Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA
Source
youtube.com
Author
AI Engineer
Date
Context
Making contracts searchable requires preserving how their tables connect prices, products and conditions. Reading a page as successive lines can lose the relationships expressed by merged cells and nested columns. Docusign's presentation with NVIDIA describes a roughly 900-million-parameter vision-LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → built to extract that structure.
The model is designed to return reading order, document structure and preserved tables in one pass. That gives downstream systems structured agreement data to query, including terms whose meaning would be lost in a flattened text extraction.
Key takeaways · AI-distilled
Docusign processes about a million agreements a day across nearly two million paying customers, and agreements routinely reference each other, so answering a simple question can mean traversing twenty years of related documents.
The specific failure mode is tables: pricing tiers, rate cards, and service levels are exactly the terms people need, and generic extraction reading a page line by line destroys the merged cells and nested columns that hold them.
Docusign and NVIDIA built a roughly 900-million-parameter vision-language model as a pure extractor, not a generator, replacing a stack of separate layout and table models with one pass returning reading order, structure, and preserved tables.
Their small, purpose-built extractor ran table extraction about 20 times faster than the general-purpose alternatives they tested, with the smaller context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → also cutting latency and cost at their processing volume.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
videoVoice AI Where One Wrong Question Breaks the Rules - Shiv Indap, Upstart & Brooke Hopkins, Coval
video78% Wanted This AI Feature. Almost Nobody Used It - Supriya Subudhi, Bumble
videoWhen Running Your Own AI Models Beats Paying for APIs - Steven Latré, imec
videoHow to Decide What Your AI Agent Should Remember - Wojciech Pluta, Oracle
