Perplexity
Short answer
Perplexity measures how surprised a model is by a body of text. Lower is better, and it is mainly used to check that quantisation has not damaged a model.
It is a useful internal gauge and a poor product claim. Perplexity on a standard corpus tells you very little about whether a model will answer your question about your PDF correctly.
Nothing to do with the company of the same name, which makes searching for either of them tiresome.
Related terms
Other ways people ask this
- what is perplexity metric