Press Space to continue
Press Space to continue
Most LLMs sound confident whether they are right or wrong. That's fine until it's a mortgage form or a health claim, and the "confidently wrong" number costs someone real money. @nutrientdocs 's answer: don't just extract the value, extract the proof. Every field comes with a bounding box, a confidence score, and a page index. Hover it, see the exact pixels it came from. No hunting through the doc to check if the model made it up. They open-sourced the grounding model, published the…
jdrhyne@jdrhyneJul 14
Everyone keep talking about how LLMs are good enough now. The LLMs have always sounded confident. I don't buy it. After years running and building document workflows @nutrientdocs (formerly @PSPDFKit), the breakage is in the numbers. A model confidently hallucinates answers from a filing table it never fully saw because a 512-token window chopped the table in half. So we open-sourced grounding-en, a 0.4B model on @huggingface that scores whether a claim is actually entailed by the source. On our open benchmark our two models sit at #1 and #2. The real gap is number grounding: .92 AUC vs .45–.66 from the usual NLI models and we run 8,192 tokens so the table stays intact. Apache-2.0, with a free demo and a benchmark anyone can run. A document workflow gets trustworthy not when the model sounds confident, but when the claim can be checked. If it can't be grounded, it doesn't ship.