mag-pdf
Copyright (c) Magure

This product includes software developed at Magure, licensed under the
Apache License, Version 2.0. See LICENSE.

================================================================================
Third-party software
================================================================================

liteparse
  The sole runtime dependency of the base package. A Rust core with PyO3
  bindings, distributed as a binary wheel. Its manylinux wheel statically
  compiles Tesseract OCR (its CycloneDX SBOM records `tesseract-rs 0.2.0`) and
  bundles libpdfium. Consult the liteparse distribution for its own licence and
  for the licences of the components it vendors, including Tesseract (Apache-2.0)
  and PDFium (BSD-3-Clause).

docling-ibm-models
  Used by the optional `tables` extra to run TableFormer. Not vendored; it is a
  declared dependency and carries its own licence. Note that `docling` itself is
  NOT a dependency of this package.

================================================================================
Model weights
================================================================================

The `tables` extra downloads model weights at runtime, and the published Docker
image BAKES THEM IN - which means the image redistributes them. Attribution and
licence status of each:

TableFormer
  Source:   https://huggingface.co/docling-project/docling-models
  Revision: v2.3.0 (pinned)
  Files:    model_artifacts/tableformer/{fast,accurate}
  Licence:  the model card declares CDLA-Permissive-2.0 and Apache-2.0.
            Both permit redistribution.

DocLayout-YOLO page gate
  Source:   https://huggingface.co/Nexaaman/kb-yolo-gate
  Revision: c4ff7dae4fb150292b8ae26f93175dc3640aac5d (pinned)
  File:     yolo_gate.onnx
  Licence:  *** NOT DECLARED. ***
            As of this writing that repository states no licence, no licence
            tag, and no upstream lineage. It is a personal namespace, not an
            organisation account.

            Two consequences follow, and neither is resolved here:

            1. No licence declared is not the same as permissively licensed.
               Absent a grant, redistribution rights are not established, and
               the Docker image redistributes this file.
            2. DocLayout-YOLO models in general derive from YOLOv10, which is
               AGPL-3.0. If this ONNX export shares that lineage, copyleft terms
               would attach. The lineage has NOT been verified either way.

            Resolving this means confirming the export's provenance with its
            author and either publishing an explicit licence on that repository
            or mirroring a model whose terms are known. `TablesConfig.yolo_repo`
            and `TablesConfig.yolo_revision` exist precisely so a different
            mirror can be substituted without a code change.

            This was raised before publication and the risk was accepted by the
            image publisher. It is recorded here so it is not rediscovered as a
            surprise.
