import index
 from scipy.ndimage import find_objects, label
 
 from docling.datamodel.base_models import BoundingBox, CoordOrigin, OcrCell, Page
+from docling.datamodel.pipeline_options import OcrOptions
 
 _log = logging.getLogger(__name__)
 
 
 class BaseOcrModel:
-    def __init__(self, config):
-        self.config = config
-        self.enabled = config["enabled"]
+    def __init__(self, enabled: bool, options: OcrOptions):
+        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) -> Tuple[bool, List[BoundingBox]]:

docling/models/ds_glm_model.py

@@ -2,7 +2,7 @@ import copy
 import random
 
 from deepsearch_glm.nlp_utils import init_nlp_model
-from deepsearch_glm.utils.ds_utils import to_legacy_document_format
+from deepsearch_glm.utils.doc_utils import to_legacy_document_format
 from deepsearch_glm.utils.load_pretrained_models import load_pretrained_nlp_models
 from docling_core.types import BaseText
 from docling_core.types import Document as DsDocument
@@ -10,18 +10,22 @@ from docling_core.types import Ref
 from PIL import ImageDraw
 
 from docling.datamodel.base_models import BoundingBox, Cluster, CoordOrigin
-from docling.datamodel.document import ConvertedDocument
+from docling.datamodel.document import ConversionResult
 
 
 class GlmModel:
     def __init__(self, config):
         self.config = config
+        self.model_names = self.config.get(
+            "model_names", ""
+        )  # "language;term;reference"
         load_pretrained_nlp_models()
-        model = init_nlp_model(model_names="language;term;reference")
+        # model = init_nlp_model(model_names="language;term;reference")
+        model = init_nlp_model(model_names=self.model_names)
         self.model = model
 
-    def __call__(self, document: ConvertedDocument) -> DsDocument:
-        ds_doc = document.to_ds_document()
+    def __call__(self, conv_res: ConversionResult) -> DsDocument:
+        ds_doc = conv_res._to_ds_document()
         ds_doc_dict = ds_doc.model_dump(by_alias=True)
 
         glm_doc = self.model.apply_on_doc(ds_doc_dict)
@@ -34,7 +38,7 @@ class GlmModel:
         # DEBUG code:
         def draw_clusters_and_cells(ds_document, page_no):
             clusters_to_draw = []
-            image = copy.deepcopy(document.pages[page_no].image)
+            image = copy.deepcopy(conv_res.pages[page_no].image)
             for ix, elem in enumerate(ds_document.main_text):
                 if isinstance(elem, BaseText):
                     prov = elem.prov[0]
@@ -56,7 +60,7 @@ class GlmModel:
                             bbox=BoundingBox.from_tuple(
                                 coord=prov.bbox,
                                 origin=CoordOrigin.BOTTOMLEFT,
-                            ).to_top_left_origin(document.pages[page_no].size.height),
+                            ).to_top_left_origin(conv_res.pages[page_no].size.height),
                         )
                     )
 

docling/models/easyocr_model.py

@@ -4,21 +4,33 @@ from typing import Iterable
 import numpy
 
 from docling.datamodel.base_models import BoundingBox, CoordOrigin, OcrCell, Page
+from docling.datamodel.pipeline_options import EasyOcrOptions
 from docling.models.base_ocr_model import BaseOcrModel
 
 _log = logging.getLogger(__name__)
 
 
 class EasyOcrModel(BaseOcrModel):
-    def __init__(self, config):
-        super().__init__(config)
+    def __init__(self, enabled: bool, options: EasyOcrOptions):
+        super().__init__(enabled=enabled, options=options)
+        self.options: EasyOcrOptions
 
         self.scale = 3  # multiplier for 72 dpi == 216 dpi.
 
         if self.enabled:
-            import easyocr
+            try:
+                import easyocr
+            except ImportError:
+                raise ImportError(
+                    "EasyOCR is not installed. Please install it via `pip install easyocr` to use this OCR engine. "
+                    "Alternatively, Docling has support for other OCR engines. See the documentation."
+                )
 
-            self.reader = easyocr.Reader(config["lang"])
+            self.reader = easyocr.Reader(
+                lang_list=self.options.lang,
+                model_storage_directory=self.options.model_storage_directory,
+                download_enabled=self.options.download_enabled,
+            )
 
     def __call__(self, page_batch: Iterable[Page]) -> Iterable[Page]:
 
@@ -31,6 +43,9 @@ class EasyOcrModel(BaseOcrModel):
 
             all_ocr_cells = []
             for ocr_rect in ocr_rects:
+                # Skip zero area boxes
+                if ocr_rect.area() == 0:
+                    continue
                 high_res_image = page._backend.get_page_image(
                     scale=self.scale, cropbox=ocr_rect
                 )

docling/models/layout_model.py

@@ -33,6 +33,7 @@ class LayoutModel:
         "Page-footer",
         "Code",
         "List-item",
+        # "Title"
         # "Formula",
     ]
     PAGE_HEADER_LABELS = ["Page-header", "Page-footer"]
@@ -69,9 +70,7 @@ class LayoutModel:
             "Key-Value Region": 0.45,
         }
 
-        CLASS_REMAPPINGS = {
-            "Document Index": "Table",
-        }
+        CLASS_REMAPPINGS = {"Document Index": "Table", "Title": "Section-header"}
 
         _log.debug("================= Start postprocess function ====================")
         start_time = time.time()
@@ -277,6 +276,7 @@ class LayoutModel:
                     bbox=BoundingBox.model_validate(pred_item),
                     cells=[],
                 )
+
                 clusters.append(cluster)
 
             # Map cells to clusters