Introduce a unified TransformersExtractionModel that supports multiple
prompt styles via an ExtractionPromptStyle enum. This replaces the
need for separate model classes per VLM.
- Add ExtractionPromptStyle enum (NUEXTRACT, GRANITE_VISION)
- Add prompt_utils.py with style-specific prompt builders
- Add TransformersExtractionModel with prompt-style dispatch
- Add GRANITE_VISION_4_1_TRANSFORMERS model spec
- Add extraction_prompt_style field to VlmExtractionPipelineOptions
Signed-off-by: Ben Wiesel <benwiesel@ibm.com>
Co-authored-by: Ben Wiesel <benwiesel@ibm.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Remove eager materialization from docling-service batch submission
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Update lock
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Make convert_all evaluate Iterable input lazily, remove raises_on_error
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Make convert_all use async generator like submit_and_retrieve_many
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* fix mypy fast check
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* upgrade packages
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* update test GT data
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Update test GT from linux machine
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Reset all GT test data and uv.lock
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
---------
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* fix(docx): handle missing chr attribute in groupChr OMML elements
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* fix(docx): escape spaces in OMML limit text for proper LaTeX rendering
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* fix(docx): fix inline equation reconstruction to prevent tag corruption
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* chore(docx): add type hints and docstrings to OMML module
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* fix(docx): fix genfrac formatting and eliminate grouping function warnings
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* fix(docx): handle unmapped characters in OMML % formatting
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
---------
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* ci: prototype tach-based modular skipping
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: modularize ubuntu setup and refine gating
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: adopt metaxy-inspired governance helpers
- replace custom aggregate check with re-actors/alls-green
- set FORCE_JAVASCRIPT_ACTIONS_TO_NODE24 on every workflow
- keep PR concurrency alive when the graphite:merge label is present
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: tune checks and pin action versions
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: split CI suites and heavy examples
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* DCO Remediation Commit for Georg Heiler <georg.kf.heiler@gmail.com>
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: ecaa4777886157d5c2a7b3893c3a820983089dbf
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: d15416f3ca94ac97af2a8317cd6404208db9d896
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: sharpen tach graph and per-suite path filters
- Split docling.pipeline into per-pipeline tach modules
(asr, vlm, standard_pdf, threaded_standard_pdf, legacy_standard_pdf,
extraction_vlm, base, base_extraction, simple) so pytest --tach-base
impact analysis can attribute changes to a specific pipeline rather
than the whole package.
- Split the asr- and vlm-specific docling.datamodel option files
(asr_model_specs, pipeline_options_asr_model, vlm_engine_options,
vlm_model_specs, pipeline_options_vlm_model, layout_model_specs,
stage_model_specs, backend_options) into their own tach modules so
a narrow spec/options change no longer marks the full datamodel as
impacted.
- Narrow the per-suite pipeline path filters in checks.yml to the
concrete pipeline files relevant to each suite, so editing
vlm_pipeline.py only triggers the vlm matrix cell and editing
asr_pipeline.py only the asr one.
- Rekey the model cache in setup-ubuntu-ci to include runner.os and
hashFiles(uv.lock, pyproject.toml), with ordered restore-keys
fallbacks so a lockfile bump no longer silently stales the cache.
Metaxy parity note: layered tach enforcement (layer = "...") is
blocked by existing backend<->datamodel and utils<->stages cycles;
depot runners, nox dynamic matrices, devenv/nix, dprint and ty are
not applicable to docling's stack. All pinned action SHAs are on
their latest release as of this commit.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: introduce pipeline and orchestration tach layers
Earlier notes claimed layers were blocked. That was only true for the
cyclic core (backend<->datamodel, utils<->stages). The boundary
*above* core is clean:
- No module under docling/backend, docling/datamodel, docling/models,
docling/utils, docling/exceptions, or docling/chunking imports
anything from docling.pipeline (verified by grep).
- No module anywhere in docling/ imports from docling.cli,
docling.document_converter, docling.document_extractor, or
docling.service_client (also verified).
So we can introduce two real layers on top of the cyclic core:
- "pipeline" — docling.pipeline and all nine concrete pipelines
(base, simple, base_extraction, asr, vlm,
extraction_vlm, standard_pdf,
threaded_standard_pdf, legacy_standard_pdf).
- "orchestration" — docling.cli, docling.document_converter,
docling.document_extractor, and
docling.experimental.pipeline.
Unlayered modules stay "below" both layers (tach allows them to be
depended on freely) and continue to carry the declared-but-cyclic
backend<->datamodel and utils<->stages edges.
A VLM-only layer was explored but rejected: only
docling.pipeline.vlm_pipeline and docling.pipeline.extraction_vlm_pipeline
could be cleanly layered as "vlm", because the matching datamodel
options (pipeline_options_vlm_model, vlm_engine_options,
vlm_model_specs) and model stages (vlm_convert, vlm_pipeline_models)
sit inside the datamodel/models cycle and cannot be promoted to a
higher layer without first breaking that cycle. Layering only the
two pipeline files is not worth the extra config.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: expand tach layers to entrypoints/pipeline/models/core
Follow-up to the two-layer attempt. After verifying via grep that
nothing in datamodel/utils/backend imports from
docling.models.{extraction,factories,plugins,vlm_pipeline_models}
or from the "upper" stages (page_assemble, page_preprocessing,
reading_order, picture_description, vlm_convert), those nine
modules can be promoted out of the cyclic core into a dedicated
"models" layer.
The resulting order (highest first):
- entrypoints — cli, document_converter, document_extractor,
experimental.pipeline
- pipeline — docling.pipeline + the nine concrete pipelines
- models — model factories, extraction, plugins,
vlm_pipeline_models, and the five "upper" stages
- core — datamodel*, backend*, utils, exceptions, chunking,
models (base), models.utils, inference_engines.*,
the six "core stages" that utils cycles with
(chart_extraction, code_formula, layout, ocr,
picture_classifier, table_structure), and the
experimental.* and service_client modules
Rename the previous "orchestration" layer to "entrypoints" to
match the common docling vocabulary. Every module now carries an
explicit layer tag instead of relying on implicit unlayered
behaviour, so future additions must pick a layer deliberately.
A VLM layer, a stand-alone inference-engines layer, and separating
datamodel from backend all remain blocked by the bidirectional
backend<->datamodel and utils<->core-stages edges; those need a
code-level refactor first.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: refine tach client and foundation layers
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: add optional windows and macos smoke lanes
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: normalize reusable workflow boolean inputs
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: replace external all-green action
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: use org-allowed setup-uv action
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: install compiler toolchain for ML tests
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* DCO Remediation Commit for Georg Heiler <georg.kf.heiler@gmail.com>
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: bb714afb42cd1b29ab073a7f59cc72874ff2fdcd
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: a1f2761da8f72bfed636bd571ebf77b42c8771b6
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* DCO Remediation Commit for Georg Heiler <georg.kf.heiler@gmail.com>
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: cc6551b54c5bf4815ae9cd57cf43a98928a74be0
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: b21b0e7ca12b552dbdd54fac1bda113719c286f1
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: simplify ML pytest suite patterns
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: gate heavy examples on label, add job timeouts
- ci-heavy-examples: run only on main push, schedule, workflow_dispatch,
or when a PR is labeled tests:full / tests:heavy-examples. Drops the
path-based auto-trigger so that common edits to pyproject.toml,
uv.lock, or .github/actions do not kick off the 45-60min matrix on
every PR push. Collapses the changes job into a job-level if gate and
adds timeout-minutes: 90.
- checks.yml: add timeout-minutes to every job so stuck runners cannot
burn the full 6h default.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: tolerate cancelled allowed-skip jobs in check aggregator
Intentional cancellations (manual cancel, concurrency replacement) on
jobs that are already in ALLOWED_SKIPS should not mark the overall
workflow red. Treat `cancelled` the same as `skipped` when the job is
listed as an allowed skip; any unexpected cancellation of a required
job still fails.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* docs: make minimal vlm example portable
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* DCO Remediation Commit for Georg Heiler <georg.kf.heiler@gmail.com>
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: 2135051da3ed73d4b8a9130f584f40b56155af1a
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: 4f6d1d7960f7418d0cde6425ae61538da84fda40
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: install workspace packages in CI syncs
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* DCO Remediation Commit for Georg Heiler <georg.kf.heiler@gmail.com>
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: 492fa9883d4de6d98ebcb40fa863eafe2facff3c
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: 3eefae71643f9ca3df0264690c0c6eb1f67f06f1
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* DCO Remediation Commit for Georg Heiler <georg.kf.heiler@gmail.com>
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: fe8c9689a0ee94f36eb826da8e2177ef87404f5e
I, Georg Heiler <georg.kf.heiler@gmail.com>, hereby add my Signed-off-by to this commit: eabdd24a6734ec873cdaac857718aef2473677e7
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: remove unused graphite concurrency exception
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: document test labels and gate cross-platform lanes
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: select ml tests with pytest markers
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: fix marker selector typing
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: simplify ml suite scheduling
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: mark cross-platform smoke tests
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: reuse test trigger for ml matrix
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: tighten full ci aggregation
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: share required job result check
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
---------
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* ci: add phase 1 path-based workflow skipping
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: add fast pull_request_target lint checks
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: keep pr fast checks cheap
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: expand full matrix triggers
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: enable same-repo and merge queue checks
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: harden pull_request_target fetch inputs
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: address phase 1 workflow review
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: grant reusable checks permissions
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: temporarily enable pr fast checks validation
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: allow first run of pr fast checks
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: load pr fast check script for first validation
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: format pr fast check script
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: guard temporary pr fast check script fallback
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: use pr metadata for temporary fast check validation
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: remove temporary pr fast checks trigger
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: disable duplicate pull request runs
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: run fast pr checks without path trigger filter
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: add job timeouts in checks.yml
Cap every job so a stuck runner cannot burn the 6h default. Limits:
changes=5, lint=20, run-tests-1/2=45, run-examples=60,
test-pip-install-*=30, build/test-package=15.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: restore pull request workflow triggers
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* ci: run lint on pull requests
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
---------
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* fix(pptx): skip malformed picture shapes instead of aborting conversion
MsPowerpointDocumentBackend._handle_pictures reads embedded image bytes via python-pptx's shape.image accessor. On PPTX files with slightly malformed <p:pic> shapes, shape.image raises three exceptions that the existing (UnidentifiedImageError, OSError, ValueError) clause does not catch, so one bad picture aborts conversion of the entire presentation:
- InvalidXmlError when <p:blipFill> is missing
- KeyError when <a:blip r:embed> points to an unknown relationship
- AttributeError when the embedded part's content-type isn't an image
These files open normally in Keynote and Google Drive, so the backend should handle them as gracefully as it already handles truncated or unreadable image payloads.
This follows the same pattern as #2914, which extended the same except tuple with ValueError to handle linked (external) image references. The three cases above are the remaining shape.image failure modes that still escape.
Extend the except tuple to cover the three cases and log the same warning used for other unreadable images, leaving the rest of the presentation to convert normally. Add a regression fixture with one malformed picture per failure mode plus a focused test.
Fixes#3371
Signed-off-by: pateltejas <tejas226@hotmail.com>
* refactor(pptx): use warnings.warn for malformed picture skips
Address PR review feedback: use Python's warnings module with UserWarning to signal the skip to callers instead of logging.Logger.warning, matching the pattern used in msword_backend for "Skipping external image reference". This makes the skip visible via standard warning filters and catchable in tests.
Update the regression test to assert the warning is emitted via pytest.warns, which also suppresses the message during the test run so it doesn't clutter suite output.
Signed-off-by: pateltejas <tejas226@hotmail.com>
---------
Signed-off-by: pateltejas <tejas226@hotmail.com>
* chore: Update .gitignore with local dirs of AI agents
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* feat: Extend KserveV2OcrModel and kserve_v2_grpc.py to support the new version of Triton-RapidOCR
model where the language is the first input parameter:
- The gRPC client has been extended to encode BYTE input, needed for String types.
- An additional test ensures to have proper BYTE encoding/decoding.
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* feat: Add test for the KServe-Triton integration: WIP
- The test currently supports only the gRPC KServe client
- Extend the ground-truth test data.
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* fix: Simplify code in kserve test
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* chore: Rename test file
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* feat: Extend the kserve_v2 implementation to support binary data in the HTTP interface.
- Decouple functions for binary encoding/decoding inside the kserve_v2_utils.py and share for both HTTP and gRPC.
- Introduce use_binary_data init parameter in KserveV2OptionsMixin
- Improve tests
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* fix: Put back the field grpc_use_binary_data of KserveV2OptionsMixin as a deprecated alias to use_binary_data
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
---------
Signed-off-by: Nikos Livathinos <nli@zurich.ibm.com>
* fix(docx): handle unsupported limit functions gracefully in OMML conversion
Replace RuntimeError with graceful fallback for unknown limit functions in do_limlow().
Add argmax and argmin to LIM_FUNC dictionary for proper LaTeX rendering.
Fixes conversion failures when Word documents contain mathematical operators
not previously supported in the limit function dictionary.
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* test(docx): regenerate ground truth files
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
---------
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* feat(docx): add checkbox parsing support to MsWordDocumentBackend
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* refactor(docx): remove duplicate code in text element handling
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* docs(docx): update checkbox method docstrings
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* refactor(docx): use self._BLIP_NAMESPACES for w14 namespace in checkbox methods
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
---------
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
fix(html): preserve fragment-only anchor links during path resolution
Fragment-only hrefs (e.g. href="#section1") were resolved as filesystem
paths when source_uri was set, breaking internal document navigation.
Add '#' to the skip-resolution prefixes in _resolve_relative_path() so
fragment links pass through unchanged.
Partially addresses #2929
Signed-off-by: aatrey56 <aatrey.sahay@gmail.com>
* Add ResponseFormat.DOCLANG and parsing branch in VLM pipeline
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* fix: Remove bogus preamble from VLM chat template
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* fix: Add include_stop_str_in_output in allowed VLLM sampling params
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* allow vllm model_impl to be defined
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* make VLLM model_impl default to auto
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
---------
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* fix: prevent XXE and decompression bomb in METS-GBS processing
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* refactor: enforce resource limits for METS-GBS tar extraction
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
---------
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
feat(docx): Extract VML images with v:imagedata elements
Add VML image support with EMF/WMF conversion and consolidate image handler code.
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* Add explicit usage exceeded error handling
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Make websocket more resilient
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
---------
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
Add a new table structure model using IBM Granite Vision to extract table
structure from document images via OTSL token generation.
Changes:
- Add `GraniteVisionTableStructureOptions` with configurable model repo,
device, batch size, and crop padding options
- Implement `GraniteVisionTableStructureModel` that uses a VLM pipeline to
generate OTSL tokens from cropped table images, then parses them into
`TableData` with cells, rows, and columns
- Register the model in `table_structure_engines` alongside existing engines
- Add example script `docs/examples/granite_vision_table_structure.py`
- Add tests covering options, model enable/disable, OTSL parsing (including
self-closing tags xcel/srow/ecel), and invalid-backend error handling
- Update model catalog docs and CI workflow accordingly
Signed-off-by: Eli Schwartz <eli.shw@gmail.com>
* fix(latex): fully unwrap deeply nested formatting macros
Two related bugs when formatting macros are nested:
1. `\textcolor{color}{...}` extracted the color name alongside the
text content because `_nodes_to_text` fell through to the generic
else branch, which concatenates all arguments. E.g.
`\section{\textcolor{blue}{\textbf{[SEP]}}}` produced heading text
"blue [SEP]" instead of "[SEP]".
2. `\textsc`, `\textsf`, `\textrm`, `\textnormal`, `\mbox` and
`\textcolor`/`\colorbox` are listed in MACROS_STRUCTURAL, so when
encountered mid-sentence `_process_macro_node_inline` flushed the
text buffer and called `_process_macro`, which creates a new doc
node. This broke inline paragraphs into fragments.
Fix:
- Add explicit handlers for MACROS_TEXT_STYLE and textcolor/colorbox
in `_process_macro_node_inline` (before the MACROS_STRUCTURAL flush
path) so they are accumulated inline like MACROS_TEXT_FORMATTING.
- Add matching handlers in `_nodes_to_text` so colour names are
skipped and only the text-content argument is returned.
Fixes#3207
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Signed-off-by: Smeet23 <smeetagrawal2003@gmail.com>
* fix(latex): fully unwrap deeply nested formatting macros
Two related bugs when formatting macros are nested inside each other:
1. `\textcolor{color}{...}` extracted the color name alongside the
text content because `_nodes_to_text` fell through to the generic
else branch, which concatenates all arguments. E.g.
`\section{\textcolor{blue}{\textbf{[SEP]}}}` produced heading text
"blue [SEP]" instead of "[SEP]".
2. `\textsc`, `\textsf`, `\textrm`, `\textnormal`, `\mbox` and
`\textcolor`/`\colorbox` are listed in MACROS_STRUCTURAL, so when
encountered mid-sentence `_process_macro_node_inline` flushed the
text buffer and called `_process_macro`, which creates a new doc
node. This broke inline paragraphs into fragments.
Fix:
- Add MACROS_COLOR_INLINE constant for textcolor/colorbox to keep
all macro classifications in one place (constants.py).
- Add explicit handlers for MACROS_TEXT_STYLE and MACROS_COLOR_INLINE
in `_process_macro_node_inline` (before the MACROS_STRUCTURAL flush
path) so they are accumulated inline like MACROS_TEXT_FORMATTING.
- Merge the identical MACROS_TEXT_FORMATTING and MACROS_TEXT_STYLE
branches in `_nodes_to_text` into a single branch.
- Use argnlist[-1] instead of reversed() iteration for
MACROS_COLOR_INLINE since the text content is always the last arg,
consistent with _extract_macro_arg.
Fixes#3207
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Signed-off-by: Smeet23 <smeetagrawal2003@gmail.com>
* refactor(latex): extract _macro_node_to_text to reduce complexity
Split the macro-handling branch of `_nodes_to_text` into a dedicated
`_macro_node_to_text` helper so that cyclomatic complexity stays within
the ruff C901 limit (was 31, now < 30 for both methods).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Smeet23 <smeetagrawal2003@gmail.com>
* fix(latex): migrate nested-formatting test to tests/test_latex/
Upstream reorganised all latex tests from tests/test_backend_latex.py
into tests/test_latex/. Move test_latex_nested_formatting_macros to
tests/test_latex/test_macros.py and fix ruff-reported style nits.
Signed-off-by: Smeet23 <smeetagrawal2003@gmail.com>
---------
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Signed-off-by: Smeet23 <smeetagrawal2003@gmail.com>
Co-authored-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(docx) Handle inline formulas in list items
Fixes issue where inline formulas in list items were ignored during conversion.
Added helper methods to eliminate code duplication.
Updated test data with list items containing inline equations.
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* fix(docx): collect element refs in _add_inline_equations_to_parent
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
---------
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
fix(format): detect plain-text (.txt/.text/.qmd) as Markdown
Fixesdocling-project/docling#3259
Previously, .txt files were rejected with "File format not allowed"
because _mime_from_extension intentionally returns None for .txt
(the extension is ambiguous between XML_USPTO and MD), and the
text/plain branch in _guess_from_content only checked for the USPTO
PATN\r\n marker before giving up.
This change lets _guess_from_content fall back to InputFormat.MD for
text/plain content when the file extension is in the MD format's
extension list (md/txt/text/qmd/rmd/Rmd). Crucially, the fallback is
extension-gated: unknown extensions like .xyz continue to return None
and raise ConversionError as before, preserving the expected behavior
in tests/test_invalid_input.py.
Changes:
- _guess_from_content accepts an optional ext parameter
- _guess_format now captures the extension into obj_ext and passes
it into _guess_from_content
- test_input_doc.py asserts that .txt streams and files are detected
as InputFormat.MD (not None)
Signed-off-by: CrepuscularIRIS <serenitygp@qq.com>
* fix(pptx): handle NotImplementedError from shape.shape_type
python-pptx raises NotImplementedError from Shape.shape_type for
<p:sp> elements that aren't placeholders, autoshapes, textboxes, or
freeforms (e.g. shapes with empty <p:spPr> from Google Slides exports,
LibreOffice, or Keynote). handle_groups() and handle_shapes() access
shape_type without catching this, crashing the entire conversion.
Add a _safe_shape_type() helper that returns None on
NotImplementedError, so unrecognized shapes skip only the GROUP
recursion and PICTURE extraction while text and table extraction
proceed normally.
Fixes#3308
Signed-off-by: Tejas Patel <tejas226@hotmail.com>
* Fix lint
Signed-off-by: Tejas Patel <tejas226@hotmail.com>
---------
Signed-off-by: Tejas Patel <tejas226@hotmail.com>
* fix(ocr): support language selection for RapidOCR engine
Allows specifying 'english' or 'chinese' via the --ocr-lang flag and automatically downloads the correct models.
Signed-off-by: DevAbdullah90 <abdullahkashif12b3@gmail.com>
* fix(ocr): fix linting and add unit tests for RapidOCR language selection
Signed-off-by: DevAbdullah90 <abdullahkashif12b3@gmail.com>
* fix(ocr): align RapidOCR english assets with 3.8 mobile models
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* fix(ocr): restore RapidOCR default model compatibility
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* fix(examples): disable OCR in code formula comparison
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
---------
Signed-off-by: DevAbdullah90 <abdullahkashif12b3@gmail.com>
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
Co-authored-by: DevAbdullah90 <abdullahkashif12b3@gmail.com>
* fix(docx): isolate list state in table cells
Lists with the same numId in different table cells were incorrectly
merged. Added context manager to isolate list state during cell
processing. Includes test cases and updated ground truth files.
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* style(docx): modernize type hints to use PEP 604 union syntax
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
---------
Signed-off-by: Cesar Berrospi Ramis <ceb@zurich.ibm.com>
* feat: explicitly handle tikzpicture as atomic figure nodes
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* fix: resolve copilot review on recursion depth and text labels
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* test: cover tikz fallback label and depth guard
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* fix: address copilot tikz parent and end-tag edge cases
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* feat(latex): store TikZ source in Picture meta code field
Handle tikzpicture environments as real Picture nodes via add_picture and
store raw TikZ in PictureMeta.code (CodeMetaField) with language metadata.
Also update LaTeX backend tests to assert picture meta code behavior and
bump docling-core minimum version to 2.71.0 (includes PictureMeta.code).
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* fix(latex): make TikZ code metadata compatible across docling-core versions
Avoid hard import of CodeMetaField to prevent import-time crashes when older
docling-core is resolved. Use PictureMeta.code when available and retain legacy
CODE fallback otherwise. Update TikZ tests for dual-path behavior and refresh
uv.lock to pin docling-core 2.71.0.
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* refactor(latex): simplify TikZ parent assignment
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* fix(latex): remove legacy PictureMeta guard and use TIKZ code metadata
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* fix: require docling-core 2.73.0 for TIKZ code label
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* test: remove legacy tikz code-meta compatibility branch
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
---------
Signed-off-by: StealthTensor <rayabharaputrinai@gmail.com>
* Move client SDK to docling
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Add client SDK examples
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* Mark client SDK as experimental
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
---------
Signed-off-by: Christoph Auer <cau@zurich.ibm.com>
* feat(vlm): add nanonets ocr2 onboarding
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* feat(vlm): add vLLM and API runtimes for Nanonets-OCR2
Extend the Nanonets-OCR2 preset with vLLM + remote API paths so all
standard docling runtimes (Transformers, MLX, vLLM, API, LM Studio,
OpenAI-compatible) work out of the box. Drop the restricted
supported_engines set to match the GLM-OCR / LightOnOCR / Falcon-OCR
pattern, add top-level torch_dtype on the Transformers override, and
register NANONETS_OCR2_VLLM / NANONETS_OCR2_VLLM_API /
NANONETS_OCR2_LMSTUDIO_API legacy specs plus VlmModelType enum entries.
Folds in the remote-API scope that was on the superseded PR #3275.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
---------
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix(pdf): extend ligature map with Dutch IJ and PUA glyph U+F0A0
Add two entries missing from the PDF text sanitizer's ligature map:
- U+0132 (IJ) → "IJ" and U+0133 (ij) → "ij": Latin capital/small ligature
IJ, used in Dutch (e.g. IJssel, IJ becomes IJ at the start of words).
- U+F0A0 → "": a Private-Use Area glyph emitted by some PDF fonts as a
spurious character with no textual meaning; it is silently discarded.
The _LIGATURE_RE pattern is updated to match these new code points.
Closes#2882
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
* style: apply ruff formatter fixes
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
* fix: remove accidentally included msexcel tests from ligature branch
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
---------
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Co-authored-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Commands like \vspace{-1mm} and \hspace{0.2cm} were being filtered
at the command level but their argument values were leaking through as
plain text nodes. Ensure that when a spacing/ignored command is
encountered, its arguments are also suppressed.
- Add vspace, hspace, vspace*, hspace*, addvspace to MACROS_SPACING
- Guard against spacing/ignored macros in _process_macro_node_inline
so their brace arguments are not extracted as inline text
- Guard against spacing/ignored macros in _nodes_to_text so dimension
values do not leak when processing footnotes, captions, etc.
- Update ground truth files to reflect corrected output
Fixes#3240
Signed-off-by: Smeet Agrawal <smeetagrawal23@gmail.com>
Co-authored-by: Smeet Agrawal <smeetagrawal23@gmail.com>
* feat(vlm): add PARTIAL_SUCCESS status for VLM pipeline pages
Override _determine_status in VlmPipeline to detect partial failures
from VLM inference. Pages with truncated output (LENGTH) or filtered
content (CONTENT_FILTERED) now correctly report PARTIAL_SUCCESS with
descriptive error messages, matching the pattern used by
ExtractionVlmPipeline and AsrPipeline.
Includes @override decorator per coding conventions and 7 unit tests
covering all VlmStopReason variants.
Closes docling-project#2583
Signed-off-by: aatrey56 <aatrey.sahay@gmail.com>
* style(vlm): fix import ordering for typing_extensions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: aatrey56 <aatrey.sahay@gmail.com>
---------
Signed-off-by: aatrey56 <aatrey.sahay@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add support for Falcon-OCR (tiiuae/Falcon-OCR)
Add Falcon-OCR as a VLM convert model for document OCR and markdown
conversion. Falcon-OCR is a 0.3B parameter model from TII that supports
full-page text extraction, formula (LaTeX), and table (HTML) recognition.
Changes:
- Add VLM_CONVERT_FALCON_OCR preset (stage_model_specs.py)
- Add legacy specs: FALCON_OCR_TRANSFORMERS, FALCON_OCR_VLLM,
FALCON_OCR_VLLM_API (vlm_model_specs.py)
- Register preset with VlmConvertOptions (pipeline_options.py)
- Fix trust_remote_code propagation from VlmModelSpec to engine options
in auto_inline_engine.py and from_preset(), which also fixes the
existing Phi-4 preset
- Add test suite (tests/test_falcon_ocr_vlm.py)
Closes: docling-project/docling#3236
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* style: apply ruff formatting
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
* fix: align Falcon-OCR presets with review feedback
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
---------
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
feat: add support for LightOnOCR-2-1B VLM model
Add LightOnOCR-2-1B as a new VLM model option for OCR and markdown conversion.
Includes transformers, vLLM, and API configurations, stage model preset,
and tests.
Signed-off-by: Georg Heiler <georg.kf.heiler@gmail.com>
Maksym Lysak
David Wallace
benvizel
Brighton
Christoph Auer
geoHeil
Panos Vagenas
Qiefan Jiang
Cesar Berrospi Ramis
pateltejas
Nikos Livathinos
Aatrey Sahay
Matvei Smirnov
EliSchwartz
Smeet Agrawal
Yarizakura
Hemanth Battu
Rayabharapu Trinai
Aditya Sasidhar
Said Gürbüz
vwe-ibm
Smeet Agrawal