Overview

This section outlines the end-to-end process of building generative extraction models for unstructured and complex documents.

Building high-performing models is an iterative process that requires continuous prompt engineering and output validation.

The model building process

1. Upload a sample of representative documents for the use case:
   • Identify what data points (fields) you want to extract and how they relate to each other (field groups).
   • Provide overall extraction instructions.
   • Create the corresponding extraction schema and provide field group and field instructions.

2. Validate and improve extractions:
   • Review the initial predictions, and notice how well they are performing.
   • Modify the prompt instructions and test the impact on the performance of the model.
   • Confirm or correct extractions to gather ground truth for performance evaluation.