docling-core 2.7.1__tar.gz → 2.9.0__tar.gz

{docling_core-2.7.1 → docling_core-2.9.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: docling-core
-Version: 2.7.1
+Version: 2.9.0
 Summary: A python library to define and validate data types in Docling.
 Home-page: https://ds4sd.github.io/
 License: MIT
@@ -25,13 +25,16 @@ Classifier: Topic :: Database
 Classifier: Topic :: Scientific/Engineering :: Information Analysis
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Classifier: Typing :: Typed
+Provides-Extra: chunking
 Requires-Dist: jsonref (>=1.1.0,<2.0.0)
 Requires-Dist: jsonschema (>=4.16.0,<5.0.0)
 Requires-Dist: pandas (>=2.1.4,<3.0.0)
 Requires-Dist: pillow (>=10.3.0,<11.0.0)
 Requires-Dist: pydantic (>=2.6.0,<3.0.0,!=2.10.0,!=2.10.1,!=2.10.2)
 Requires-Dist: pyyaml (>=5.1,<7.0.0)
+Requires-Dist: semchunk (>=2.2.0,<3.0.0) ; extra == "chunking"
 Requires-Dist: tabulate (>=0.9.0,<0.10.0)
+Requires-Dist: transformers (>=4.34.0,<5.0.0) ; extra == "chunking"
 Requires-Dist: typing-extensions (>=4.12.2,<5.0.0)
 Project-URL: Repository, https://github.com/DS4SD/docling-core
 Description-Content-Type: text/markdown

{docling_core-2.7.1 → docling_core-2.9.0}/docling_core/transforms/chunker/__init__.py

@@ -7,6 +7,7 @@
 
 from docling_core.transforms.chunker.base import BaseChunk, BaseChunker, BaseMeta
 from docling_core.transforms.chunker.hierarchical_chunker import (
+    DocChunk,
     DocMeta,
     HierarchicalChunker,
 )

{docling_core-2.7.1 → docling_core-2.9.0}/docling_core/transforms/chunker/base.py

@@ -4,6 +4,7 @@
 #
 
 """Define base classes for chunking."""
+import json
 from abc import ABC, abstractmethod
 from typing import Any, ClassVar, Iterator
 
@@ -11,6 +12,8 @@ from pydantic import BaseModel
 
 from docling_core.types.doc import DoclingDocument as DLDocument
 
+DFLT_DELIM = "\n"
+
 
 class BaseMeta(BaseModel):
     """Chunk metadata base class."""
@@ -45,6 +48,8 @@ class BaseChunk(BaseModel):
 class BaseChunker(BaseModel, ABC):
     """Chunker base class."""
 
+    delim: str = DFLT_DELIM
+
     @abstractmethod
     def chunk(self, dl_doc: DLDocument, **kwargs) -> Iterator[BaseChunk]:
         """Chunk the provided document.
@@ -59,3 +64,32 @@ class BaseChunker(BaseModel, ABC):
             Iterator[BaseChunk]: iterator over extracted chunks
         """
         raise NotImplementedError()
+
+    def serialize(self, chunk: BaseChunk) -> str:
+        """Serialize the given chunk. This base implementation is embedding-targeted.
+
+        Args:
+            chunk: chunk to serialize
+
+        Returns:
+            str: the serialized form of the chunk
+        """
+        meta = chunk.meta.export_json_dict()
+
+        items = []
+        for k in meta:
+            if k not in chunk.meta.excluded_embed:
+                if isinstance(meta[k], list):
+                    items.append(
+                        self.delim.join(
+                            [
+                                d if isinstance(d, str) else json.dumps(d)
+                                for d in meta[k]
+                            ]
+                        )
+                    )
+                else:
+                    items.append(json.dumps(meta[k]))
+        items.append(chunk.text)
+
+        return self.delim.join(items)

{docling_core-2.7.1 → docling_core-2.9.0}/docling_core/transforms/chunker/hierarchical_chunker.py

@@ -104,7 +104,7 @@ class DocMeta(BaseMeta):
 
 
 class DocChunk(BaseChunk):
-    """Data model for Hierarchical Chunker chunks."""
+    """Data model for document chunks."""
 
     meta: DocMeta
 
@@ -119,7 +119,6 @@ class HierarchicalChunker(BaseChunker):
     """
 
     merge_list_items: bool = True
-    delim: str = "\n"
 
     @classmethod
     def _triplet_serialize(cls, table_df: DataFrame) -> str:

docling_core-2.9.0/docling_core/transforms/chunker/hybrid_chunker.py

@@ -0,0 +1,272 @@
+#
+# Copyright IBM Corp. 2024 - 2024
+# SPDX-License-Identifier: MIT
+#
+
+"""Hybrid chunker implementation leveraging both doc structure & token awareness."""
+
+import warnings
+from typing import Iterable, Iterator, Optional, Union
+
+from pydantic import BaseModel, ConfigDict, PositiveInt, TypeAdapter, model_validator
+from typing_extensions import Self
+
+try:
+    import semchunk
+    from transformers import AutoTokenizer, PreTrainedTokenizerBase
+except ImportError:
+    raise RuntimeError(
+        "Module requires 'chunking' extra; to install, run: "
+        "`pip install 'docling-core[chunking]'`"
+    )
+
+from docling_core.transforms.chunker import (
+    BaseChunk,
+    BaseChunker,
+    DocChunk,
+    DocMeta,
+    HierarchicalChunker,
+)
+from docling_core.types import DoclingDocument
+from docling_core.types.doc.document import TextItem
+
+
+class HybridChunker(BaseChunker):
+    r"""Chunker doing tokenization-aware refinements on top of document layout chunking.
+
+    Args:
+        tokenizer: The tokenizer to use; either instantiated object or name or path of
+            respective pretrained model
+        max_tokens: The maximum number of tokens per chunk. If not set, limit is
+            resolved from the tokenizer
+        merge_peers: Whether to merge undersized chunks sharing same relevant metadata
+    """
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    tokenizer: Union[PreTrainedTokenizerBase, str]
+    max_tokens: int = None  # type: ignore[assignment]
+    merge_peers: bool = True
+
+    _inner_chunker: HierarchicalChunker = HierarchicalChunker()
+
+    @model_validator(mode="after")
+    def _patch_tokenizer_and_max_tokens(self) -> Self:
+        self._tokenizer = (
+            self.tokenizer
+            if isinstance(self.tokenizer, PreTrainedTokenizerBase)
+            else AutoTokenizer.from_pretrained(self.tokenizer)
+        )
+        if self.max_tokens is None:
+            self.max_tokens = TypeAdapter(PositiveInt).validate_python(
+                self._tokenizer.model_max_length
+            )
+        return self
+
+    def _count_tokens(self, text: Optional[Union[str, list[str]]]):
+        if text is None:
+            return 0
+        elif isinstance(text, list):
+            total = 0
+            for t in text:
+                total += self._count_tokens(t)
+            return total
+        return len(self._tokenizer.tokenize(text, max_length=None))
+
+    class _ChunkLengthInfo(BaseModel):
+        total_len: int
+        text_len: int
+        other_len: int
+
+    def _doc_chunk_length(self, doc_chunk: DocChunk):
+        text_length = self._count_tokens(doc_chunk.text)
+        headings_length = self._count_tokens(doc_chunk.meta.headings)
+        captions_length = self._count_tokens(doc_chunk.meta.captions)
+        total = text_length + headings_length + captions_length
+        return self._ChunkLengthInfo(
+            total_len=total,
+            text_len=text_length,
+            other_len=total - text_length,
+        )
+
+    def _make_chunk_from_doc_items(
+        self, doc_chunk: DocChunk, window_text: str, window_start: int, window_end: int
+    ):
+        meta = DocMeta(
+            doc_items=doc_chunk.meta.doc_items[window_start : window_end + 1],
+            headings=doc_chunk.meta.headings,
+            captions=doc_chunk.meta.captions,
+        )
+        new_chunk = DocChunk(text=window_text, meta=meta)
+        return new_chunk
+
+    def _merge_text(self, t1, t2):
+        if t1 == "":
+            return t2
+        elif t2 == "":
+            return t1
+        else:
+            return f"{t1}{self.delim}{t2}"
+
+    def _split_by_doc_items(self, doc_chunk: DocChunk) -> list[DocChunk]:
+        if doc_chunk.meta.doc_items is None or len(doc_chunk.meta.doc_items) <= 1:
+            return [doc_chunk]
+        length = self._doc_chunk_length(doc_chunk)
+        if length.total_len <= self.max_tokens:
+            return [doc_chunk]
+        else:
+            chunks = []
+            window_start = 0
+            window_end = 0
+            window_text = ""
+            window_text_length = 0
+            other_length = length.other_len
+            num_items = len(doc_chunk.meta.doc_items)
+            while window_end < num_items:
+                doc_item = doc_chunk.meta.doc_items[window_end]
+                if isinstance(doc_item, TextItem):
+                    text = doc_item.text
+                else:
+                    raise RuntimeError("Non-TextItem split not implemented yet")
+                text_length = self._count_tokens(text)
+                if (
+                    text_length + window_text_length + other_length < self.max_tokens
+                    and window_end < num_items - 1
+                ):
+                    # Still room left to add more to this chunk AND still at least one
+                    # item left
+                    window_end += 1
+                    window_text_length += text_length
+                    window_text = self._merge_text(window_text, text)
+                elif text_length + window_text_length + other_length < self.max_tokens:
+                    # All the items in the window fit into the chunk and there are no
+                    # other items left
+                    window_text = self._merge_text(window_text, text)
+                    new_chunk = self._make_chunk_from_doc_items(
+                        doc_chunk, window_text, window_start, window_end
+                    )
+                    chunks.append(new_chunk)
+                    window_end = num_items
+                elif window_start == window_end:
+                    # Only one item in the window and it doesn't fit into the chunk. So
+                    # we'll just make it a chunk for now and it will get split in the
+                    # plain text splitter.
+                    window_text = self._merge_text(window_text, text)
+                    new_chunk = self._make_chunk_from_doc_items(
+                        doc_chunk, window_text, window_start, window_end
+                    )
+                    chunks.append(new_chunk)
+                    window_start = window_end + 1
+                    window_end = window_start
+                    window_text = ""
+                    window_text_length = 0
+                else:
+                    # Multiple items in the window but they don't fit into the chunk.
+                    # However, the existing items must have fit or we wouldn't have
+                    # gotten here. So we put everything but the last item into the chunk
+                    # and then start a new window INCLUDING the current window end.
+                    new_chunk = self._make_chunk_from_doc_items(
+                        doc_chunk, window_text, window_start, window_end - 1
+                    )
+                    chunks.append(new_chunk)
+                    window_start = window_end
+                    window_text = ""
+                    window_text_length = 0
+            return chunks
+
+    def _split_using_plain_text(
+        self,
+        doc_chunk: DocChunk,
+    ) -> list[DocChunk]:
+        lengths = self._doc_chunk_length(doc_chunk)
+        if lengths.total_len <= self.max_tokens:
+            return [DocChunk(**doc_chunk.export_json_dict())]
+        else:
+            # How much room is there for text after subtracting out the headers and
+            # captions:
+            available_length = self.max_tokens - lengths.other_len
+            sem_chunker = semchunk.chunkerify(
+                self._tokenizer, chunk_size=available_length
+            )
+            if available_length <= 0:
+                warnings.warn(
+                    f"Headers and captions for this chunk are longer than the total amount of size for the chunk, chunk will be ignored: {doc_chunk.text=}"  # noqa
+                )
+                return []
+            text = doc_chunk.text
+            segments = sem_chunker.chunk(text)
+            chunks = [DocChunk(text=s, meta=doc_chunk.meta) for s in segments]
+            return chunks
+
+    def _merge_chunks_with_matching_metadata(self, chunks: list[DocChunk]):
+        output_chunks = []
+        window_start = 0
+        window_end = 0
+        num_chunks = len(chunks)
+        while window_end < num_chunks:
+            chunk = chunks[window_end]
+            lengths = self._doc_chunk_length(chunk)
+            headings_and_captions = (chunk.meta.headings, chunk.meta.captions)
+            ready_to_append = False
+            if window_start == window_end:
+                # starting a new block of chunks to potentially merge
+                current_headings_and_captions = headings_and_captions
+                window_text = chunk.text
+                window_other_length = lengths.other_len
+                window_text_length = lengths.text_len
+                window_items = chunk.meta.doc_items
+                window_end += 1
+                first_chunk_of_window = chunk
+            elif (
+                headings_and_captions == current_headings_and_captions
+                and window_text_length + window_other_length + lengths.text_len
+                <= self.max_tokens
+            ):
+                # there is room to include the new chunk so add it to the window and
+                # continue
+                window_text = self._merge_text(window_text, chunk.text)
+                window_text_length += lengths.text_len
+                window_items = window_items + chunk.meta.doc_items
+                window_end += 1
+            else:
+                ready_to_append = True
+
+            if ready_to_append or window_end == num_chunks:
+                # no more room OR the start of new metadata.  Either way, end the block
+                # and use the current window_end as the start of a new block
+                if window_start + 1 == window_end:
+                    # just one chunk so use it as is
+                    output_chunks.append(first_chunk_of_window)
+                else:
+                    new_meta = DocMeta(
+                        doc_items=window_items,
+                        headings=current_headings_and_captions[0],
+                        captions=current_headings_and_captions[1],
+                    )
+                    new_chunk = DocChunk(
+                        text=window_text,
+                        meta=new_meta,
+                    )
+                    output_chunks.append(new_chunk)
+                # no need to reset window_text, etc. because that will be reset in the
+                # next iteration in the if window_start == window_end block
+                window_start = window_end
+
+        return output_chunks
+
+    def chunk(self, dl_doc: DoclingDocument, **kwargs) -> Iterator[BaseChunk]:
+        r"""Chunk the provided document.
+
+        Args:
+            dl_doc (DLDocument): document to chunk
+
+        Yields:
+            Iterator[Chunk]: iterator over extracted chunks
+        """
+        res: Iterable[DocChunk]
+        res = self._inner_chunker.chunk(dl_doc=dl_doc, **kwargs)  # type: ignore
+        res = [x for c in res for x in self._split_by_doc_items(c)]
+        res = [x for c in res for x in self._split_using_plain_text(c)]
+        if self.merge_peers:
+            res = self._merge_chunks_with_matching_metadata(res)
+        return iter(res)

{docling_core-2.7.1 → docling_core-2.9.0}/docling_core/types/doc/document.py

@@ -380,6 +380,7 @@ class DocumentOrigin(BaseModel):
         "application/vnd.openxmlformats-officedocument.presentationml.template",
         "application/vnd.openxmlformats-officedocument.presentationml.slideshow",
         "application/vnd.openxmlformats-officedocument.presentationml.presentation",
+        "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
         "text/asciidoc",
         "text/markdown",
     ]
@@ -445,7 +446,7 @@ class ImageRef(BaseModel):
     mimetype: str
     dpi: int
     size: Size
-    uri: Union[AnyUrl, Path]
+    uri: Union[AnyUrl, Path] = Field(union_mode="left_to_right")
     _pil: Optional[PILImage.Image] = None
 
     @property
@@ -1668,7 +1669,7 @@ class DoclingDocument(BaseModel):
         self,
         root: Optional[NodeItem] = None,
         with_groups: bool = False,
-        traverse_pictures: bool = True,
+        traverse_pictures: bool = False,
         page_no: Optional[int] = None,
         _level: int = 0,  # fixed parameter, carries through the node nesting level
     ) -> typing.Iterable[Tuple[NodeItem, int]]:  # tuple of node and level
@@ -1685,30 +1686,31 @@ class DoclingDocument(BaseModel):
         if not root:
             root = self.body
 
+        # Yield non-group items or group items when with_groups=True
         if not isinstance(root, GroupItem) or with_groups:
             if isinstance(root, DocItem):
-                if page_no is not None:
-                    for prov in root.prov:
-                        if prov.page_no == page_no:
-                            yield root, _level
-                else:
+                if page_no is None or any(
+                    prov.page_no == page_no for prov in root.prov
+                ):
                     yield root, _level
             else:
                 yield root, _level
 
+        # Handle picture traversal - only traverse children if requested
+        if isinstance(root, PictureItem) and not traverse_pictures:
+            return
+
         # Traverse children
         for child_ref in root.children:
             child = child_ref.resolve(self)
-
             if isinstance(child, NodeItem):
-                # If the child is a NodeItem, recursively traverse it
-                if not isinstance(child, PictureItem) or traverse_pictures:
-                    yield from self.iterate_items(
-                        child,
-                        _level=_level + 1,
-                        with_groups=with_groups,
-                        page_no=page_no,
-                    )
+                yield from self.iterate_items(
+                    child,
+                    with_groups=with_groups,
+                    traverse_pictures=traverse_pictures,
+                    page_no=page_no,
+                    _level=_level + 1,
+                )
 
     def _clear_picture_pil_cache(self):
         """Clear cache storage of all images."""
@@ -1864,7 +1866,7 @@ class DoclingDocument(BaseModel):
 
         """
         with open(filename, "r") as f:
-            return cls.model_validate(json.loads(f.read()))
+            return cls.model_validate_json(f.read())
 
     def save_as_yaml(
         self,
@@ -2115,10 +2117,30 @@ class DoclingDocument(BaseModel):
         # Bold, Italic, or Bold-Italic
         # Hence, any underscore that we print into Markdown is coming from document text
         # That means we need to escape it, to properly reflect content in the markdown
+        # However, we need to preserve underscores in image URLs
+        # to maintain their validity
+        # For example: ![image](path/to_image.png) should remain unchanged
         def escape_underscores(text):
-            # Replace "_" with "\_" only if it's not already escaped
-            escaped_text = re.sub(r"(?<!\\)_", r"\_", text)
-            return escaped_text
+            """Escape underscores but leave them intact in the URL.."""
+            # Firstly, identify all the URL patterns.
+            url_pattern = r"!\[.*?\]\((.*?)\)"
+            parts = []
+            last_end = 0
+
+            for match in re.finditer(url_pattern, text):
+                # Text to add before the URL (needs to be escaped)
+                before_url = text[last_end : match.start()]
+                parts.append(re.sub(r"(?<!\\)_", r"\_", before_url))
+
+                # Add the full URL part (do not escape)
+                parts.append(match.group(0))
+                last_end = match.end()
+
+            # Add the final part of the text (which needs to be escaped)
+            if last_end < len(text):
+                parts.append(re.sub(r"(?<!\\)_", r"\_", text[last_end:]))
+
+            return "".join(parts)
 
         mdtext = escape_underscores(mdtext)
 

{docling_core-2.7.1 → docling_core-2.9.0}/docling_core/types/legacy_doc/base.py

@@ -140,6 +140,7 @@ class BaseCell(AliasModel):
     obj_type: str = Field(
         alias="type", json_schema_extra=es_field(type="keyword", ignore_above=8191)
     )
+    payload: Optional[dict] = None
 
     def get_location_tokens(
         self,

docling_core-2.9.0/docling_core/utils/legacy.py

@@ -0,0 +1,346 @@
+#
+# Copyright IBM Corp. 2024 - 2024
+# SPDX-License-Identifier: MIT
+#
+
+"""Utilities for converting between legacy and new document format."""
+
+import hashlib
+import uuid
+from typing import Union
+
+from docling_core.types.doc import (
+    DocItem,
+    DocItemLabel,
+    DoclingDocument,
+    PictureItem,
+    SectionHeaderItem,
+    TableCell,
+    TableItem,
+    TextItem,
+)
+from docling_core.types.doc.document import ListItem
+from docling_core.types.legacy_doc.base import (
+    BaseCell,
+    BaseText,
+    Figure,
+    GlmTableCell,
+    PageDimensions,
+    PageReference,
+    Prov,
+    Ref,
+)
+from docling_core.types.legacy_doc.base import Table as DsSchemaTable
+from docling_core.types.legacy_doc.base import TableCell as DsTableCell
+from docling_core.types.legacy_doc.document import (
+    CCSDocumentDescription as DsDocumentDescription,
+)
+from docling_core.types.legacy_doc.document import CCSFileInfoObject as DsFileInfoObject
+from docling_core.types.legacy_doc.document import ExportedCCSDocument as DsDocument
+
+
+def _create_hash(string: str):
+    hasher = hashlib.sha256()
+    hasher.update(string.encode("utf-8"))
+
+    return hasher.hexdigest()
+
+
+def doc_item_label_to_legacy_type(label: DocItemLabel):
+    """Convert the DocItemLabel to the legacy type."""
+    _label_to_ds_type = {
+        DocItemLabel.TITLE: "title",
+        DocItemLabel.DOCUMENT_INDEX: "table-of-contents",
+        DocItemLabel.SECTION_HEADER: "subtitle-level-1",
+        DocItemLabel.CHECKBOX_SELECTED: "checkbox-selected",
+        DocItemLabel.CHECKBOX_UNSELECTED: "checkbox-unselected",
+        DocItemLabel.CAPTION: "caption",
+        DocItemLabel.PAGE_HEADER: "page-header",
+        DocItemLabel.PAGE_FOOTER: "page-footer",
+        DocItemLabel.FOOTNOTE: "footnote",
+        DocItemLabel.TABLE: "table",
+        DocItemLabel.FORMULA: "equation",
+        DocItemLabel.LIST_ITEM: "paragraph",
+        DocItemLabel.CODE: "paragraph",
+        DocItemLabel.PICTURE: "figure",
+        DocItemLabel.TEXT: "paragraph",
+        DocItemLabel.PARAGRAPH: "paragraph",
+    }
+    if label in _label_to_ds_type:
+        return _label_to_ds_type[label]
+    return label.value
+
+
+def doc_item_label_to_legacy_name(label: DocItemLabel):
+    """Convert the DocItemLabel to the legacy name."""
+    _reverse_label_name_mapping = {
+        DocItemLabel.CAPTION: "Caption",
+        DocItemLabel.FOOTNOTE: "Footnote",
+        DocItemLabel.FORMULA: "Formula",
+        DocItemLabel.LIST_ITEM: "List-item",
+        DocItemLabel.PAGE_FOOTER: "Page-footer",
+        DocItemLabel.PAGE_HEADER: "Page-header",
+        DocItemLabel.PICTURE: "Picture",
+        DocItemLabel.SECTION_HEADER: "Section-header",
+        DocItemLabel.TABLE: "Table",
+        DocItemLabel.TEXT: "Text",
+        DocItemLabel.TITLE: "Title",
+        DocItemLabel.DOCUMENT_INDEX: "Document Index",
+        DocItemLabel.CODE: "Code",
+        DocItemLabel.CHECKBOX_SELECTED: "Checkbox-Selected",
+        DocItemLabel.CHECKBOX_UNSELECTED: "Checkbox-Unselected",
+        DocItemLabel.FORM: "Form",
+        DocItemLabel.KEY_VALUE_REGION: "Key-Value Region",
+        DocItemLabel.PARAGRAPH: "paragraph",
+    }
+    if label in _reverse_label_name_mapping:
+        return _reverse_label_name_mapping[label]
+    return label.value
+
+
+def docling_document_to_legacy(doc: DoclingDocument, fallback_filaname: str = "file"):
+    """Convert a DoclingDocument to the legacy format."""
+    title = ""
+    desc: DsDocumentDescription = DsDocumentDescription(logs=[])
+
+    if doc.origin is not None:
+        document_hash = _create_hash(str(doc.origin.binary_hash))
+        filename = doc.origin.filename
+    else:
+        document_hash = _create_hash(str(uuid.uuid4()))
+        filename = fallback_filaname
+
+    page_hashes = [
+        PageReference(
+            hash=_create_hash(document_hash + ":" + str(p.page_no - 1)),
+            page=p.page_no,
+            model="default",
+        )
+        for p in doc.pages.values()
+    ]
+
+    file_info = DsFileInfoObject(
+        filename=filename,
+        document_hash=document_hash,
+        num_pages=len(doc.pages),
+        page_hashes=page_hashes,
+    )
+
+    main_text: list[Union[Ref, BaseText]] = []
+    tables: list[DsSchemaTable] = []
+    figures: list[Figure] = []
+    equations: list[BaseCell] = []
+    footnotes: list[BaseText] = []
+    page_headers: list[BaseText] = []
+    page_footers: list[BaseText] = []
+
+    # TODO: populate page_headers page_footers from doc.furniture
+
+    embedded_captions = set()
+    for ix, (item, level) in enumerate(doc.iterate_items(doc.body)):
+
+        if isinstance(item, (TableItem, PictureItem)) and len(item.captions) > 0:
+            caption = item.caption_text(doc)
+            if caption:
+                embedded_captions.add(caption)
+
+    for item, level in doc.iterate_items():
+        if isinstance(item, DocItem):
+            item_type = item.label
+
+            if isinstance(item, (TextItem, ListItem, SectionHeaderItem)):
+
+                if isinstance(item, ListItem) and item.marker:
+                    text = f"{item.marker} {item.text}"
+                else:
+                    text = item.text
+
+                # Can be empty.
+                prov = [
+                    Prov(
+                        bbox=p.bbox.as_tuple(),
+                        page=p.page_no,
+                        span=[0, len(item.text)],
+                    )
+                    for p in item.prov
+                ]
+                main_text.append(
+                    BaseText(
+                        text=text,
+                        obj_type=doc_item_label_to_legacy_type(item.label),
+                        name=doc_item_label_to_legacy_name(item.label),
+                        prov=prov,
+                    )
+                )
+
+                # skip captions of they are embedded in the actual
+                # floating object
+                if item_type == DocItemLabel.CAPTION and text in embedded_captions:
+                    continue
+
+            elif isinstance(item, TableItem) and item.data:
+                index = len(tables)
+                ref_str = f"#/tables/{index}"
+                main_text.append(
+                    Ref(
+                        name=doc_item_label_to_legacy_name(item.label),
+                        obj_type=doc_item_label_to_legacy_type(item.label),
+                        ref=ref_str,
+                    ),
+                )
+
+                # Initialise empty table data grid (only empty cells)
+                table_data = [
+                    [
+                        DsTableCell(
+                            text="",
+                            # bbox=[0,0,0,0],
+                            spans=[[i, j]],
+                            obj_type="body",
+                        )
+                        for j in range(item.data.num_cols)
+                    ]
+                    for i in range(item.data.num_rows)
+                ]
+
+                # Overwrite cells in table data for which there is actual cell content.
+                for cell in item.data.table_cells:
+                    for i in range(
+                        min(cell.start_row_offset_idx, item.data.num_rows),
+                        min(cell.end_row_offset_idx, item.data.num_rows),
+                    ):
+                        for j in range(
+                            min(cell.start_col_offset_idx, item.data.num_cols),
+                            min(cell.end_col_offset_idx, item.data.num_cols),
+                        ):
+                            celltype = "body"
+                            if cell.column_header:
+                                celltype = "col_header"
+                            elif cell.row_header:
+                                celltype = "row_header"
+                            elif cell.row_section:
+                                celltype = "row_section"
+
+                            def _make_spans(cell: TableCell, table_item: TableItem):
+                                for rspan in range(
+                                    min(
+                                        cell.start_row_offset_idx,
+                                        table_item.data.num_rows,
+                                    ),
+                                    min(
+                                        cell.end_row_offset_idx,
+                                        table_item.data.num_rows,
+                                    ),
+                                ):
+                                    for cspan in range(
+                                        min(
+                                            cell.start_col_offset_idx,
+                                            table_item.data.num_cols,
+                                        ),
+                                        min(
+                                            cell.end_col_offset_idx,
+                                            table_item.data.num_cols,
+                                        ),
+                                    ):
+                                        yield [rspan, cspan]
+
+                            spans = list(_make_spans(cell, item))
+                            table_data[i][j] = GlmTableCell(
+                                text=cell.text,
+                                bbox=(
+                                    cell.bbox.as_tuple()
+                                    if cell.bbox is not None
+                                    else None
+                                ),  # check if this is bottom-left
+                                spans=spans,
+                                obj_type=celltype,
+                                col=j,
+                                row=i,
+                                row_header=cell.row_header,
+                                row_section=cell.row_section,
+                                col_header=cell.column_header,
+                                row_span=[
+                                    cell.start_row_offset_idx,
+                                    cell.end_row_offset_idx,
+                                ],
+                                col_span=[
+                                    cell.start_col_offset_idx,
+                                    cell.end_col_offset_idx,
+                                ],
+                            )
+
+                # Compute the caption
+                caption = item.caption_text(doc)
+
+                tables.append(
+                    DsSchemaTable(
+                        text=caption,
+                        num_cols=item.data.num_cols,
+                        num_rows=item.data.num_rows,
+                        obj_type=doc_item_label_to_legacy_type(item.label),
+                        data=table_data,
+                        prov=[
+                            Prov(
+                                bbox=p.bbox.as_tuple(),
+                                page=p.page_no,
+                                span=[0, 0],
+                            )
+                            for p in item.prov
+                        ],
+                    )
+                )
+
+            elif isinstance(item, PictureItem):
+                index = len(figures)
+                ref_str = f"#/figures/{index}"
+                main_text.append(
+                    Ref(
+                        name=doc_item_label_to_legacy_name(item.label),
+                        obj_type=doc_item_label_to_legacy_type(item.label),
+                        ref=ref_str,
+                    ),
+                )
+
+                # Compute the caption
+                caption = item.caption_text(doc)
+
+                figures.append(
+                    Figure(
+                        prov=[
+                            Prov(
+                                bbox=p.bbox.as_tuple(),
+                                page=p.page_no,
+                                span=[0, len(caption)],
+                            )
+                            for p in item.prov
+                        ],
+                        obj_type=doc_item_label_to_legacy_type(item.label),
+                        text=caption,
+                        # data=[[]],
+                    )
+                )
+
+    page_dimensions = [
+        PageDimensions(page=p.page_no, height=p.size.height, width=p.size.width)
+        for p in doc.pages.values()
+    ]
+
+    legacy_doc: DsDocument = DsDocument(
+        name=title,
+        description=desc,
+        file_info=file_info,
+        main_text=main_text,
+        equations=equations,
+        footnotes=footnotes,
+        page_headers=page_headers,
+        page_footers=page_footers,
+        tables=tables,
+        figures=figures,
+        page_dimensions=page_dimensions,
+    )
+
+    return legacy_doc
+
+
+# def legacy_to_docling_document(legacy_doc: DsDocument) -> DoclingDocument:
+#     """Convert a legacy document to DoclingDocument."""

{docling_core-2.7.1 → docling_core-2.9.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "docling-core"
-version = "2.7.1"
+version = "2.9.0"
 description = "A python library to define and validate data types in Docling."
 license = "MIT"
 authors = [
@@ -55,6 +55,11 @@ pandas = "^2.1.4"
 pillow = "^10.3.0"
 pyyaml = ">=5.1,<7.0.0"
 typing-extensions = "^4.12.2"
+transformers = { version = "^4.34.0", optional = true }
+semchunk = { version = "^2.2.0", optional = true }
+
+[tool.poetry.extras]
+chunking = ["transformers", "semchunk"]
 
 [tool.poetry.group.dev.dependencies]
 black = "^24.4.2"
@@ -121,6 +126,7 @@ module = [
     "jsonschema.*",
     "requests.*",
     "tabulate.*",
+    "transformers.*",
     "yaml.*",
 ]
 ignore_missing_imports = true