Add NemotronParseBBox.to_original_coordinates for raw nemotron-parse output

[jamesbraza]

Why

nemotron-parse predicts bounding boxes in the space of its preprocessed input: each page is resized (aspect-preserving) and centered onto a target 2048x1648 (HxW) canvas with white padding, so the boxes are normalized to that padded canvas rather than to the page.

NVIDIA's hosted NIM undoes that and returns input-image coordinates, which is why PaperQA's current decode is correct against NIM. A consumer of the model's raw output, though, e.g. a self-hosted vLLM or HuggingFace transformers endpoint, receives canvas-space coordinates and draws every box shifted inward and compressed, worst at low DPI where the padding dominates. This adds a first-class way for such consumers to recover original-image coordinates, mirroring what NVIDIA's postprocessing.transform_bbox_to_original and example.py do.

Notes

• Opt-in, and intentionally not used on the existing NIM path. NIM already returns input-image coordinates, so applying this there would double-transform. to_original_coordinates sits alongside to_page_coordinates for the raw-output case only.
• Relation to Upsizing PDF pages to match nemotron-parse training data #1271: that (closed) PR addressed the same mismatch from the input side, upsizing/pre-fitting pages to the training aspect ratio before sending. This is the complementary output side, and leaves the request path untouched.
• Tests assert parity against NVIDIA's reference across the no-resize, downsized, and width-resized regimes, plus round-trip recovery and a low-DPI heavy-padding case.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a first-class conversion utility for consumers of raw nemotron-parse outputs by inverting the model’s resize + centered-padding preprocessing, so canvas-normalized bounding boxes can be mapped back into original image pixel coordinates (matching NVIDIA’s reference behavior). This complements the existing to_page_coordinates path (used for NIM, which already returns input-image coordinates).

Changes:

• Add NEMOTRON_PARSE_TARGET_HEIGHT/WIDTH constants for Nemotron-Parse’s native input canvas size.
• Add NemotronParseBBox.to_original_coordinates() to convert padded-canvas normalized bboxes to original-image pixel coordinates.
• Add tests asserting parity with NVIDIA’s reference transform across multiple resize/padding regimes and a round-trip recovery check.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
packages/paper-qa-nemotron/src/paperqa_nemotron/api.py	Introduces target canvas constants and to_original_coordinates() bbox back-projection.
packages/paper-qa-nemotron/tests/test_api.py	Adds parity and round-trip tests for the new coordinate transform.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.