Model type

In Typless, we offer two types of models: Pretrained and Trainable.

Pretrained models

Pretrained models work out of the box — no dataset building or training required before they start extracting data. They are based on Large Language Models, and therefore they understand the semantic meaning of the document they are processing.

These models, however, need special instructions for each field that is being processed. We use Field description to textually explain what we want to extract in the document type field. The model will then look for this field on the document.

However, this does not mean that providing feedback for the model does not improve its performance. The model is fine-tuned for each individual supplier based on the feedback provided. Additionally, Supplier description can be utilized to further steer the model into correct data extraction with instructions for that particular Supplier.

Trainable models

If you are not satisfied with the results of the Pretrained models, you can always build your own dataset and train the Trainable models. These models need the confirmed data of at least one example of each Supplier in order to start correctly extracting the data. They rely more heavily on the feedback provided, and therefore follow the instructions in the feedback better. Trainable models might give more predictable/consistent results for stable document layouts, while Pretrained models handles variety better.