Pre-Trained Models

What are pre-trained Document AI models?

Pre-trained Document AI models are AI models developed and trained on large volumes of real-world documents before deployment, enabling organizations to process specific document types accurately from day one, without providing their own training data or running model training projects.

In intelligent document processing (IDP), pre-trained models cover the data extraction, classification, and validation tasks for common document types: invoices, purchase orders, contracts, identity documents, insurance forms, financial statements, and hundreds of others.

The value of pre-trained models: Time-to-value and accuracy

The value of pre-trained models is time-to-value and accuracy: rather than spending weeks or months collecting training documents and building custom models, organizations can deploy pre-trained extraction models immediately and achieve production-grade accuracy for covered document types out of the box.

Pre-trained models can typically be extended or fine-tuned on organization-specific document variations: unusual layouts, custom fields, domain-specific terminology, to improve accuracy for edge cases without requiring a full custom training project.

Pre-trained models vs Generative AI

Pre-trained Document AI models differ from zero-shot extraction with general-purpose generative AI in several critical respects.

• Pre-trained Document AI models: These models are deterministic, purpose-built, and optimized for a specific document task. A pre-trained intelligent document processing model is engineered to extract a field consistently and accurately. It uses confidence scoring that signals when human review is warranted.
• Generative AI models: These models apply broad language understanding to infer what a field might contain. This can produce outputs that vary between runs and hallucinate values that do not present in the document.

Combining pre-trained models with generative AI

The most effective Document AI architectures combine both: pre-trained models for deterministic accuracy on known document types, augmented by generative AI for contextual interpretation, data enrichment.