Embeddings

We offer multiple embedding methods to encode the text and layout information of the PDFs. The following components can be added to a pipeline or composed together, and contain preprocessing and postprocessing logic to convert and batch documents.

Factory name	Description
simple-text-embedding	A module that embeds the textual features of the blocks.
embedding-combiner	Encodes boxes using a combination of multiple encoders
sub-box-cnn-pooler	Pools the output of a CNN over the elements of a box (like words)
box-layout-embedding	Encodes the layout of the boxes
box-transformer	Contextualizes box representations using a transformer
huggingface-embedding	Box representations using a Huggingface multi-modal model.

Layers

These components are not to be confused with layers, which are standard PyTorch modules that can be used to build trainable components, such as the ones described here.

---