vera-ingest-docling

Optional Docling ingest pipeline for VERA. Registers the docling provider (default variant hybrid) under the vera.ingest_pipelines entry-point group.

The pipeline uses Docling's DocumentConverter for layout-aware PDF parsing and HybridChunker for token-aware chunks. Readable chunk text is stored for keyword search; contextualized text from HybridChunker.contextualize() is used for embeddings.

Install

python -m pip install "vera-ingest-docling>=0.3.0"

From a repository checkout with uv (workspace .venv for CLI and tests):

uv sync --extra docling

Non-desktop users can install the CLI extra:

pip install "vera-cli[docling]>=0.3.0"

Python 3.10 or newer is required. The package depends on Docling's rapidocr extra so RapidOCR and onnxruntime are installed for OCR. RapidOCR ONNX weights come with that extra; first conversion may still download Docling layout models (about 380 MB: Heron ONNX + TableFormer accurate). Set DOCLING_ARTIFACTS_PATH to a local cache (or prefetch layout models offline) for air-gapped runs. Incomplete caches are not treated as ready. Hub progress is visible on CLI stderr. This extra is not bundled in the 0.3.0 desktop installer and is not listed in Convert.

Usage

vera convert "manual.pdf" "manual.vera" --parser docling
# or explicitly:
vera convert "manual.pdf" "manual.vera" --parser docling:hybrid

from vera_ingest import convert

convert("manual.pdf", "manual.vera", parser="docling")

Notes

• Defaults: chunk_size=500 whitespace tokens (not LLM subword tokens), ocr_mode=auto, ocr_language=en, pdf_backend=docling_parse. Docling does not advertise overlap or ocr_dpi, so those legacy convert/CLI aliases are not forwarded. The Tesseract --ocr-language alias is also not forwarded; Docling keeps en.
• Prefer pipeline_options= / --pipeline-option KEY=VALUE for provider-owned settings; --chunk-size and --ocr* remain compatibility aliases for pipelines that accept them.
• OCR modes map to Docling/RapidOCR: off, auto (default), and force (full-page OCR). ocr_language expects a RapidOCR-native code (en, fr, cyrillic, ...) — Tesseract-style codes such as eng are not translated. The shared --ocr-language CLI default (eng) is not forwarded; pass --pipeline-option ocr_language=fr (or another RapidOCR code) when you need a non-default language.
• Torch model compilation is disabled so Windows does not need MSVC cl.exe.
• Picture crops are stored as figure attachments and linked onto a nearby same-page chunk so search --figures can return them. Docling's HybridChunker omits pictures from chunk text.
• On page-level memory errors (bad_alloc), VERA retries failed pages then falls back to whole-document pypdfium2, then page-batch pypdfium2 if that still raises. Force the backend with --pipeline-option pdf_backend=pypdfium2. Conversion rejects only when recovery is exhausted. Failures include the underlying exception and print it on sidecar stderr.

See the vera-ingest-docling documentation and conversion guide.