NVIDIA is pitching a new wave of document-intelligence systems built on its open Nemotron models to turn dense corporate paperwork into live, queryable data. The company says the stack—combining parsing, embeddings and reranking models, packaged as NIM microservices on GPUs—can extract tables and figures, ground answers with citations and scale across large and changing repositories. Early adopters include Justt, which uses Nemotron Parse to automate chargeback evidence assembly and optimize dispute decisions; Docusign, evaluating the models to deepen contract understanding and reduce manual corrections; and Edison Scientific, which is using the tools to accelerate literature synthesis in science. NVIDIA points to strong benchmark showings on MTEB, MMTEB and ViDoRe V3 and advocates routing tasks across open and frontier models to balance accuracy and cost. The push underscores how AI agents are moving document stores from static archives to auditable knowledge systems for finance, legal and research workloads.
Related articles:
— Google Cloud Document AI: Turning Unstructured Documents Into Structured Data
— Amazon Textract: Automatically Extract Text and Data From Scanned Documents
— Retrieval‑Augmented Generation (RAG) Explained
— Justt: Chargeback Automation for Merchants
