AI EngineerVideoSource Linked

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

Docusign and NVIDIA use a small extraction model to preserve complex agreement tables as structured data, showing where specialized parsers can beat generic VLMs in agent pipelines.

AI Engineer · Sep 16, 2026

Open Source Open Markdown Open JSON

Source Summary

Docusign says it processes **about 1 million agreements daily** across 1.9 million paying customers. Pricing, SLA, and rate-card tables often defeat line-oriented extraction, especially with merged or nested cells.

Practical Implication

Use specialized extraction before retrieval when an agent must reason over table structure. The NeMo Retriever parse model preserves layout, reading order, and table relationships; Docusign reports **20× faster table extraction** than the open models it tested.

Agent-Ready Context

Docusign says it processes **about 1 million agreements daily** across 1.9 million paying customers. Pricing, SLA, and rate-card tables often defeat line-oriented extraction, especially with merged or nested cells.

Use specialized extraction before retrieval when an agent must reason over table structure. The NeMo Retriever parse model preserves layout, reading order, and table relationships; Docusign reports **20× faster table extraction** than the open models it tested.

The talk does not provide the underlying scores, model list, or production error rates. Docusign still uses a hybrid pipeline with OCR for other fields, while quantization and multi-token generation were described as future optimization paths.

Connected Context · Feed7 Judgment

This moves document parsing into the model-selection path: when meaning depends on layout and merged or nested cells, retrieval quality is capped by structure lost during extraction. It supports a specialized, hybrid pipeline rather than one universal parser, while the missing model list, scores, and error rates make the reported speedup insufficient to establish production accuracy or general superiority.

When Does Bigger Help? A Controlled Study of LLM Scale for Ontology LearningThe ontology study reinforces the same selection rule: model scale alone is a weak proxy, so the parser should be evaluated on the exact structural task rather than chosen by size.Compression at the Edge — NVIDIA, Unsloth, HuggingFace, OllamaDocusign’s proposed quantization path inherits the compression warning that each resulting artifact must be tested in its real extraction harness because speed, precision, and quality interact.The Structure of Quantization Damage in LLMs: Why the Next Bit Should Be Spent GloballyIf the parser is quantized, the causal finding favors testing broadly distributed extra precision before assuming that protecting a few layers will preserve extraction quality.Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist DomainsThe Greek RAG results independently reinforce staged, workload-specific evaluation: upstream adaptation—retrieval there, structured parsing here—can determine the quality of grounded answers downstream.

Context Map

modeldata#model-selection#retrieval#open-models

Uncertainty

The talk does not provide the underlying scores, model list, or production error rates. Docusign still uses a hybrid pipeline with OCR for other fields, while quantization and multi-token generation were described as future optimization paths.