For prompting engineering, we add one or two additional prompt completion pairs before our real question. My question is: Are those prompt completion pairs fixed or will be randomly sampled from the dataset every time we ask a new question?
Those additional prompt helpers are supposed to provide a context to the model so it can provide a better answer within that context.