From 1a902ea5c8b68bc65bff22616fa4b3c4fade2eff Mon Sep 17 00:00:00 2001
From: kevin9327 <5299031+kevin9327@users.noreply.github.com>
Date: Wed, 14 Oct 2026 09:44:26 +0900
Subject: [PATCH 1/2] feat(pdf): join CJK text lines without a space

A line break between two Han, kana or CJK/full-width punctuation
characters is not a word boundary, so sanitize_text no longer inserts a
space there. Clusters that contain Hangul keep the space, since Korean
separates words with spaces.

Part B of #4672.

Signed-off-by: kevin9327 <5299031+kevin9327@users.noreply.github.com>
---
 .../skills/docling/references/extraction.md   | 23 ++++---
 .../skills/docling/references/python-sdk.md   | 66 +++++++++++--------
 .../.agents/skills/docling/references/rag.md  |  4 +-
 .../docling/references/service-client.md      |  4 +-
 .../page_assemble/page_assemble_model.py      | 52 ++++++++++++++-
 tests/test_page_assemble_model.py             | 64 ++++++++++++++++++
 6 files changed, 172 insertions(+), 41 deletions(-)

diff --git a/docling/.agents/skills/docling/references/extraction.md b/docling/.agents/skills/docling/references/extraction.md
index c7a53655c8..10cdb30681 100644
--- a/docling/.agents/skills/docling/references/extraction.md
+++ b/docling/.agents/skills/docling/references/extraction.md
@@ -33,25 +33,30 @@ a Pydantic model **instance** (`Union[str, dict, BaseModel, Type[BaseModel]]`).
 
 ```python
 # 1. JSON-ish string
-result = extractor.extract(source="invoice.pdf",
-                           template='{"bill_no": "string", "total": "float"}')
+result = extractor.extract(
+    source="invoice.pdf", template='{"bill_no": "string", "total": "float"}'
+)
 
 # 2. dict template
-result = extractor.extract(source="invoice.pdf",
-                           template={"bill_no": "string", "total": "float"})
+result = extractor.extract(
+    source="invoice.pdf", template={"bill_no": "string", "total": "float"}
+)
 
 # 3. Pydantic model class (recommended — typed, self-documenting)
 from pydantic import BaseModel
 
+
 class Invoice(BaseModel):
     bill_no: str
     total: float
 
+
 result = extractor.extract(source="invoice.pdf", template=Invoice)
 
 # 4. Pydantic instance (fields double as examples / defaults)
-result = extractor.extract(source="invoice.pdf",
-                           template=Invoice(bill_no="INV-0001", total=0.0))
+result = extractor.extract(
+    source="invoice.pdf", template=Invoice(bill_no="INV-0001", total=0.0)
+)
 ```
 
 Prefer a **Pydantic model class** for durable schemas — it documents intent and
@@ -70,7 +75,9 @@ result = extractor.extract(source="invoice.pdf", template=Invoice)
 
 if result.status in (ConversionStatus.SUCCESS, ConversionStatus.PARTIAL_SUCCESS):
     for page in result.pages:
-        print(page.page_no, page.extracted_data)   # e.g. {"bill_no": "...", "total": 42.0}
+        print(
+            page.page_no, page.extracted_data
+        )  # e.g. {"bill_no": "...", "total": 42.0}
 else:
     print("extraction failed:", result.errors)
 ```
@@ -81,7 +88,7 @@ else:
 for result in extractor.extract_all(
     source=["a.pdf", "b.pdf", "https://example.com/c.pdf"],
     template=Invoice,
-    raises_on_error=False,     # keep going past individual failures
+    raises_on_error=False,  # keep going past individual failures
 ):
     print(result.input.file.name, result.status)
 ```
diff --git a/docling/.agents/skills/docling/references/python-sdk.md b/docling/.agents/skills/docling/references/python-sdk.md
index 6b764830bb..ab6928e65c 100644
--- a/docling/.agents/skills/docling/references/python-sdk.md
+++ b/docling/.agents/skills/docling/references/python-sdk.md
@@ -11,11 +11,11 @@ controlled by **`PipelineOptions`** subclasses passed per input format.
 from docling.document_converter import DocumentConverter
 
 converter = DocumentConverter()
-result = converter.convert("report.pdf")   # local path or http(s) URL
+result = converter.convert("report.pdf")  # local path or http(s) URL
 
-doc = result.document                       # a DoclingDocument
+doc = result.document  # a DoclingDocument
 print(doc.export_to_markdown())
-data = doc.export_to_dict()                 # structured / lossless
+data = doc.export_to_dict()  # structured / lossless
 ```
 
 `result.status` reports success/failure; `result.document` is always a
@@ -40,11 +40,11 @@ from docling.datamodel.base_models import InputFormat
 from docling.datamodel.pipeline_options import PdfPipelineOptions
 
 pipeline_options = PdfPipelineOptions(
-    do_ocr=True,                  # OCR scanned/image regions (False to skip)
-    do_table_structure=True,      # detect table structure (False = faster)
-    do_code_enrichment=False,     # code understanding
+    do_ocr=True,  # OCR scanned/image regions (False to skip)
+    do_table_structure=True,  # detect table structure (False = faster)
+    do_code_enrichment=False,  # code understanding
     do_formula_enrichment=False,  # formula understanding
-    generate_page_images=False,   # keep page raster images on the document
+    generate_page_images=False,  # keep page raster images on the document
 )
 
 converter = DocumentConverter(
@@ -76,15 +76,20 @@ Useful `PdfPipelineOptions` / base fields:
 
 ```python
 from docling.datamodel.pipeline_options import (
-    PdfPipelineOptions, RapidOcrOptions, TesseractOcrOptions, OcrMacOptions,
+    PdfPipelineOptions,
+    RapidOcrOptions,
+    TesseractOcrOptions,
+    OcrMacOptions,
 )
 
 # Default is EasyOCR:
 opts = PdfPipelineOptions(do_ocr=True)
 # Alternatives:
-opts = PdfPipelineOptions(do_ocr=True, ocr_options=RapidOcrOptions())    # lightweight
-opts = PdfPipelineOptions(do_ocr=True, ocr_options=TesseractOcrOptions())# system Tesseract
-opts = PdfPipelineOptions(do_ocr=True, ocr_options=OcrMacOptions())      # macOS only
+opts = PdfPipelineOptions(do_ocr=True, ocr_options=RapidOcrOptions())  # lightweight
+opts = PdfPipelineOptions(
+    do_ocr=True, ocr_options=TesseractOcrOptions()
+)  # system Tesseract
+opts = PdfPipelineOptions(do_ocr=True, ocr_options=OcrMacOptions())  # macOS only
 ```
 
 Each engine is an optional dependency — see [slim-packaging.md](slim-packaging.md).
@@ -97,7 +102,7 @@ from docling.datamodel.pipeline_options import KserveV2OcrOptions
 
 opts = PdfPipelineOptions(
     do_ocr=True,
-    enable_remote_services=True,   # REQUIRED for KserveV2OcrOptions
+    enable_remote_services=True,  # REQUIRED for KserveV2OcrOptions
     ocr_options=KserveV2OcrOptions(url="localhost:8001", model_name="rapidocr"),
 )
 ```
@@ -129,7 +134,9 @@ from docling.datamodel.accelerator_options import AcceleratorDevice, Accelerator
 from docling.datamodel.pipeline_options import PdfPipelineOptions
 
 opts = PdfPipelineOptions(
-    accelerator_options=AcceleratorOptions(device=AcceleratorDevice.CUDA, num_threads=8),
+    accelerator_options=AcceleratorOptions(
+        device=AcceleratorDevice.CUDA, num_threads=8
+    ),
 )
 ```
 
@@ -208,7 +215,7 @@ tables through the VLM. Reduces hallucination on text-heavy pages.
 ```python
 pipeline_options = VlmPipelineOptions(
     vlm_options=vlm_model_specs.GRANITEDOCLING_TRANSFORMERS,
-    force_backend_text=True,     # hybrid
+    force_backend_text=True,  # hybrid
     generate_page_images=True,
 )
 ```
@@ -237,7 +244,7 @@ vlm_opts = ApiVlmOptions(
 pipeline_options = VlmPipelineOptions(
     vlm_options=vlm_opts,
     generate_page_images=True,
-    enable_remote_services=True,   # REQUIRED — gates all outbound HTTP
+    enable_remote_services=True,  # REQUIRED — gates all outbound HTTP
 )
 converter = DocumentConverter(
     format_options={
@@ -272,9 +279,9 @@ tokenizer = HuggingFaceTokenizer.from_pretrained(
 chunker = HybridChunker(tokenizer=tokenizer, merge_peers=True)
 
 for chunk in chunker.chunk(result.document):
-    embed_text = chunker.contextualize(chunk)   # heading-prefixed text to embed
-    print(chunk.meta.headings)                  # heading breadcrumb
-    print(chunk.meta.origin.page_no)            # source page
+    embed_text = chunker.contextualize(chunk)  # heading-prefixed text to embed
+    print(chunk.meta.headings)  # heading breadcrumb
+    print(chunk.meta.origin.page_no)  # source page
 ```
 
 OpenAI tokenizer (requires `pip install 'docling-core[chunking-openai]'`):
@@ -301,11 +308,11 @@ for item, level in doc.iterate_items():
         print(f"{'#' * level} {item.text}")
 
 for table in doc.tables:
-    print(table.export_to_dataframe())   # pandas DataFrame
+    print(table.export_to_dataframe())  # pandas DataFrame
     print(table.export_to_markdown())
 
 for picture in doc.pictures:
-    print(picture.caption_text(doc))     # caption if present
+    print(picture.caption_text(doc))  # caption if present
 ```
 
 ## Batch conversion & error handling
@@ -318,7 +325,7 @@ from docling.document_converter import DocumentConverter
 from docling.datamodel.base_models import ConversionStatus, InputFormat
 
 converter = DocumentConverter(
-    allowed_formats=[InputFormat.PDF, InputFormat.DOCX],   # reject everything else
+    allowed_formats=[InputFormat.PDF, InputFormat.DOCX],  # reject everything else
 )
 
 sources = ["a.pdf", "b.docx", "https://example.com/c.pdf"]
@@ -376,7 +383,7 @@ converter = DocumentConverter(
     }
 )
 result = converter.convert("interview.wav")
-print(result.document.export_to_markdown())   # timestamped transcript
+print(result.document.export_to_markdown())  # timestamped transcript
 ```
 
 The CLI can also transcribe: `docling interview.wav --to md --output /tmp/`.
@@ -398,7 +405,7 @@ from docling.datamodel.pipeline_options import PdfPipelineOptions
 opts = PdfPipelineOptions(
     generate_page_images=True,
     generate_picture_images=True,
-    images_scale=2.0,            # higher-resolution rasters
+    images_scale=2.0,  # higher-resolution rasters
 )
 ```
 
@@ -437,11 +444,13 @@ from docling.datamodel.backend_options import DeclarativeBackendOptions
 from docling.datamodel.base_models import InputFormat
 from docling.document_converter import DocumentConverter, DoclingJSONFormatOption
 
-converter = DocumentConverter(format_options={
-    InputFormat.JSON_DOCLING: DoclingJSONFormatOption(
-        backend_options=DeclarativeBackendOptions(enable_local_fetch=True)
-    ),
-})
+converter = DocumentConverter(
+    format_options={
+        InputFormat.JSON_DOCLING: DoclingJSONFormatOption(
+            backend_options=DeclarativeBackendOptions(enable_local_fetch=True)
+        ),
+    }
+)
 ```
 
 Or save with `ImageRefMode.EMBEDDED` so the JSON carries its images.
@@ -493,5 +502,6 @@ equivalent is `docling report.pdf --artifacts-path /models`.
 
 ```python
 from importlib.metadata import version
+
 print(version("docling"), version("docling-core"))
 ```
diff --git a/docling/.agents/skills/docling/references/rag.md b/docling/.agents/skills/docling/references/rag.md
index b549283125..915e05154f 100644
--- a/docling/.agents/skills/docling/references/rag.md
+++ b/docling/.agents/skills/docling/references/rag.md
@@ -27,9 +27,9 @@ from langchain_docling.loader import ExportType
 
 loader = DoclingLoader(
     file_path=["report.pdf", "https://example.com/paper.pdf"],
-    export_type=ExportType.DOC_CHUNKS,   # chunked; or ExportType.MARKDOWN
+    export_type=ExportType.DOC_CHUNKS,  # chunked; or ExportType.MARKDOWN
 )
-docs = loader.load()                     # list[langchain_core.documents.Document]
+docs = loader.load()  # list[langchain_core.documents.Document]
 # each chunk keeps metadata: source, headings, page numbers
 ```
 
diff --git a/docling/.agents/skills/docling/references/service-client.md b/docling/.agents/skills/docling/references/service-client.md
index 37c1b8c450..dc59a709e8 100644
--- a/docling/.agents/skills/docling/references/service-client.md
+++ b/docling/.agents/skills/docling/references/service-client.md
@@ -54,7 +54,7 @@ DOCLING_SERVICE_API_KEY=your-api-key    # omit if the service is unauthenticated
 ```python
 from docling.service_client import DoclingServiceClient
 
-client = DoclingServiceClient(url=..., api_key=...)   # or rely on env / .env
+client = DoclingServiceClient(url=..., api_key=...)  # or rely on env / .env
 result = client.convert(source="path/to/report.pdf")  # local path or http(s) URL
 print(result.document.export_to_markdown())
 ```
@@ -67,7 +67,7 @@ from docling.datamodel.service.options import ConvertDocumentsOptions
 
 result = client.convert(
     source="report.pdf",
-    options=ConvertDocumentsOptions(...),   # per-request pipeline overrides
+    options=ConvertDocumentsOptions(...),  # per-request pipeline overrides
 )
 ```
 
diff --git a/docling/models/stages/page_assemble/page_assemble_model.py b/docling/models/stages/page_assemble/page_assemble_model.py
index a7bd223f63..5ca9fdb852 100644
--- a/docling/models/stages/page_assemble/page_assemble_model.py
+++ b/docling/models/stages/page_assemble/page_assemble_model.py
@@ -5,7 +5,7 @@
 import re
 from collections.abc import Iterable
 from pathlib import Path
-from typing import Dict, List, Optional, Union
+from typing import Dict, Final, List, Optional, Union
 
 import numpy as np
 from docling_core.types.doc import BoundingBox
@@ -33,6 +33,33 @@
 
 _log = logging.getLogger(__name__)
 
+# Scripts written without spaces between words: Han ideographs, kana and
+# CJK / full-width punctuation. A line break between two such characters is
+# not a word boundary, so the lines are joined without a space. Hangul is
+# deliberately absent: Korean separates words with spaces.
+_CJK_NO_SPACE_RE = re.compile(
+    "["
+    "\u2e80-\u2fdf"  # CJK radicals, Kangxi radicals
+    "\u3000-\u303f"  # CJK symbols and punctuation
+    "\u3040-\u309f"  # Hiragana
+    "\u30a0-\u30ff"  # Katakana
+    "\u31c0-\u31ef"  # CJK strokes
+    "\u31f0-\u31ff"  # Katakana phonetic extensions
+    "\u3400-\u4dbf"  # CJK Unified Ideographs Extension A
+    "\u4e00-\u9fff"  # CJK Unified Ideographs
+    "\uf900-\ufaff"  # CJK Compatibility Ideographs
+    "\ufe30-\ufe4f"  # CJK compatibility forms (vertical punctuation)
+    "\uff01-\uff60"  # Full-width ASCII variants and punctuation
+    "\uff61-\uff9f"  # Half-width CJK punctuation and katakana
+    "\U00020000-\U0003134f"  # CJK Unified Ideographs Extensions B-H
+    "]"
+)
+# Lowest code point in _CJK_NO_SPACE_RE: a cheap pre-check for Latin text.
+_CJK_MIN_CHAR: Final = "\u2e80"
+_HANGUL_RE = re.compile(
+    "[\u1100-\u11ff\u3130-\u318f\ua960-\ua97f\uac00-\ud7af\ud7b0-\ud7ff]"
+)
+
 # Ligature normalization map.
 # Covers the Unicode Alphabetic Presentation Forms block (U+FB00-U+FB06),
 # the Latin capital/small ligature IJ (U+0132/U+0133), and the Private-Use
@@ -123,12 +150,23 @@ def sanitize_text(self, lines: List[str]) -> str:
         hyphen-prefixed token -- so it is kept and its lines are joined like
         any other line break.
 
+        A break between two CJK characters (Han, kana or CJK / full-width
+        punctuation) is not a word boundary either, since those scripts put
+        no spaces between words, so it is closed up as well -- unless the
+        cluster contains Hangul: Korean separates words with spaces, Hanja
+        included.
+
         The joined text is then normalized: a few typographic characters are
         replaced by ASCII equivalents and ligatures are expanded.
         """
         if len(lines) == 0:
             return ""
 
+        # Korean text mixes Hanja into space-separated words, so a cluster
+        # that contains Hangul keeps the space at every line break. Only
+        # checked once a CJK/CJK break is found, so other text pays nothing.
+        has_hangul: Optional[bool] = None
+
         for ix, line in enumerate(lines[1:]):
             prev_line = lines[ix]
 
@@ -148,6 +186,18 @@ def sanitize_text(self, lines: List[str]) -> str:
                     lines[ix] = prev_line[:-1]
                 elif not hyphen_attached_to_word:
                     lines[ix] += " "
+            elif (
+                prev_line
+                and line
+                and prev_line[-1] >= _CJK_MIN_CHAR
+                and line[0] >= _CJK_MIN_CHAR
+                and _CJK_NO_SPACE_RE.match(prev_line[-1])
+                and _CJK_NO_SPACE_RE.match(line[0])
+            ):
+                if has_hangul is None:
+                    has_hangul = any(_HANGUL_RE.search(ln) for ln in lines)
+                if has_hangul:
+                    lines[ix] += " "
             else:
                 lines[ix] += " "
 
diff --git a/tests/test_page_assemble_model.py b/tests/test_page_assemble_model.py
index 93ef7bb9a8..861657080f 100644
--- a/tests/test_page_assemble_model.py
+++ b/tests/test_page_assemble_model.py
@@ -308,3 +308,67 @@ def test_no_page_size_returns_none(self):
         page.parsed_page = pp
         bbox = BoundingBox(l=10, t=10, r=90, b=20)
         assert PageAssembleModel._match_hyperlink(bbox, page) is None
+
+
+class TestSanitizeTextCjkLineJoin:
+    """Line breaks inside CJK text are not word boundaries (issue #4672)."""
+
+    @pytest.mark.parametrize(
+        ("lines", "expected"),
+        [
+            # Japanese: kanji / kana on both sides of the break
+            (
+                ["衆議院議員早稲田夕季君提出", "インクルーシブ防災を実現する"],
+                "衆議院議員早稲田夕季君提出インクルーシブ防災を実現する",
+            ),
+            # Chinese
+            (
+                ["我们提出了一种新的文档", "转换方法。"],
+                "我们提出了一种新的文档转换方法。",
+            ),
+            # full-width punctuation ends or starts a line
+            (["避難所における、", "良好な生活環境"], "避難所における、良好な生活環境"),
+            (["政府としては", "「避難所」"], "政府としては「避難所」"),
+            # half-width katakana and an ideograph from Extension B
+            (["ｶﾀｶﾅ", "𠀋"], "ｶﾀｶﾅ𠀋"),
+            # more than two lines
+            (["令和", "元年", "十月"], "令和元年十月"),
+        ],
+    )
+    def test_cjk_lines_join_without_space(self, model, lines, expected):
+        assert model.sanitize_text(lines) == expected
+
+    @pytest.mark.parametrize(
+        ("lines", "expected"),
+        [
+            # Latin on one side of the break keeps the space
+            (["docling は", "PDF を変換する"], "docling は PDF を変換する"),
+            (["変換ツール", "docling"], "変換ツール docling"),
+            (["Docling", "文書変換"], "Docling 文書変換"),
+            # digits are not CJK
+            (["令和元年", "10月"], "令和元年 10月"),
+            # plain Latin text is unchanged
+            (["hello", "world"], "hello world"),
+        ],
+    )
+    def test_mixed_script_keeps_space(self, model, lines, expected):
+        assert model.sanitize_text(lines) == expected
+
+    def test_korean_keeps_space(self, model):
+        """Korean separates words with spaces; a line break is a word boundary."""
+        assert (
+            model.sanitize_text(["문서 변환", "도구입니다"]) == "문서 변환 도구입니다"
+        )
+
+    def test_korean_with_hanja_keeps_space(self, model):
+        """Hanja inside a Korean cluster are space-separated words too."""
+        assert (
+            model.sanitize_text(["대한민국 憲法", "第一條 대한민국은"])
+            == "대한민국 憲法 第一條 대한민국은"
+        )
+
+    def test_hyphenation_still_applies(self, model):
+        assert model.sanitize_text(["文書 algo-", "rithms"]) == "文書 algorithms"
+
+    def test_empty_line_between_cjk_lines(self, model):
+        assert model.sanitize_text(["日本", "", "語"]) == "日本  語"

From 130bb27ffc00f230c0a40f34302482e14165c77e Mon Sep 17 00:00:00 2001
From: kevin9327 <5299031+kevin9327@users.noreply.github.com>
Date: Sun, 18 Oct 2026 09:21:34 +0900
Subject: [PATCH 2/2] chore: drop unrelated reformatting of the docling skill
 references

Keep this PR to the CJK line-join change in sanitize_text.

Signed-off-by: kevin9327 <5299031+kevin9327@users.noreply.github.com>
---
 .../skills/docling/references/extraction.md   | 23 +++----
 .../skills/docling/references/python-sdk.md   | 66 ++++++++-----------
 .../.agents/skills/docling/references/rag.md  |  4 +-
 .../docling/references/service-client.md      |  4 +-
 4 files changed, 40 insertions(+), 57 deletions(-)

diff --git a/docling/.agents/skills/docling/references/extraction.md b/docling/.agents/skills/docling/references/extraction.md
index 10cdb30681..c7a53655c8 100644
--- a/docling/.agents/skills/docling/references/extraction.md
+++ b/docling/.agents/skills/docling/references/extraction.md
@@ -33,30 +33,25 @@ a Pydantic model **instance** (`Union[str, dict, BaseModel, Type[BaseModel]]`).
 
 ```python
 # 1. JSON-ish string
-result = extractor.extract(
-    source="invoice.pdf", template='{"bill_no": "string", "total": "float"}'
-)
+result = extractor.extract(source="invoice.pdf",
+                           template='{"bill_no": "string", "total": "float"}')
 
 # 2. dict template
-result = extractor.extract(
-    source="invoice.pdf", template={"bill_no": "string", "total": "float"}
-)
+result = extractor.extract(source="invoice.pdf",
+                           template={"bill_no": "string", "total": "float"})
 
 # 3. Pydantic model class (recommended — typed, self-documenting)
 from pydantic import BaseModel
 
-
 class Invoice(BaseModel):
     bill_no: str
     total: float
 
-
 result = extractor.extract(source="invoice.pdf", template=Invoice)
 
 # 4. Pydantic instance (fields double as examples / defaults)
-result = extractor.extract(
-    source="invoice.pdf", template=Invoice(bill_no="INV-0001", total=0.0)
-)
+result = extractor.extract(source="invoice.pdf",
+                           template=Invoice(bill_no="INV-0001", total=0.0))
 ```
 
 Prefer a **Pydantic model class** for durable schemas — it documents intent and
@@ -75,9 +70,7 @@ result = extractor.extract(source="invoice.pdf", template=Invoice)
 
 if result.status in (ConversionStatus.SUCCESS, ConversionStatus.PARTIAL_SUCCESS):
     for page in result.pages:
-        print(
-            page.page_no, page.extracted_data
-        )  # e.g. {"bill_no": "...", "total": 42.0}
+        print(page.page_no, page.extracted_data)   # e.g. {"bill_no": "...", "total": 42.0}
 else:
     print("extraction failed:", result.errors)
 ```
@@ -88,7 +81,7 @@ else:
 for result in extractor.extract_all(
     source=["a.pdf", "b.pdf", "https://example.com/c.pdf"],
     template=Invoice,
-    raises_on_error=False,  # keep going past individual failures
+    raises_on_error=False,     # keep going past individual failures
 ):
     print(result.input.file.name, result.status)
 ```
diff --git a/docling/.agents/skills/docling/references/python-sdk.md b/docling/.agents/skills/docling/references/python-sdk.md
index ab6928e65c..6b764830bb 100644
--- a/docling/.agents/skills/docling/references/python-sdk.md
+++ b/docling/.agents/skills/docling/references/python-sdk.md
@@ -11,11 +11,11 @@ controlled by **`PipelineOptions`** subclasses passed per input format.
 from docling.document_converter import DocumentConverter
 
 converter = DocumentConverter()
-result = converter.convert("report.pdf")  # local path or http(s) URL
+result = converter.convert("report.pdf")   # local path or http(s) URL
 
-doc = result.document  # a DoclingDocument
+doc = result.document                       # a DoclingDocument
 print(doc.export_to_markdown())
-data = doc.export_to_dict()  # structured / lossless
+data = doc.export_to_dict()                 # structured / lossless
 ```
 
 `result.status` reports success/failure; `result.document` is always a
@@ -40,11 +40,11 @@ from docling.datamodel.base_models import InputFormat
 from docling.datamodel.pipeline_options import PdfPipelineOptions
 
 pipeline_options = PdfPipelineOptions(
-    do_ocr=True,  # OCR scanned/image regions (False to skip)
-    do_table_structure=True,  # detect table structure (False = faster)
-    do_code_enrichment=False,  # code understanding
+    do_ocr=True,                  # OCR scanned/image regions (False to skip)
+    do_table_structure=True,      # detect table structure (False = faster)
+    do_code_enrichment=False,     # code understanding
     do_formula_enrichment=False,  # formula understanding
-    generate_page_images=False,  # keep page raster images on the document
+    generate_page_images=False,   # keep page raster images on the document
 )
 
 converter = DocumentConverter(
@@ -76,20 +76,15 @@ Useful `PdfPipelineOptions` / base fields:
 
 ```python
 from docling.datamodel.pipeline_options import (
-    PdfPipelineOptions,
-    RapidOcrOptions,
-    TesseractOcrOptions,
-    OcrMacOptions,
+    PdfPipelineOptions, RapidOcrOptions, TesseractOcrOptions, OcrMacOptions,
 )
 
 # Default is EasyOCR:
 opts = PdfPipelineOptions(do_ocr=True)
 # Alternatives:
-opts = PdfPipelineOptions(do_ocr=True, ocr_options=RapidOcrOptions())  # lightweight
-opts = PdfPipelineOptions(
-    do_ocr=True, ocr_options=TesseractOcrOptions()
-)  # system Tesseract
-opts = PdfPipelineOptions(do_ocr=True, ocr_options=OcrMacOptions())  # macOS only
+opts = PdfPipelineOptions(do_ocr=True, ocr_options=RapidOcrOptions())    # lightweight
+opts = PdfPipelineOptions(do_ocr=True, ocr_options=TesseractOcrOptions())# system Tesseract
+opts = PdfPipelineOptions(do_ocr=True, ocr_options=OcrMacOptions())      # macOS only
 ```
 
 Each engine is an optional dependency — see [slim-packaging.md](slim-packaging.md).
@@ -102,7 +97,7 @@ from docling.datamodel.pipeline_options import KserveV2OcrOptions
 
 opts = PdfPipelineOptions(
     do_ocr=True,
-    enable_remote_services=True,  # REQUIRED for KserveV2OcrOptions
+    enable_remote_services=True,   # REQUIRED for KserveV2OcrOptions
     ocr_options=KserveV2OcrOptions(url="localhost:8001", model_name="rapidocr"),
 )
 ```
@@ -134,9 +129,7 @@ from docling.datamodel.accelerator_options import AcceleratorDevice, Accelerator
 from docling.datamodel.pipeline_options import PdfPipelineOptions
 
 opts = PdfPipelineOptions(
-    accelerator_options=AcceleratorOptions(
-        device=AcceleratorDevice.CUDA, num_threads=8
-    ),
+    accelerator_options=AcceleratorOptions(device=AcceleratorDevice.CUDA, num_threads=8),
 )
 ```
 
@@ -215,7 +208,7 @@ tables through the VLM. Reduces hallucination on text-heavy pages.
 ```python
 pipeline_options = VlmPipelineOptions(
     vlm_options=vlm_model_specs.GRANITEDOCLING_TRANSFORMERS,
-    force_backend_text=True,  # hybrid
+    force_backend_text=True,     # hybrid
     generate_page_images=True,
 )
 ```
@@ -244,7 +237,7 @@ vlm_opts = ApiVlmOptions(
 pipeline_options = VlmPipelineOptions(
     vlm_options=vlm_opts,
     generate_page_images=True,
-    enable_remote_services=True,  # REQUIRED — gates all outbound HTTP
+    enable_remote_services=True,   # REQUIRED — gates all outbound HTTP
 )
 converter = DocumentConverter(
     format_options={
@@ -279,9 +272,9 @@ tokenizer = HuggingFaceTokenizer.from_pretrained(
 chunker = HybridChunker(tokenizer=tokenizer, merge_peers=True)
 
 for chunk in chunker.chunk(result.document):
-    embed_text = chunker.contextualize(chunk)  # heading-prefixed text to embed
-    print(chunk.meta.headings)  # heading breadcrumb
-    print(chunk.meta.origin.page_no)  # source page
+    embed_text = chunker.contextualize(chunk)   # heading-prefixed text to embed
+    print(chunk.meta.headings)                  # heading breadcrumb
+    print(chunk.meta.origin.page_no)            # source page
 ```
 
 OpenAI tokenizer (requires `pip install 'docling-core[chunking-openai]'`):
@@ -308,11 +301,11 @@ for item, level in doc.iterate_items():
         print(f"{'#' * level} {item.text}")
 
 for table in doc.tables:
-    print(table.export_to_dataframe())  # pandas DataFrame
+    print(table.export_to_dataframe())   # pandas DataFrame
     print(table.export_to_markdown())
 
 for picture in doc.pictures:
-    print(picture.caption_text(doc))  # caption if present
+    print(picture.caption_text(doc))     # caption if present
 ```
 
 ## Batch conversion & error handling
@@ -325,7 +318,7 @@ from docling.document_converter import DocumentConverter
 from docling.datamodel.base_models import ConversionStatus, InputFormat
 
 converter = DocumentConverter(
-    allowed_formats=[InputFormat.PDF, InputFormat.DOCX],  # reject everything else
+    allowed_formats=[InputFormat.PDF, InputFormat.DOCX],   # reject everything else
 )
 
 sources = ["a.pdf", "b.docx", "https://example.com/c.pdf"]
@@ -383,7 +376,7 @@ converter = DocumentConverter(
     }
 )
 result = converter.convert("interview.wav")
-print(result.document.export_to_markdown())  # timestamped transcript
+print(result.document.export_to_markdown())   # timestamped transcript
 ```
 
 The CLI can also transcribe: `docling interview.wav --to md --output /tmp/`.
@@ -405,7 +398,7 @@ from docling.datamodel.pipeline_options import PdfPipelineOptions
 opts = PdfPipelineOptions(
     generate_page_images=True,
     generate_picture_images=True,
-    images_scale=2.0,  # higher-resolution rasters
+    images_scale=2.0,            # higher-resolution rasters
 )
 ```
 
@@ -444,13 +437,11 @@ from docling.datamodel.backend_options import DeclarativeBackendOptions
 from docling.datamodel.base_models import InputFormat
 from docling.document_converter import DocumentConverter, DoclingJSONFormatOption
 
-converter = DocumentConverter(
-    format_options={
-        InputFormat.JSON_DOCLING: DoclingJSONFormatOption(
-            backend_options=DeclarativeBackendOptions(enable_local_fetch=True)
-        ),
-    }
-)
+converter = DocumentConverter(format_options={
+    InputFormat.JSON_DOCLING: DoclingJSONFormatOption(
+        backend_options=DeclarativeBackendOptions(enable_local_fetch=True)
+    ),
+})
 ```
 
 Or save with `ImageRefMode.EMBEDDED` so the JSON carries its images.
@@ -502,6 +493,5 @@ equivalent is `docling report.pdf --artifacts-path /models`.
 
 ```python
 from importlib.metadata import version
-
 print(version("docling"), version("docling-core"))
 ```
diff --git a/docling/.agents/skills/docling/references/rag.md b/docling/.agents/skills/docling/references/rag.md
index 915e05154f..b549283125 100644
--- a/docling/.agents/skills/docling/references/rag.md
+++ b/docling/.agents/skills/docling/references/rag.md
@@ -27,9 +27,9 @@ from langchain_docling.loader import ExportType
 
 loader = DoclingLoader(
     file_path=["report.pdf", "https://example.com/paper.pdf"],
-    export_type=ExportType.DOC_CHUNKS,  # chunked; or ExportType.MARKDOWN
+    export_type=ExportType.DOC_CHUNKS,   # chunked; or ExportType.MARKDOWN
 )
-docs = loader.load()  # list[langchain_core.documents.Document]
+docs = loader.load()                     # list[langchain_core.documents.Document]
 # each chunk keeps metadata: source, headings, page numbers
 ```
 
diff --git a/docling/.agents/skills/docling/references/service-client.md b/docling/.agents/skills/docling/references/service-client.md
index dc59a709e8..37c1b8c450 100644
--- a/docling/.agents/skills/docling/references/service-client.md
+++ b/docling/.agents/skills/docling/references/service-client.md
@@ -54,7 +54,7 @@ DOCLING_SERVICE_API_KEY=your-api-key    # omit if the service is unauthenticated
 ```python
 from docling.service_client import DoclingServiceClient
 
-client = DoclingServiceClient(url=..., api_key=...)  # or rely on env / .env
+client = DoclingServiceClient(url=..., api_key=...)   # or rely on env / .env
 result = client.convert(source="path/to/report.pdf")  # local path or http(s) URL
 print(result.document.export_to_markdown())
 ```
@@ -67,7 +67,7 @@ from docling.datamodel.service.options import ConvertDocumentsOptions
 
 result = client.convert(
     source="report.pdf",
-    options=ConvertDocumentsOptions(...),  # per-request pipeline overrides
+    options=ConvertDocumentsOptions(...),   # per-request pipeline overrides
 )
 ```