It's possible to detect, at least, binary features from dataset values as they would have only 2 possible values. The same can happen for one-hot encoded features, as a set of features (generally sequentially) have specific rules like: just one active value per multiple columns.
However, we would need the dataset to detect that. For a single input instance, these assumptions are more difficult to make as we do not have sufficient information.
Therefore, this improvement could initially: takes a dataset and automatically identify feature types.