Nutrient Vision API: Document extraction beyond OCR

Traditional OCR reads text off a page — but it loses tables, ignores reading order, and can’t handle handwriting or equations. Vision API goes further. It analyzes document layout, detects semantic elements, and returns structured, machine-readable data with coordinates mapping every value back to its exact location in the source.

In this presentation, we walk through how Vision API works, what each extraction mode is built for, and where it fits in your document processing stack.

• Intelligent content recognition (ICR) — Local AI models analyze document layout and extract structured content: tables with cell-level coordinates, handwriting, equations in LaTeX format, and key-value pairs from forms — all processed on your servers with zero network requests.
• VLM-enhanced ICR — For the most difficult documents. Combines local ICR with a cloud vision language model (Claude, OpenAI, or AWS Bedrock) for improved accuracy on degraded scans, complex financial tables, and handwritten forms. You choose which documents route to the cloud and which provider to use.
• AI-generated image descriptions — Generate natural language descriptions of photographs, diagrams, and charts for WCAG and Section 508 accessibility compliance. Use a cloud provider or run a local VLM server for fully on-premises processing.
• Deployment models — How to run Vision API fully locally for air-gapped and regulated environments, or selectively connect cloud providers for enhanced accuracy — and how to switch between modes with a single configuration change.

Vision API is available now in the Java and Python SDKs, with Document Engine and DWS Processor API support coming later this quarter.

Watch on demand

Learn how Vision API handles real documents containing tables, handwriting, and formulas — and how the same SDK gives you full control over where processing happens and what accuracy tradeoffs to make.