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

This post details the development of a purpose-built, small vision language model (900M parameters) for extracting structured data, specifically tables and semantic structure, from legal agreements. The authors highlight the trade-off between model size and performance, demonstrating that a smaller, specialized model achieved ~20x faster table extraction than general alternatives. The model's design as an 'extractor' rather than a 'generator' is emphasized, along with its ability to handle complex document relationships and preserve table integrity. The post also touches on the scale of the problem (millions of agreements daily) and the engineering challenges of processing such volumes.