Top P is a model setting that determines how many different phrases a model considers when it’s trying to predict the next string in an output. Much like the Temperature setting, it determines diversity, with a higher Top P being more diverse. For example, setting the Top P value to 0.5 instructs the model to evaluate the top 50 most likely words of phrases. It ranges from 0 to 1, with 0.7 as a default.
For agents meant to deliver accurate information, a lower Top P provides more consistent and predictable results. For agents meant to brainstorm ideas and emulate creativity, a higher Top P provides more variation and new ideas.