{ "video": { "id": "_gvamfT8H-w", "title": "Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA", "duration": 1102, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "_gvamfT8H-w", "title": "Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA", "one_liner": "Docusign processes a million agreements a day and partnered with NVIDIA to use Nemotron Parse — an ~850–900M-parameter single-shot VLM extractor — to turn the pricing tables, SLAs and rate cards trapped in those contracts into structured, queryable data 20x faster than other open-source models.", "summary": "Hiral Shah (Docusign) and Sean Sodha (NVIDIA) argue that enterprise agreement data is a database nobody can query: Docusign has 1.9M paying customers, a billion users and processes a million agreements a day, while a Deloitte study they cite puts $2 trillion of negotiated value inside agreements that nobody ever capitalizes on. The hard part isn't text, it's tables — pricing tiers, SKUs, SLAs and rate cards — where traditional extraction tools and generic VLMs fail because they read line by line and break merged cells and nested structure. Docusign uses NVIDIA's Nemotron Parse, a tiny single-shot vision-language extractor, inside its Agreement Manager product, demoed live turning an uploaded order form into highlighted metadata, structured order details, a CSV download and an API. Their stated lesson is that small purpose-built models beat big general ones at enterprise scale: lower context means lower latency and lower cost, and Nemotron extracted 20x more tables per second than the open-source alternatives they tested.", "key_points": [ "Scale framing: 1.9 million paying Docusign customers, a billion users, a million agreements processed per day; a Deloitte study they cite finds $2 trillion of negotiated value captured in agreements that nobody goes back and extracts, because it requires human reading, disconnected systems and manual workflows.", "Agreements aren't flat, they're hierarchical — one agreement governs another — so answering a simple question like 'what did we contract for in total tokens with Claude?' requires reasoning across a corpus spanning 10–20 years of business, not one document.", "The failure mode is tables specifically: 'vital terms' (pricing tiers, SKUs, SLAs, rate cards) live in tables, and traditional document extraction tools and generic VLMs read text line by line, which breaks merged cells, nested tables, merged rows/columns and unbounded cell structure.", "Nemotron Parse is a very small C-RADIO-based VLM at roughly 850–900 million parameters — not a multi-billion-parameter heavy VLM — that replaces a stack of small task models (e.g. YOLOX for table or page-element detection) with one single-shot pass emitting semantic formatting, layout, text, reading order and preserved table structure. It's an extractor, not a generator, served via NVIDIA NIM or vLLM.", "NVIDIA's Nemotron Retriever line covers embedding, reranking and document-extraction models, open-sourcing weights, datasets, quantization/distillation/pruning techniques, blueprints and agent skills; the team cites leaderboard wins on ViDoRe V1/V2/V3, MTEB and MMTEB. Table extraction was benchmarked on RD-TableBench against popular open-source models, tracking the Pareto curve of accuracy versus performance.", "Live demo of Docusign Agreement Manager: upload an order form, the jobs engine processes it with AI in seconds, rich metadata and commercial terms are extracted, structured and highlighted with click-to-jump to the source section; NVIDIA's model breaks the dense pricing tables into order details, downloadable as CSV for finance/procurement and available through an API.", "Measured results they state: Nemotron was 20x faster than other open-source models on tables extracted per second, and lower context meant lower latency and lower cost to deliver at their scale.", "Architecture is deliberately hybrid, not fully agentic: tables go through layout-aware extraction while other fields, metadata and general text still go through OCR — two routes in one pipeline. Pre-processing is required because querying thousands of documents reactively won't meet latency, and downstream systems (e.g. pushing a pricing table into Coupa) need bulk structured data, not one-document-at-a-time agent work. A blog details the at-scale approach.", "Roadmap: the model is FP16 today with paths to FP8 and NVFP4 in the coming months (pushing users to Blackwell), multi-token generation is a future optimization since it currently emits one token at a time, and accuracy is the current focus before performance. Docusign and NVIDIA will extend from extraction into Nemotron Retriever (finding the right page at all) and the NVIDIA Agent Toolkit for production-scale agents." ], "takeaways": [ "Don't throw a generic multi-billion-parameter VLM at document tables — use a purpose-built extractor. Docusign's headline lesson is that a small, purpose-built model per job let them accelerate and bring things to market much faster.", "Optimize for throughput economics at enterprise scale: a sub-1B-parameter model with lower context means lower latency and lower cost, and tables-per-second is the metric that matters when you process a million documents a day.", "Split your pipeline by content type rather than picking one extractor — run layout-aware table extraction for tables and keep OCR for other fields, metadata and body text.", "Decide upfront whether your use case is high-throughput or low-latency: pre-process and index the corpus when queries span thousands of documents, but batch sizes, concurrency and technique all change for a single uploaded contract you want to Q&A immediately.", "Make the output actually usable downstream — structured export (CSV) and an API so procurement, finance and legal can push extracted pricing tables into systems like Coupa, rather than stopping at a chat answer.", "Solve retrieval and extraction as two separate problems: first find the right document for a query, then find the right information inside it." ], "topics": [ "document-extraction", "table-extraction", "vlm", "retrieval", "enterprise-ai", "small-models", "ocr", "inference-optimization" ], "tools": [ "Docusign", "Docusign IAM (Intelligent Agreement Management)", "Docusign Agreement Manager", "NVIDIA", "Nemotron", "Nemotron Parse", "Nemotron Retriever", "NVIDIA NIM", "vLLM", "C-RADIO", "YOLOX", "NVIDIA Agent Toolkit", "NVIDIA Blackwell", "NVFP4 / FP8 / FP16", "RD-TableBench", "ViDoRe (V1/V2/V3)", "MTEB / MMTEB", "Deloitte", "Coupa", "Claude" ], "quotes": [ { "text": "there's $2 trillion captured in this agreement negotiated value that no one capitalizes no one goes back and gets that um data back", "at": "02:19", "url": "https://www.youtube.com/watch?v=_gvamfT8H-w&t=139s" }, { "text": "simple thing what did we contract for the total tokens right with claude no one knows that's captured inside this agreement in different forms and fashion", "at": "03:45", "url": "https://www.youtube.com/watch?v=_gvamfT8H-w&t=225s" }, { "text": "traditional document extraction tools or a generic VM BLM completely fail here like you know we've tried we've definitely done this because they're reading text line by line which breaks a lot of that concept within the table", "at": "04:15", "url": "https://www.youtube.com/watch?v=_gvamfT8H-w&t=255s" }, { "text": "when we think about how many tables can it extract per second Neotron was 20x faster", "at": "12:16", "url": "https://www.youtube.com/watch?v=_gvamfT8H-w&t=736s" }, { "text": "It's not a generator. It's more of an extractor at the end of the day.", "at": "07:55", "url": "https://www.youtube.com/watch?v=_gvamfT8H-w&t=475s" } ], "words": 3671 }, "summary_url": "/#_gvamfT8H-w", "transcript": { "html": "/transcripts/_gvamfT8H-w.html", "txt": "/transcripts/_gvamfT8H-w.txt", "vtt": "/transcripts/_gvamfT8H-w.vtt" } }