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

Docusign and NVIDIA partnered to build a purpose-built small VLM (Nemotron Parse) for extracting structured data - especially complex tables - from massive volum...

By Sean Weldon

Your Agreements Are a Database You Can't Query: Purpose-Built VLMs for Enterprise Document Extraction

Abstract

Enterprise document understanding remains bottlenecked by the extraction of structured information from unstructured, hierarchically organized artifacts, particularly tables in which commercially decisive terms are typically encoded. This synthesis examines a collaboration between Docusign and NVIDIA that produced Nemotron Parse, a purpose-built small vision-language model (VLM) of approximately 850-900 million parameters optimized for single-pass extraction of text, layout, reading order, and preserved table structure. The analysis situates this work within Docusign's intelligent agreement management (IAM) platform, which processes roughly one million agreements per day, and within NVIDIA's open-source Nemotron Retriever initiative addressing two-stage retrieval at petabyte scale. Benchmarking indicates a 20x throughput advantage in tables extracted per second relative to comparable open-source models. Findings suggest that task-specific model specialization, hybrid pipelines combining VLMs with conventional optical character recognition (OCR), and workload-aware compute strategies deliver superior cost-latency-accuracy trade-offs compared with generic large-model approaches.

1. Introduction

Contractual agreements constitute a substantial but largely inert corpus of enterprise knowledge. A Deloitte study cited in this work estimates $2 trillion in negotiated value remains trapped in agreements that organizations never operationalize. The mechanism of this loss is straightforward: agreements arrive as unstructured artifacts - PDF and PNG files - and are hierarchically interdependent, with master agreements governing subordinate order forms, amendments, and statements of work. Consequently, legal, procurement, and sales functions expend substantial manual effort locating elementary facts such as renewal dates, liability caps, or pricing commitments.

Key terminology is established here for clarity. A vision-language model (VLM) is a neural architecture that jointly processes image and text modalities, in this context to interpret document layout and content simultaneously. Table structure preservation refers to the retention of row, column, merged-cell, and boundary relationships during extraction, as opposed to flattening a table into sequential text. The central research question addressed here concerns model sizing and specialization: whether extraction of structured agreement data is better served by general-purpose, multi-billion-parameter VLMs or by compact, purpose-built extractors.

The evidence presented favors the latter, particularly when throughput, unit economics, and table fidelity govern production viability. This synthesis proceeds by establishing the problem scale and platform context (Section 2), analyzing the architecture and design philosophy of Nemotron Parse alongside the broader retrieval stack (Section 3), distilling actionable engineering findings (Section 4), and discussing implications and open problems (Sections 5-6).

2. Background and Related Work

Docusign reports 1.9 million paying customers, approximately one billion users, and a processing volume near one million agreements per day. At this magnitude, per-document inference cost and latency become first-order architectural constraints rather than secondary optimizations. Underpinning Docusign's intelligent agreement management (IAM) platform is a proprietary agreement data model that structures extracted information at two levels of granularity - the individual agreement and the aggregate organizational portfolio - enabling both document-scoped question answering and cross-portfolio analytics.

This work also draws on NVIDIA's Nemotron family, an open-source initiative publishing datasets, compression techniques (quantization, distillation, pruning), and deployment blueprints. Within this family, Nemotron Retriever addresses a two-step retrieval problem common to enterprise search: first identifying the correct document within petabyte-scale corpora, then extracting the correct information within that document. NVIDIA's retrieval team has topped leaderboards including Vidori V1/V2/V3 and MTEB/MMTEB, establishing credibility for the embedding and reranking components that complement the extraction model examined here.

3. Core Analysis

3.1 Failure Modes of Prior Extraction Approaches

Traditional document extraction tooling and generic VLMs were observed to fail on tabular content because they process text line by line, a strategy that disregards two-dimensional table semantics. This line-by-line approach breaks merged cells, misaligns column boundaries, and destroys the row-column relationships that encode pricing tiers, SKUs, SLAs, and rate cards - precisely the terms of greatest commercial consequence in agreements. Because these terms are conventionally expressed in tabular form, extraction failures at the table level propagate directly into downstream agreement intelligence errors, motivating a design response centered on structure preservation rather than generic text recognition.

3.2 Architecture Design Principles of Nemotron Parse

Nemotron Parse is architected as an extractor rather than a generator, distinguishing it from typical generative VLMs. At approximately 850-900 million parameters, it is substantially smaller than multi-billion-parameter VLM alternatives, yet it is designed to output semantic formatting, layout, text, reading order, and preserved table structure in a single inference pass. This all-in-one design consolidates functionality previously distributed across multiple specialized models, such as YOLO X for page-element detection, into one unified pipeline - reducing orchestration complexity and cumulative latency.

The model can be served via NVIDIA NIM or accessed as a standard LLM endpoint, and it currently operates at FP16 precision, with a stated roadmap toward FP8 and NVFP4 quantization on Blackwell-generation hardware. The underlying architecture is encoder-decoder, generating output one token at a time; a planned optimization path introduces multi-token generation to reduce decoding latency, indicating that current throughput figures represent a floor rather than a ceiling for this design.

3.3 Applied Integration: The Agreement Manager Demonstration

The practical integration of Nemotron Parse was demonstrated through Docusign's Agreement Manager interface, in which uploading an order form triggers automatic AI processing and metadata extraction within seconds. Key commercial terms are structured, highlighted, and linked back to their source sections within the document, with Nemotron specifically powering the decomposition of dense pricing tables into discrete, structured order details. Extracted data is made available for download as CSV or through API access for finance and procurement workflows, illustrating an end-to-end path from unstructured input to queryable output.

4. Technical Insights

Several implementation-level findings emerge from this partnership that generalize beyond the specific agreement-intelligence use case.

5. Discussion

The findings synthesized here support a broader argument within applied AI engineering: that task-specific model specialization can outperform generic scaling for narrowly defined extraction problems, particularly where throughput and unit economics are binding constraints. The 20x throughput advantage attributed to Nemotron Parse did not derive from architectural novelty in isolation, but from the explicit reframing of the model as an extractor rather than a generator - an assignment of the model's inductive biases toward the structural regularities of documents rather than open-ended text generation.

This has implications for the ongoing industry debate over consolidation versus specialization of foundation models. While large multi-billion-parameter VLMs offer generality across tasks, the evidence here suggests that production systems operating at enterprise scale may benefit from decomposing a workflow into specialized components - an extraction model, a hybrid OCR layer, retrieval and reranking models - rather than relying on a single monolithic model. A remaining knowledge gap concerns the transferability of Nemotron Parse's table-extraction gains to document types with different structural conventions than commercial agreements, such as scientific literature or financial filings, which was not directly addressed in the source material.

The roadmap toward Nemotron Retriever integration - extending from page-level extraction to document- and corpus-level retrieval - suggests that the partnership views extraction as one stage within a larger agentic pipeline, foreshadowing integration with the NVIDIA agent toolkit for production-scale agentic use cases.

6. Conclusion

This synthesis has examined how Docusign and NVIDIA addressed the structural failure of generic document extraction tools on tabular agreement data through the development of Nemotron Parse, a compact, purpose-built VLM achieving a reported 20x throughput improvement over comparable open-source alternatives. The core contribution lies not merely in the model's parameter efficiency but in its architectural reframing as a single-pass extractor and its integration into a hybrid pipeline alongside traditional OCR.

For practitioners, the practical takeaway is that model selection for enterprise document workflows should be driven by task specificity and production-scale economics rather than default preference for larger, general-purpose models. Future work, as indicated by the stated roadmap, will extend this extraction capability into full retrieval and agentic pipelines, suggesting that document intelligence architectures are converging toward modular systems combining specialized extraction, retrieval, and reasoning components.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub