Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA

Organizations face a significant challenge in extracting valuable, negotiated data from agreements. Despite careful negotiation, crucial information like pricing tiers, SKUs, and service level agreements (SLAs) remains locked within unstructured documents such as PDFs, often across disconnected systems. This trapped data represents an aggregate value of approximately two trillion dollars that companies fail to capitalize on due to the manual effort required for human review and extraction.

DocuSign operates at a massive scale, processing about one million agreements daily for 1.9 million paying customers and a billion users. This sheer volume necessitates an AI-first approach to intelligent agreement management, spanning the entire agreement lifecycle from creation to post-signing insights. A primary obstacle in this process is the extraction of data from tables, which often defeats traditional document extraction tools and generic Vision Language Models (VLMs). These tools struggle with the complexities of tables, such as merged cells, nested columns, and non-standard boundaries, leading to inaccurate data extraction and significant operational overhead for legal, procurement, and sales teams.

To address this, DocuSign partnered with NVIDIA to leverage purpose-built models. Sean Sodha from NVIDIA introduced the NeMo Retriever initiative, which focuses on building world-class open-source models, datasets, and techniques for retrieval, including embedding, reranking, and document extraction. The core problem addressed with DocuSign is finding the right information within a document, specifically tackling complex table extraction.

The solution is the NeMo Parse model, an unusually small (approximately 850-900 million parameters) vision language model. Unlike typical multi-billion parameter VLMs, NeMo Parse is designed as an extractor, not a generator. It functions as an “all-in-one package,” taking a document and, in a single shot, returning semantic formatting, layout, text, reading order, and crucially, preserved table structure. This approach replaces a stack of separate layout and table models, offering a streamlined and efficient solution.

Key benefits of NeMo Parse observed in testing include:

• Accuracy: Capable of handling the vast permutations of tables, including nested structures and merged cells/rows/columns.
• Efficiency: Its smaller size and lower context requirement translate to significantly lower latency and cost, which is critical for processing millions of agreements daily.
• Speed: NeMo Parse demonstrated a 20x faster table extraction rate compared to general open-source alternatives, allowing DocuSign to process agreements at an unprecedented scale.

A demo showcased DocuSign’s “Agreement Manager” platform, where uploading an order form immediately triggers AI processing. The platform, powered by NVIDIA’s model, extracts a rich set of structured metadata, including key terms, commercial details, and granular pricing data from dense tables. This data is then instantly usable, exportable to CSV, or accessible via API, transforming agreements into actionable insights for various business functions.

The collaboration’s key learnings emphasize the value of purpose-built models for specific tasks, noting their ability to accelerate market delivery. The importance of model efficiency for large-scale enterprise applications was highlighted, where lower context directly impacts latency and cost. Looking ahead, the partnership aims to deepen, focusing on enhancing the NeMo Retriever efforts to improve document discovery, integrate with the NVIDIA Agent Toolkit, and scale into production-grade AI agents.

In a Q&A session, the speakers addressed the future of OCR, suggesting a hybrid approach where pre-processing (including OCR) remains crucial for querying vast datasets and enabling downstream integration with systems like procurement platforms. While real-time, low-latency Q&A on single documents might evolve, the need for structured, queryable data at scale necessitates upfront computation. They also noted that while the model currently runs on FP16, there are plans for further quantization to FP8 and NVFP4, and potential optimizations for multi-token generation to enhance performance, always prioritizing accuracy first.

• Tags:
• Feature
• NVIDIA NeMo
• Model Optimization
• Small Language Models
• Vision Language Models
• Document AI
• Information Extraction
• Agreement Management
• Table Extraction
• DocuSign
• Enterprise Solutions
• Unstructured Data Processing