docling 0.1.0__py3-none-any.whl

docling/backend/abstract_backend.py

@@ -0,0 +1,55 @@
+from abc import ABC, abstractmethod
+from io import BytesIO
+from pathlib import Path
+from typing import Any, Iterable, Optional, Union
+
+from PIL import Image
+
+
+class PdfPageBackend(ABC):
+    def __init__(self, page_obj: Any) -> object:
+        pass
+
+    @abstractmethod
+    def get_text_in_rect(self, bbox: "BoundingBox") -> str:
+        pass
+
+    @abstractmethod
+    def get_text_cells(self) -> Iterable["Cell"]:
+        pass
+
+    @abstractmethod
+    def get_page_image(
+        self, scale: int = 1, cropbox: Optional["BoundingBox"] = None
+    ) -> Image.Image:
+        pass
+
+    @abstractmethod
+    def get_size(self) -> "PageSize":
+        pass
+
+    @abstractmethod
+    def unload(self):
+        pass
+
+
+class PdfDocumentBackend(ABC):
+    @abstractmethod
+    def __init__(self, path_or_stream: Iterable[Union[BytesIO, Path]]):
+        pass
+
+    @abstractmethod
+    def load_page(self, page_no: int) -> PdfPageBackend:
+        pass
+
+    @abstractmethod
+    def page_count(self) -> int:
+        pass
+
+    @abstractmethod
+    def is_valid(self) -> bool:
+        pass
+
+    @abstractmethod
+    def unload(self):
+        pass

docling/backend/pypdfium2_backend.py

@@ -0,0 +1,223 @@
+import random
+from io import BytesIO
+from pathlib import Path
+from typing import Iterable, List, Optional, Union
+
+import pypdfium2 as pdfium
+from PIL import Image, ImageDraw
+from pypdfium2 import PdfPage
+
+from docling.backend.abstract_backend import PdfDocumentBackend, PdfPageBackend
+from docling.datamodel.base_models import BoundingBox, Cell, CoordOrigin, PageSize
+
+
+class PyPdfiumPageBackend(PdfPageBackend):
+    def __init__(self, page_obj: PdfPage):
+        super().__init__(page_obj)
+        self._ppage = page_obj
+        self.text_page = None
+
+    def get_text_in_rect(self, bbox: BoundingBox) -> str:
+        if not self.text_page:
+            self.text_page = self._ppage.get_textpage()
+
+        if bbox.coord_origin != CoordOrigin.BOTTOMLEFT:
+            bbox = bbox.to_bottom_left_origin(self.get_size().height)
+
+        text_piece = self.text_page.get_text_bounded(*bbox.as_tuple())
+
+        return text_piece
+
+    def get_text_cells(self) -> Iterable[Cell]:
+        if not self.text_page:
+            self.text_page = self._ppage.get_textpage()
+
+        cells = []
+        cell_counter = 0
+
+        page_size = self.get_size()
+
+        for i in range(self.text_page.count_rects()):
+            rect = self.text_page.get_rect(i)
+            text_piece = self.text_page.get_text_bounded(*rect)
+            x0, y0, x1, y1 = rect
+            cells.append(
+                Cell(
+                    id=cell_counter,
+                    text=text_piece,
+                    bbox=BoundingBox(
+                        l=x0, b=y0, r=x1, t=y1, coord_origin=CoordOrigin.BOTTOMLEFT
+                    ).to_top_left_origin(page_size.height),
+                )
+            )
+            cell_counter += 1
+
+        # PyPdfium2 produces very fragmented cells, with sub-word level boundaries, in many PDFs.
+        # The cell merging code below is to clean this up.
+        def merge_horizontal_cells(
+            cells: List[Cell],
+            horizontal_threshold_factor: float = 1.0,
+            vertical_threshold_factor: float = 0.5,
+        ) -> List[Cell]:
+            if not cells:
+                return []
+
+            def group_rows(cells: List[Cell]) -> List[List[Cell]]:
+                rows = []
+                current_row = [cells[0]]
+                row_top = cells[0].bbox.t
+                row_bottom = cells[0].bbox.b
+                row_height = cells[0].bbox.height
+
+                for cell in cells[1:]:
+                    vertical_threshold = row_height * vertical_threshold_factor
+                    if (
+                        abs(cell.bbox.t - row_top) <= vertical_threshold
+                        and abs(cell.bbox.b - row_bottom) <= vertical_threshold
+                    ):
+                        current_row.append(cell)
+                        row_top = min(row_top, cell.bbox.t)
+                        row_bottom = max(row_bottom, cell.bbox.b)
+                        row_height = row_bottom - row_top
+                    else:
+                        rows.append(current_row)
+                        current_row = [cell]
+                        row_top = cell.bbox.t
+                        row_bottom = cell.bbox.b
+                        row_height = cell.bbox.height
+
+                if current_row:
+                    rows.append(current_row)
+
+                return rows
+
+            def merge_row(row: List[Cell]) -> List[Cell]:
+                merged = []
+                current_group = [row[0]]
+
+                for cell in row[1:]:
+                    prev_cell = current_group[-1]
+                    avg_height = (prev_cell.bbox.height + cell.bbox.height) / 2
+                    if (
+                        cell.bbox.l - prev_cell.bbox.r
+                        <= avg_height * horizontal_threshold_factor
+                    ):
+                        current_group.append(cell)
+                    else:
+                        merged.append(merge_group(current_group))
+                        current_group = [cell]
+
+                if current_group:
+                    merged.append(merge_group(current_group))
+
+                return merged
+
+            def merge_group(group: List[Cell]) -> Cell:
+                if len(group) == 1:
+                    return group[0]
+
+                merged_text = "".join(cell.text for cell in group)
+                merged_bbox = BoundingBox(
+                    l=min(cell.bbox.l for cell in group),
+                    t=min(cell.bbox.t for cell in group),
+                    r=max(cell.bbox.r for cell in group),
+                    b=max(cell.bbox.b for cell in group),
+                )
+                return Cell(id=group[0].id, text=merged_text, bbox=merged_bbox)
+
+            rows = group_rows(cells)
+            merged_cells = [cell for row in rows for cell in merge_row(row)]
+
+            for i, cell in enumerate(merged_cells, 1):
+                cell.id = i
+
+            return merged_cells
+
+        def draw_clusters_and_cells():
+            image = self.get_page_image()
+            draw = ImageDraw.Draw(image)
+            for c in cells:
+                x0, y0, x1, y1 = c.bbox.as_tuple()
+                cell_color = (
+                    random.randint(30, 140),
+                    random.randint(30, 140),
+                    random.randint(30, 140),
+                )
+                draw.rectangle([(x0, y0), (x1, y1)], outline=cell_color)
+            image.show()
+
+        # before merge:
+        # draw_clusters_and_cells()
+
+        cells = merge_horizontal_cells(cells)
+
+        # after merge:
+        # draw_clusters_and_cells()
+
+        return cells
+
+    def get_page_image(
+        self, scale: int = 1, cropbox: Optional[BoundingBox] = None
+    ) -> Image.Image:
+
+        page_size = self.get_size()
+
+        if not cropbox:
+            cropbox = BoundingBox(
+                l=0,
+                r=page_size.width,
+                t=0,
+                b=page_size.height,
+                coord_origin=CoordOrigin.TOPLEFT,
+            )
+            padbox = BoundingBox(
+                l=0, r=0, t=0, b=0, coord_origin=CoordOrigin.BOTTOMLEFT
+            )
+        else:
+            padbox = cropbox.to_bottom_left_origin(page_size.height)
+            padbox.r = page_size.width - padbox.r
+            padbox.t = page_size.height - padbox.t
+
+        image = (
+            self._ppage.render(
+                scale=scale * 1.5,
+                rotation=0,  # no additional rotation
+                crop=padbox.as_tuple(),
+            )
+            .to_pil()
+            .resize(size=(round(cropbox.width * scale), round(cropbox.height * scale)))
+        )  # We resize the image from 1.5x the given scale to make it sharper.
+
+        return image
+
+    def get_size(self) -> PageSize:
+        return PageSize(width=self._ppage.get_width(), height=self._ppage.get_height())
+
+    def unload(self):
+        self._ppage = None
+        self.text_page = None
+
+
+class PyPdfiumDocumentBackend(PdfDocumentBackend):
+    def __init__(self, path_or_stream: Iterable[Union[BytesIO, Path]]):
+        super().__init__(path_or_stream)
+
+        if isinstance(path_or_stream, Path):
+            self._pdoc = pdfium.PdfDocument(path_or_stream)
+        elif isinstance(path_or_stream, BytesIO):
+            self._pdoc = pdfium.PdfDocument(
+                path_or_stream
+            )  # TODO Fix me, won't accept bytes.
+
+    def page_count(self) -> int:
+        return len(self._pdoc)
+
+    def load_page(self, page_no: int) -> PdfPage:
+        return PyPdfiumPageBackend(self._pdoc[page_no])
+
+    def is_valid(self) -> bool:
+        return self.page_count() > 0
+
+    def unload(self):
+        self._pdoc.close()
+        self._pdoc = None

docling/datamodel/base_models.py

@@ -0,0 +1,247 @@
+from enum import Enum, auto
+from io import BytesIO
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from PIL.Image import Image
+from pydantic import BaseModel, ConfigDict, model_validator
+
+from docling.backend.abstract_backend import PdfPageBackend
+
+
+class ConversionStatus(str, Enum):
+    PENDING = auto()
+    STARTED = auto()
+    FAILURE = auto()
+    SUCCESS = auto()
+    SUCCESS_WITH_ERRORS = auto()
+
+
+class DocInputType(str, Enum):
+    PATH = auto()
+    STREAM = auto()
+
+
+class CoordOrigin(str, Enum):
+    TOPLEFT = auto()
+    BOTTOMLEFT = auto()
+
+
+class PageSize(BaseModel):
+    width: float = 0.0
+    height: float = 0.0
+
+
+class BoundingBox(BaseModel):
+    l: float  # left
+    t: float  # top
+    r: float  # right
+    b: float  # bottom
+
+    coord_origin: CoordOrigin = CoordOrigin.TOPLEFT
+
+    @property
+    def width(self):
+        return self.r - self.l
+
+    @property
+    def height(self):
+        return abs(self.t - self.b)
+
+    def as_tuple(self):
+        if self.coord_origin == CoordOrigin.TOPLEFT:
+            return (self.l, self.t, self.r, self.b)
+        elif self.coord_origin == CoordOrigin.BOTTOMLEFT:
+            return (self.l, self.b, self.r, self.t)
+
+    @classmethod
+    def from_tuple(cls, coord: Tuple[float], origin: CoordOrigin):
+        if origin == CoordOrigin.TOPLEFT:
+            return BoundingBox(
+                l=coord[0], t=coord[1], r=coord[2], b=coord[3], coord_origin=origin
+            )
+        elif origin == CoordOrigin.BOTTOMLEFT:
+            return BoundingBox(
+                l=coord[0], b=coord[1], r=coord[2], t=coord[3], coord_origin=origin
+            )
+
+    def area(self) -> float:
+        return (self.r - self.l) * (self.b - self.t)
+
+    def intersection_area_with(self, other: "BoundingBox") -> float:
+        # Calculate intersection coordinates
+        left = max(self.l, other.l)
+        top = max(self.t, other.t)
+        right = min(self.r, other.r)
+        bottom = min(self.b, other.b)
+
+        # Calculate intersection dimensions
+        width = right - left
+        height = bottom - top
+
+        # If the bounding boxes do not overlap, width or height will be negative
+        if width <= 0 or height <= 0:
+            return 0.0
+
+        return width * height
+
+    def to_bottom_left_origin(self, page_height) -> "BoundingBox":
+        if self.coord_origin == CoordOrigin.BOTTOMLEFT:
+            return self
+        elif self.coord_origin == CoordOrigin.TOPLEFT:
+            return BoundingBox(
+                l=self.l,
+                r=self.r,
+                t=page_height - self.t,
+                b=page_height - self.b,
+                coord_origin=CoordOrigin.BOTTOMLEFT,
+            )
+
+    def to_top_left_origin(self, page_height):
+        if self.coord_origin == CoordOrigin.TOPLEFT:
+            return self
+        elif self.coord_origin == CoordOrigin.BOTTOMLEFT:
+            return BoundingBox(
+                l=self.l,
+                r=self.r,
+                t=page_height - self.t,  # self.b
+                b=page_height - self.b,  # self.t
+                coord_origin=CoordOrigin.TOPLEFT,
+            )
+
+
+class Cell(BaseModel):
+    id: int
+    text: str
+    bbox: BoundingBox
+
+
+class OcrCell(Cell):
+    confidence: float
+
+
+class Cluster(BaseModel):
+    id: int
+    label: str
+    bbox: BoundingBox
+    confidence: float = 1.0
+    cells: List[Cell] = []
+
+
+class BasePageElement(BaseModel):
+    label: str
+    id: int
+    page_no: int
+    cluster: Cluster
+    text: Optional[str] = None
+
+
+class LayoutPrediction(BaseModel):
+    clusters: List[Cluster] = []
+
+
+class TableCell(BaseModel):
+    bbox: BoundingBox
+    row_span: int
+    col_span: int
+    start_row_offset_idx: int
+    end_row_offset_idx: int
+    start_col_offset_idx: int
+    end_col_offset_idx: int
+    text: str
+    column_header: bool = False
+    row_header: bool = False
+    row_section: bool = False
+
+    @model_validator(mode="before")
+    @classmethod
+    def from_dict_format(cls, data: Any) -> Any:
+        if isinstance(data, Dict):
+            text = data["bbox"].get("token", "")
+            if not len(text):
+                text_cells = data.pop("text_cell_bboxes", None)
+                if text_cells:
+                    for el in text_cells:
+                        text += el["token"] + " "
+
+                text = text.strip()
+            data["text"] = text
+
+        return data
+
+
+class TableElement(BasePageElement):
+    otsl_seq: List[str]
+    num_rows: int = 0
+    num_cols: int = 0
+    table_cells: List[TableCell]
+
+
+class TableStructurePrediction(BaseModel):
+    table_map: Dict[int, TableElement] = {}
+
+
+class TextElement(BasePageElement):
+    ...
+
+
+class FigureData(BaseModel):
+    pass
+
+
+class FigureElement(BasePageElement):
+    data: Optional[FigureData] = None
+    provenance: Optional[str] = None
+    predicted_class: Optional[str] = None
+    confidence: Optional[float] = None
+
+
+class FigureClassificationPrediction(BaseModel):
+    figure_count: int = 0
+    figure_map: Dict[int, FigureElement] = {}
+
+
+class EquationPrediction(BaseModel):
+    equation_count: int = 0
+    equation_map: Dict[int, TextElement] = {}
+
+
+class PagePredictions(BaseModel):
+    layout: LayoutPrediction = None
+    tablestructure: TableStructurePrediction = None
+    figures_classification: FigureClassificationPrediction = None
+    equations_prediction: EquationPrediction = None
+
+
+PageElement = Union[TextElement, TableElement, FigureElement]
+
+
+class AssembledUnit(BaseModel):
+    elements: List[PageElement]
+    body: List[PageElement]
+    headers: List[PageElement]
+
+
+class Page(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    page_no: int
+    page_hash: str = None
+    size: PageSize = None
+    image: Image = None
+    cells: List[Cell] = None
+    predictions: PagePredictions = PagePredictions()
+    assembled: AssembledUnit = None
+
+    _backend: PdfPageBackend = None  # Internal PDF backend
+
+
+class DocumentStream(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    filename: str
+    stream: BytesIO
+
+
+class PipelineOptions(BaseModel):
+    do_table_structure: bool = True
+    do_ocr: bool = False