How Adaboost works?

I have a basic doubt about Adaboost that pseudoalgorithms doesn’t clearly explain. At each iteration, it uses a subset (that is resampled based in the errors of the previous weak learner). After fitted, the current weak learner should predict samples to calculate the error and alpha. So the question is:

Which dataset is used for prediction at each iteration: the current resampled subset OR the original entire dataset?

Ps: when using the entire original dataset, Adaboost is performing poorly in my implementation. However i can’t confirm if using the resampled subset for prediction can bias the model.

Thanks in advance.