The guide
Text has to become numbers, and every choice in that conversion has consequences you feel much later.
Character tokenisers never hit an unknown word but make sequences long and force the model to learn spelling. Word tokenisers are compact and fall over on anything unseen. Byte pair encoding splits the difference by learning its vocabulary from the corpus, merging frequent pairs until it hits a budget, which is why modern models tokenise "unbelievable" into pieces rather than one symbol or eleven.
TF-IDF is the pre-neural way of deciding which words carry signal, and it still works surprisingly well. Positional encoding exists because attention has no inherent sense of order, so position has to be added deliberately.