docling

Docling's enrichment models convert RGBA images with PIL's convert("RGB"), which keeps the color stored under the alpha channel (usually black), so transparent areas become black instead of white. This affects document_picture_classifier.py:167, picture_description_base_model.py:106, and inference_engines/vlm/_utils.py:131. Fix: replace convert("RGB") at those three sites with _to_rgb from docling/backend/image_backend.py (added in #4511), which flattens the image onto a white background. As a workaround in your own code, composite RGBA onto white before passing images to the models, e.g. bg = Image.new("RGB", img.size, (255,255,255)); bg.paste(img, mask=img.split()[3]). Note page images already go through _to_rgb; only pictures embedded in DOCX/HTML/PPTX via prepare_element hit the buggy path.

This is a regression in Docling 2.132–2.134 (commit 42de1e2): the reading order model generates an unbounded number of page-separator candidates (e.g. ~195K for a concert-hall floor plan where each seat is a square), and the candidate merge is quadratic, so conversion effectively never finishes. The document_timeout budget was also not checked inside this phase, so the pipeline blocked forever.

Fix: upgrade to a Docling version including PR #4633. The fix (1) caps separator candidates at 5,000 per page — pages above the cap skip separator extraction with a warning and fall back to default reading order — and (2) the candidate loop and merge now abort when the remaining document_timeout budget is exhausted, recording a TIMEOUT error so conversion returns with status PARTIAL_SUCCESS instead of hanging.

Workaround if you cannot upgrade: pin to the last unaffected version with pip install docling==2.124.

Root cause: Docling's TesseractOcrModel caches one tesserocr PyTessBaseAPI reader per pipeline, and tesserocr is not thread-safe. When conversions run concurrently (settings.perf.doc_batch_concurrency > 1, or one DocumentConverter called from several threads), simultaneous SetImage/SetRectangle/GetUTF8Text calls corrupt each other's state. The process then dies with SIGSEGV (exit code 139), raises cysignals.signals.SignalError: Segmentation fault, or silently returns status success with OCR text missing.

Workarounds:

1. Force sequential batch conversion before converting:

from docling.datamodel.settings import settings
settings.perf.doc_batch_concurrency = 1

2. If you call Docling from your own threads, do not share one DocumentConverter. Either create a separate DocumentConverter per thread (each builds its own pipeline and tesserocr readers), or serialize all convert()/convert_all() calls with a threading.Lock.

3. Upgrade docling to the latest release and check issue #4940 for the upstream fix status.

Note: Python cannot catch the SIGSEGV, so one crashing document kills every conversion in the process — keep concurrency at 1 until a thread-safe fix is released.

Docling (up to 2.141.0) pins pillow>=10.0.0,<13.0.0 in four packages: docling-slim, docling-core, docling-parse, and docling-ibm-models. This blocks Pillow 13.0.0, which fixes 8 CVEs flagged by pip-audit (CVE-2026-112895, CVE-2026-112400, CVE-2026-112397, CVE-2026-112896, CVE-2026-112897, CVE-2026-112395, CVE-2026-112404, CVE-2026-112899).

Workaround: force Pillow 13 with a dependency override. With uv, add an override to your project or use an overrides file:

uv pip install docling --overrides overrides.txt   # overrides.txt contains: pillow==13.0.0

[tool.uv]
override-dependencies = ["pillow>=13.0.0"]

This is a bug in Docling's CSV backend: _sniff_dialect() in docling/backend/csv_backend.py detects the delimiter from the first line (header) only. If a header column name contains a comma (e.g., Weight (kg, net)), Python's csv.Sniffer picks , instead of ;, corrupting every cell. This commonly affects files Excel exports in European locales where ; is the column separator and , is the decimal mark. Workarounds: (1) Remove commas from header column names before converting (e.g., rename Weight (kg, net) to Weight (kg net)) — the same file is then split correctly on ;. (2) Pre-process the file: sniff the dialect using the full file content (csv.Sniffer().sniff(full_text, ',;\t|:')) and rewrite it, or convert the CSV to XLSX first and convert that with Docling. (3) Update Docling once the linked fix PR for issue #4895 is merged, which makes dialect detection use multiple rows for consistency.