docling 2.42.2__tar.gz → 2.43.0__tar.gz

{docling-2.42.2 → docling-2.43.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docling
-Version: 2.42.2
+Version: 2.43.0
 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
@@ -28,9 +28,9 @@ License-File: LICENSE
 Requires-Dist: pydantic<3.0.0,>=2.0.0
 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: docling-ibm-models<4,>=3.9.0
 Requires-Dist: filetype<2.0.0,>=1.2.0
-Requires-Dist: pypdfium2<5.0.0,>=4.30.0
+Requires-Dist: pypdfium2!=4.30.1,<5.0.0,>=4.30.0
 Requires-Dist: pydantic-settings<3.0.0,>=2.3.0
 Requires-Dist: huggingface_hub<1,>=0.23
 Requires-Dist: requests<3.0.0,>=2.32.2

{docling-2.42.2 → docling-2.43.0}/docling/backend/html_backend.py

@@ -1,6 +1,5 @@
 import logging
 import re
-import traceback
 from io import BytesIO
 from pathlib import Path
 from typing import Final, Optional, Union, cast
@@ -144,11 +143,7 @@ class HTMLDocumentBackend(DeclarativeDocumentBackend):
         )
         # reset context
         self.ctx = _Context()
-
-        try:
-            self._walk(content, doc)
-        except Exception:
-            print(traceback.format_exc())
+        self._walk(content, doc)
 
         return doc
 

{docling-2.42.2 → docling-2.43.0}/docling/backend/md_backend.py

@@ -5,7 +5,7 @@ from copy import deepcopy
 from enum import Enum
 from io import BytesIO
 from pathlib import Path
-from typing import List, Literal, Optional, Set, Union
+from typing import Literal, Optional, Union, cast
 
 import marko
 import marko.element
@@ -14,6 +14,7 @@ from docling_core.types.doc import (
     DocItemLabel,
     DoclingDocument,
     DocumentOrigin,
+    ListItem,
     NodeItem,
     TableCell,
     TableData,
@@ -89,7 +90,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
     def __init__(self, in_doc: "InputDocument", path_or_stream: Union[BytesIO, Path]):
         super().__init__(in_doc, path_or_stream)
 
-        _log.debug("MD INIT!!!")
+        _log.debug("Starting MarkdownDocumentBackend...")
 
         # Markdown file:
         self.path_or_stream = path_or_stream
@@ -131,7 +132,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
             for md_table_row in self.md_table_buffer:
                 _log.debug(md_table_row)
             _log.debug("=== TABLE END ===")
-            tcells: List[TableCell] = []
+            tcells: list[TableCell] = []
             result_table = []
             for n, md_table_row in enumerate(self.md_table_buffer):
                 data = []
@@ -232,11 +233,12 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
         element: marko.element.Element,
         depth: int,
         doc: DoclingDocument,
-        visited: Set[marko.element.Element],
+        visited: set[marko.element.Element],
         creation_stack: list[
             _CreationPayload
         ],  # stack for lazy item creation triggered deep in marko's AST (on RawText)
         list_ordered_flag_by_ref: dict[str, bool],
+        list_last_item_by_ref: dict[str, ListItem],
         parent_item: Optional[NodeItem] = None,
         formatting: Optional[Formatting] = None,
         hyperlink: Optional[Union[AnyUrl, Path]] = None,
@@ -279,7 +281,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
 
         elif (
             isinstance(element, marko.block.ListItem)
-            and len(element.children) == 1
+            and len(element.children) > 0
             and isinstance((child := element.children[0]), marko.block.Paragraph)
             and len(child.children) > 0
         ):
@@ -291,7 +293,15 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                 if parent_item
                 else False
             )
-            if len(child.children) > 1:  # inline group will be created further down
+            non_list_children: list[marko.element.Element] = [
+                item
+                for item in child.children
+                if not isinstance(item, marko.block.ListItem)
+            ]
+            if len(non_list_children) > 1:  # inline group will be created further down
+                parent_ref: Optional[str] = (
+                    parent_item.self_ref if parent_item else None
+                )
                 parent_item = self._create_list_item(
                     doc=doc,
                     parent_item=parent_item,
@@ -300,6 +310,8 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                     formatting=formatting,
                     hyperlink=hyperlink,
                 )
+                if parent_ref:
+                    list_last_item_by_ref[parent_ref] = cast(ListItem, parent_item)
             else:
                 creation_stack.append(_ListItemCreationPayload(enumerated=enumerated))
 
@@ -334,9 +346,11 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                 element.dest
             )
 
-        elif isinstance(element, marko.inline.RawText):
-            _log.debug(f" - Paragraph (raw text): {element.children}")
-            snippet_text = element.children.strip()
+        elif isinstance(element, (marko.inline.RawText, marko.inline.Literal)):
+            _log.debug(f" - RawText/Literal: {element.children}")
+            snippet_text = (
+                element.children.strip() if isinstance(element.children, str) else ""
+            )
             # Detect start of the table:
             if "|" in snippet_text or self.in_table:
                 # most likely part of the markdown table
@@ -359,6 +373,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                                 if parent_item
                                 else False
                             )
+                            parent_ref = parent_item.self_ref if parent_item else None
                             parent_item = self._create_list_item(
                                 doc=doc,
                                 parent_item=parent_item,
@@ -367,6 +382,11 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                                 formatting=formatting,
                                 hyperlink=hyperlink,
                             )
+                            if parent_ref:
+                                list_last_item_by_ref[parent_ref] = cast(
+                                    ListItem, parent_item
+                                )
+
                         elif isinstance(to_create, _HeadingCreationPayload):
                             # not keeping as parent_item as logic for correctly tracking
                             # that not implemented yet (section components not captured
@@ -458,6 +478,17 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
             element, processed_block_types
         ):
             for child in element.children:
+                if (
+                    isinstance(element, marko.block.ListItem)
+                    and isinstance(child, marko.block.List)
+                    and parent_item
+                    and list_last_item_by_ref.get(parent_item.self_ref, None)
+                ):
+                    _log.debug(
+                        f"walking into new List hanging from item of parent list {parent_item.self_ref}"
+                    )
+                    parent_item = list_last_item_by_ref[parent_item.self_ref]
+
                 self._iterate_elements(
                     element=child,
                     depth=depth + 1,
@@ -465,6 +496,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                     visited=visited,
                     creation_stack=creation_stack,
                     list_ordered_flag_by_ref=list_ordered_flag_by_ref,
+                    list_last_item_by_ref=list_last_item_by_ref,
                     parent_item=parent_item,
                     formatting=formatting,
                     hyperlink=hyperlink,
@@ -483,7 +515,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
         return False
 
     @classmethod
-    def supported_formats(cls) -> Set[InputFormat]:
+    def supported_formats(cls) -> set[InputFormat]:
         return {InputFormat.MD}
 
     def convert(self) -> DoclingDocument:
@@ -510,6 +542,7 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                 visited=set(),
                 creation_stack=[],
                 list_ordered_flag_by_ref={},
+                list_last_item_by_ref={},
             )
             self._close_table(doc=doc)  # handle any last hanging table
 
@@ -534,7 +567,6 @@ class MarkdownDocumentBackend(DeclarativeDocumentBackend):
                 ]:
                     html_str = _restore_original_html(txt=html_str, regex=regex)
                 self._html_blocks = 0
-
                 # delegate to HTML backend
                 stream = BytesIO(bytes(html_str, encoding="utf-8"))
                 in_doc = InputDocument(

{docling-2.42.2 → docling-2.43.0}/docling/datamodel/pipeline_options.py

@@ -332,3 +332,18 @@ class ProcessingPipeline(str, Enum):
     STANDARD = "standard"
     VLM = "vlm"
     ASR = "asr"
+
+
+class ThreadedPdfPipelineOptions(PdfPipelineOptions):
+    """Pipeline options for the threaded PDF pipeline with batching and backpressure control"""
+
+    # Batch sizes for different stages
+    ocr_batch_size: int = 4
+    layout_batch_size: int = 4
+    table_batch_size: int = 4
+
+    # Timing control
+    batch_timeout_seconds: float = 2.0
+
+    # Backpressure and queue control
+    queue_max_size: int = 100

{docling-2.42.2 → docling-2.43.0}/docling/datamodel/settings.py

@@ -26,18 +26,13 @@ class DocumentLimits(BaseModel):
 
 
 class BatchConcurrencySettings(BaseModel):
-    doc_batch_size: int = 2
-    doc_batch_concurrency: int = 2
-    page_batch_size: int = 4
-    page_batch_concurrency: int = 2
-    elements_batch_size: int = 16
-
-    # doc_batch_size: int = 1
-    # doc_batch_concurrency: int = 1
-    # page_batch_size: int = 1
-    # page_batch_concurrency: int = 1
-
-    # model_concurrency: int = 2
+    doc_batch_size: int = 1  # Number of documents processed in one batch. Should be >= doc_batch_concurrency
+    doc_batch_concurrency: int = 1  # Number of parallel threads processing documents. Warning: Experimental! No benefit expected without free-threaded python.
+    page_batch_size: int = 4  # Number of pages processed in one batch.
+    page_batch_concurrency: int = 1  # Currently unused.
+    elements_batch_size: int = (
+        16  # Number of elements processed in one batch, in enrichment models.
+    )
 
     # To force models into single core: export OMP_NUM_THREADS=1
 

{docling-2.42.2 → docling-2.43.0}/docling/document_converter.py

@@ -4,6 +4,7 @@ import sys
 import threading
 import time
 from collections.abc import Iterable, Iterator
+from concurrent.futures import ThreadPoolExecutor
 from functools import partial
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple, Type, Union
@@ -284,24 +285,33 @@ class DocumentConverter:
             settings.perf.doc_batch_size,  # pass format_options
         ):
             _log.info("Going to convert document batch...")
+            process_func = partial(
+                self._process_document, raises_on_error=raises_on_error
+            )
 
-            # parallel processing only within input_batch
-            # with ThreadPoolExecutor(
-            #    max_workers=settings.perf.doc_batch_concurrency
-            # ) as pool:
-            #   yield from pool.map(self.process_document, input_batch)
-            # Note: PDF backends are not thread-safe, thread pool usage was disabled.
-
-            for item in map(
-                partial(self._process_document, raises_on_error=raises_on_error),
-                input_batch,
+            if (
+                settings.perf.doc_batch_concurrency > 1
+                and settings.perf.doc_batch_size > 1
             ):
-                elapsed = time.monotonic() - start_time
-                start_time = time.monotonic()
-                _log.info(
-                    f"Finished converting document {item.input.file.name} in {elapsed:.2f} sec."
-                )
-                yield item
+                with ThreadPoolExecutor(
+                    max_workers=settings.perf.doc_batch_concurrency
+                ) as pool:
+                    for item in pool.map(
+                        process_func,
+                        input_batch,
+                    ):
+                        yield item
+            else:
+                for item in map(
+                    process_func,
+                    input_batch,
+                ):
+                    elapsed = time.monotonic() - start_time
+                    start_time = time.monotonic()
+                    _log.info(
+                        f"Finished converting document {item.input.file.name} in {elapsed:.2f} sec."
+                    )
+                    yield item
 
     def _get_pipeline(self, doc_format: InputFormat) -> Optional[BasePipeline]:
         """Retrieve or initialize a pipeline, reusing instances based on class and options."""
@@ -330,7 +340,7 @@ class DocumentConverter:
                     f"Reusing cached pipeline for {pipeline_class.__name__} with options hash {options_hash}"
                 )
 
-        return self.initialized_pipelines[cache_key]
+            return self.initialized_pipelines[cache_key]
 
     def _process_document(
         self, in_doc: InputDocument, raises_on_error: bool

{docling-2.42.2 → docling-2.43.0}/docling/models/layout_model.py

@@ -3,7 +3,7 @@ import logging
 import warnings
 from collections.abc import Iterable
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional, Union
 
 import numpy as np
 from docling_core.types.doc import DocItemLabel
@@ -148,72 +148,90 @@ class LayoutModel(BasePageModel):
     def __call__(
         self, conv_res: ConversionResult, page_batch: Iterable[Page]
     ) -> Iterable[Page]:
-        for page in page_batch:
+        # Convert to list to allow multiple iterations
+        pages = list(page_batch)
+
+        # Separate valid and invalid pages
+        valid_pages = []
+        valid_page_images: List[Union[Image.Image, np.ndarray]] = []
+
+        for page in pages:
             assert page._backend is not None
             if not page._backend.is_valid():
-                yield page
-            else:
-                with TimeRecorder(conv_res, "layout"):
-                    assert page.size is not None
-                    page_image = page.get_image(scale=1.0)
-                    assert page_image is not None
-
-                    clusters = []
-                    for ix, pred_item in enumerate(
-                        self.layout_predictor.predict(page_image)
-                    ):
-                        label = DocItemLabel(
-                            pred_item["label"]
-                            .lower()
-                            .replace(" ", "_")
-                            .replace("-", "_")
-                        )  # Temporary, until docling-ibm-model uses docling-core types
-                        cluster = Cluster(
-                            id=ix,
-                            label=label,
-                            confidence=pred_item["confidence"],
-                            bbox=BoundingBox.model_validate(pred_item),
-                            cells=[],
-                        )
-                        clusters.append(cluster)
-
-                    if settings.debug.visualize_raw_layout:
-                        self.draw_clusters_and_cells_side_by_side(
-                            conv_res, page, clusters, mode_prefix="raw"
-                        )
-
-                    # Apply postprocessing
-
-                    processed_clusters, processed_cells = LayoutPostprocessor(
-                        page, clusters, self.options
-                    ).postprocess()
-                    # Note: LayoutPostprocessor updates page.cells and page.parsed_page internally
-
-                    with warnings.catch_warnings():
-                        warnings.filterwarnings(
-                            "ignore",
-                            "Mean of empty slice|invalid value encountered in scalar divide",
-                            RuntimeWarning,
-                            "numpy",
-                        )
-
-                        conv_res.confidence.pages[page.page_no].layout_score = float(
-                            np.mean([c.confidence for c in processed_clusters])
-                        )
-
-                        conv_res.confidence.pages[page.page_no].ocr_score = float(
-                            np.mean(
-                                [c.confidence for c in processed_cells if c.from_ocr]
-                            )
-                        )
-
-                    page.predictions.layout = LayoutPrediction(
-                        clusters=processed_clusters
-                    )
-
-                if settings.debug.visualize_layout:
-                    self.draw_clusters_and_cells_side_by_side(
-                        conv_res, page, processed_clusters, mode_prefix="postprocessed"
-                    )
+                continue
 
+            assert page.size is not None
+            page_image = page.get_image(scale=1.0)
+            assert page_image is not None
+
+            valid_pages.append(page)
+            valid_page_images.append(page_image)
+
+        # Process all valid pages with batch prediction
+        batch_predictions = []
+        if valid_page_images:
+            with TimeRecorder(conv_res, "layout"):
+                batch_predictions = self.layout_predictor.predict_batch(  # type: ignore[attr-defined]
+                    valid_page_images
+                )
+
+        # Process each page with its predictions
+        valid_page_idx = 0
+        for page in pages:
+            assert page._backend is not None
+            if not page._backend.is_valid():
                 yield page
+                continue
+
+            page_predictions = batch_predictions[valid_page_idx]
+            valid_page_idx += 1
+
+            clusters = []
+            for ix, pred_item in enumerate(page_predictions):
+                label = DocItemLabel(
+                    pred_item["label"].lower().replace(" ", "_").replace("-", "_")
+                )  # Temporary, until docling-ibm-model uses docling-core types
+                cluster = Cluster(
+                    id=ix,
+                    label=label,
+                    confidence=pred_item["confidence"],
+                    bbox=BoundingBox.model_validate(pred_item),
+                    cells=[],
+                )
+                clusters.append(cluster)
+
+            if settings.debug.visualize_raw_layout:
+                self.draw_clusters_and_cells_side_by_side(
+                    conv_res, page, clusters, mode_prefix="raw"
+                )
+
+            # Apply postprocessing
+            processed_clusters, processed_cells = LayoutPostprocessor(
+                page, clusters, self.options
+            ).postprocess()
+            # Note: LayoutPostprocessor updates page.cells and page.parsed_page internally
+
+            with warnings.catch_warnings():
+                warnings.filterwarnings(
+                    "ignore",
+                    "Mean of empty slice|invalid value encountered in scalar divide",
+                    RuntimeWarning,
+                    "numpy",
+                )
+
+                conv_res.confidence.pages[page.page_no].layout_score = float(
+                    np.mean([c.confidence for c in processed_clusters])
+                )
+
+                conv_res.confidence.pages[page.page_no].ocr_score = float(
+                    np.mean([c.confidence for c in processed_cells if c.from_ocr])
+                )
+
+            page.predictions.layout = LayoutPrediction(clusters=processed_clusters)
+
+            if settings.debug.visualize_layout:
+                self.draw_clusters_and_cells_side_by_side(
+                    conv_res, page, processed_clusters, mode_prefix="postprocessed"
+                )
+
+            yield page