nutrient-document-decision · commercial

A decision model for documents. In the spirit of general decision-making systems like Jev, nutrient-document-decision reads a document once and answers typed, calibrated questions about it — yes/no, pick-one, or scored — rather than generating free text. It's built with a particular focus on document and multimodal understanding: grounding (is a claim actually supported by a source), document classification (open-vocabulary, image+OCR), and document-split (page-stream boundary detection) are the tasks we benchmark today, not the limit of what the typed-decision interface can express — it has also shown promising zero-shot generalization to adjacent judgment tasks it was never trained on, such as judging which of two OCR outputs is more accurate, and combined classification-plus-split questions asked together in one pass.

• 🎯 Try it: nutrient-document-decision-demo
• 🏆 Leaderboards: grounding · document classification · open-vocabulary classification · document split

Results

Every task is scored against: this model, the existing Nutrient specialist purpose-built and separately trained for that one task, and a general-purpose vision-language model run zero-shot (no fine-tuning, no task-specific calibration) — the floor any document-AI product has to clear.

Grounding

	nutrient-document-decision	specialist (grounding-en / grounding-multilingual)	general-purpose VLM (zero-shot)
en AUC	0.9231	0.8815	0.8306
multi AUC	0.9485	0.9652	0.8748

Document classification (image + OCR)

Track	nutrient-document-decision	specialist (document-classification-v2)	general-purpose VLM (zero-shot)
doclaynet macroF1	0.914	0.968	0.814
forms macroF1	0.997	1.000	1.000
OOD macroF1 (unseen doc types)	0.632	0.946	0.795
OOV macroF1 (unseen label wording)	0.812	0.830	0.810
Tobacco3482 macroF1	0.940	0.735	0.888

Open-vocabulary / figure classification

Track	nutrient-document-decision	specialist (doc-img-classification)	general-purpose VLM (zero-shot)
broad (top1, ~48-label taxonomy)	0.911	0.880	0.744
specialized (in-domain) top1	0.827	0.713	0.492
synonym (label-wording robustness) top1	0.833	0.730	0.740

Document split (page-stream segmentation)

Stream	nutrient-document-decision	specialist (doc-split-v2)	general-purpose VLM (zero-shot)
our200 F1	0.962	0.944	0.792
OpenPSS-short F1	0.683	0.652	0.369
OpenPSS-long F1	0.959	0.891	0.482

This model mostly beats or ties the specialist across grounding, document-split, and figure classification, and trails on a few document-classification tracks — notably OOD (unseen document types) and multilingual grounding, where a purpose-built specialist still has an edge. See the leaderboards linked above for the full, independently-reproducible picture against other systems.

Intended use & limits

Built as a decision model for document-heavy workflows (contracts, filings, forms, mixed-format page streams) where a single calibrated pass needs to answer several different typed questions about the same document — not limited to the four tasks benchmarked above. Not a general-purpose chat or reasoning assistant — it reads out typed judgments (yes/no, label choice, numeric score) rather than generating free text, and its judgment on domains very different from its benchmark coverage (e.g. truly novel, unseen document types) should be validated before relying on it unsupervised.

License & data

Commercial. Benchmarked against publicly available document-AI evaluation sets (see the linked leaderboards and benchmark datasets for exact sources and licenses per track).

📩 Get access

nutrient-document-decision is commercial and its weights are not downloadable here. To run it on-prem — contact Nutrient: nutrient.io/contact-sales.

About the author

This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.