The second port you did use for prediction contains only statistics about the data set. These are the ones used for prediction. However the full information necessary for prediction is saved as well in the PMML object (the blue data port)
I would recommend to use a Partitioning node to split into training and test data. Train the model on the training data only and afterwards apply the predictor on the test set.
You might want to go to our example server(https://www.knime.org/example-workflows), where example 002007_NaiveBayes is actually the same workflow (with different data)