A setting that controls how random or creative a model's output is: low values are more consistent, high values more varied.
Temperature is a number you set when calling a language model. At low values the model picks the most likely words, so answers are similar and predictable. At higher values it explores less likely options, making replies more varied and creative, and sometimes less accurate.
The right value depends on the task. Extracting data from an invoice or classifying customer requests needs consistency and precision, while brainstorming social media ideas benefits from variety. Using one setting for everything in your app usually gives mediocre results somewhere. Practical tip: start low for tasks with one correct answer and raise it gradually for creative tasks. Remember that a low temperature does not eliminate hallucinations; it makes answers consistent, not necessarily correct.