Make with H2O.ai : Getting Started with H2O Document AI
- Predictive AI
- H2O Document AI
Overview
This session introduces H2O Document AI, a tool for training models that extract structured data from documents like faxes, PDFs, and scanned images. Using a live demo built around a UCSF medical referral use case, the presenter shows how the tool treats documents as datasets, runs OCR to capture text and its position on the page, and supports both entity extraction and page classification models. It also covers labeling, merging annotations with OCR results to train a model, and publishing a scoring pipeline as a REST API.
Key takeaways
- How H2O Document AI handles documents that arrive in many different formats and layouts, rather than requiring a fixed template
- How the tool's OCR step (Tesseract, PDF text extraction, or the H2O-forked doc TR model) chooses the right method per page, including for scanned faxes with no embedded text
- How to label documents in the annotation interface, using labels for token-level entity extraction and classes for page classification
- How merging an annotation set with OCR results joins labels to tokens to produce the training data a model needs
- How to read model accuracy metrics, including macro versus weighted average, to spot which entity classes need more labeled examples
- How to publish a trained model as a pipeline with a custom post-processor and consume it through a REST API that returns predictions as JSON
Start owning your AI