Comment by yoavz

The extension feeds it context data in chunks (e.g. only one <p> element at a time). I could see TensorFlowJS BERT struggling with a single sequence of 87 KB -- have you tried breaking them up?