I think this is the key. A system based on just counting the words in its database will give 0% for sentence continuations not in the database, but LLMs don't look up, they extrapolate, and can provide good sentence completions for sentences that are entirely new.
Two things: 1) This library will continue predicting words too for every word that exists in the model, it just makes less sense as you keep going. That's also true of any LLM. 2) An LLM is not synonymous with "chat bot".