Document AI

This post introduces H2OVL Mississippi, a new multimodal foundation model family for OCR and Document AI. It details the model's architecture, inspired by LLaVA and InternVL, featuring a ViT-MLP-LLM setup with dynamic resolution and multi-scale adaptive cropping. The training methodology involves a two-stage process of pretraining and fine-tuning specifically for document understanding. The post also outlines the H2O Document AI platform's capabilities, including OCR, NLP, and deep learning for data extraction, automated labeling, and integration via REST APIs. It describes the five-step workflow: Ingest, Label, Train, Deploy, and Consume, highlighting features like page orientation resolution, deskewing, text formatting optimization, and color binarization during ingestion. The integration with H2O MLOps for model management and deployment options are also mentioned.