The $2 Trillion Data Problem

Organizations struggle to query their own agreements because critical data—pricing tiers, SKUs, and SLAs—is trapped in unstructured PDFs. Docusign estimates that roughly $2 trillion in negotiated value remains uncapitalized because retrieving this information requires manual human review across disconnected systems. At Docusign’s scale, processing a million agreements daily, traditional extraction methods fail because they treat documents as flat text, which destroys the semantic structure of complex, nested tables.

The Case for Purpose-Built Small Models

Rather than relying on massive, general-purpose vision language models (VLMs), Docusign and NVIDIA collaborated to build a purpose-built model within the Neotron portfolio. This model, at approximately 850-900 million parameters, is designed specifically as an extractor rather than a generator.

Key technical advantages include:

  • Efficiency: The model is 20x faster at table extraction than generic alternatives.
  • Lower Overhead: Reduced parameter count and context requirements lead to significantly lower latency and operational costs.
  • Structural Integrity: Unlike text-based extraction, this model preserves layout, reading order, and semantic table structure in a single pass, eliminating the need for separate layout and table-parsing models.

Architectural Strategy for Enterprise Scale

To handle the massive volume of agreements, the team employs a hybrid pipeline. While the purpose-built model handles complex table extraction, other metadata is processed via specialized OCR modules. This modular approach allows for high-throughput batch processing, which is essential for making historical data queryable.

Looking forward, the partnership is focusing on:

  • Optimization: Moving from FP16 to FP8 and FP4 quantization to further improve performance on NVIDIA hardware.
  • Agentic Workflows: Integrating the extraction pipeline with the NVIDIA Agent Toolkit to move from static data retrieval to dynamic, agent-driven insights.
  • Retrieval: Expanding the architecture to include the Neotron Retriever, enabling agents to first locate the correct document within a petabyte-scale corpus before extracting specific data points.