AI Assist

AI Assist is a powerful feature that can be utilized to accelerate the annotation process for custom document types where no pre-trained model is available. By leveraging AI Assist, you can annotate documents more efficiently and train machine learning (ML) models specific to their needs. This support document provides instructions on how to make use of AI Assist for custom document types and train ML models for better extraction results.

Using AI Assist for Custom Document Types:

To utilize AI Assist for custom document types, follow these steps:

1. Open document type settings and head to “Extraction”.

2. Look for “AI Assist” and enable it by clicking the toggle.

Annotate documents using AI Assist: Begin annotating your custom document type using AI Assist. As you annotate the documents, AI Assist will provide intelligent suggestions to speed up the annotation process from the first document without any initial training. All you have to do is verify if all the values extracted/suggested by AI assist are correct.

Training an ML Model for Custom Document Types:

After annotating a sufficient number of documents (at least 20 or more), you can train an ML model specific to your custom document type to generate better extraction results. Follow these steps to train an ML model.

1. Open “Models and Training”, and click “New Model” to start with the training.

2. Select the document type for which you wish to train the model.

3. Select the model type as Key-Value Model.

4. Use the annotated documents to train the model specific to your requirements.

5. Once the model is trained, link it to the document type and you are good to go. No more manual work around key value extraction.

Analyze the extraction results for the new uploads and identify any areas that require improvement. Make necessary adjustments to the model, such as refining the annotations or retraining with additional documents, to enhance its accuracy.

Note: You can also make use of AI Assist for pre-trained document types to generate results where the model is not falling short.