langchain-docling-rs

A LangChain document loader backed by docling.rs, a Rust port of IBM's docling: PDF with layout analysis, table structure and local OCR, DOCX, PPTX, XLSX, HTML, EPUB, e-mail, scans and 40-odd other formats, into LangChain Documents — one per input as Markdown, or one per chunk with docling's chunk metadata (items, page and bounding-box provenance, headings). No PyTorch; the models run through ONNX Runtime from a single native wheel.

It is langchain-docling's API, so the import is the only change:

pip install langchain-docling-rs

# was:  from langchain_docling import DoclingLoader, ExportType
from langchain_docling_rs import DoclingLoader, ExportType

docs = DoclingLoader(file_path=["https://arxiv.org/pdf/2408.09869", "notes.docx"]).load()
docs[0].page_content                      # heading path + chunk text, ready to embed
docs[0].metadata["dl_meta"]               # docling's chunk meta: items, page/bbox provenance, headings

DoclingLoader(file_path="report.pdf", export_type=ExportType.MARKDOWN).load()  # one Document per input

The package has no code of its own: it pins the docling-rs wheel (with its langchain extra) and re-exports docling_rs.langchain — DoclingLoader, ExportType, MetaExtractor, BaseMetaExtractor, ConversionBackend and PictureDescriptionLangChainOptions (describe document pictures with any LangChain chat model during conversion). The parameters, defaults and chunking are documented there; pip install "docling-rs[langchain]" and from docling_rs.langchain import … is the same thing under the engine's own name.

PDF models

Office, HTML, Markdown and e-mail need no models. PDFs and images use ONNX models for layout, tables and OCR (~720 MB), fetched once with python -c "import docling_rs; docling_rs.download_models()". Without them a PDF converts from its embedded text layer.

License

MIT. docling.rs is an independent project of the docling community (LF AI & Data); not a LangChain partner package.