Starch and proteins are the vital nutrients in buckwheat, to achieve the fast detection of the buckwheat internal composition has an important theoretical significance and application value for buckwheat breeding, processing and other steps. In the paper, forty buckwheat samples from different origins have been selected. The starch and protein content of buckwheat was determined, and the mid-infrared transmission spectrum of buckwheat has been obtained using Fourier mid-infrared spectroscopy. Forty samples were randomly divided into the prediction set and validation set with 30 and 10 samples respectively. After smoothing preprocessing, the prediction models of buckwheat starch and protein content have been established using the combination method of principal component analysis and artificial neural network, finally the models have been verified. The results showed that the correlation coefficient between the prediction value and measurement value of buckwheat starch content is 0.9029, and the relative error is smaller and its mean value is 2.33%, the method of the buckwheat starch content prediction is feasible. But the prediction for buckwheat protein content is not ideal, need to be further studied.