Detecting table grapes’s soluble solids content(SSC) is a crucial issue since berry quality and flavor are directly related to it. In recent years, as the technology of chip-level spectral sensors is becoming more and more advanced, on-chip spectral sensors with high accuracy and stability have blazed a new trail for spectral detection. In this work, a small, user-friendly, and cost-effective optical device that can detect the SSC of table grapes nondestructively has been designed, built, and tested. New generation Vis/NIR spectral sensor AS7263(sensor 1, 2) with the capacity of chip level spectral analysis, does the key work for the system. Each sensor has six digital spectral channels with an integrated Gaussian filter and anLED with the programmable current (1~100 mA). The central wavelength of the spectral channel increases uniformly from 610 to 860 nm. Moreover, LEDs can emit light at 730 or 850 nm with fullwidth half max (FWHM) of 50nm. Firstly, this optical prototype collected a spectrum from 276 grape berries in a dark room. PAL-1 was used to detect SSC, and then we calculated the SSCt based on t-distribution: SSCt0.9 and SSCt0.95. Secondly, for the original spectral data, PCA was used to extract the principal components, and 16 abnormal samples located outside the confidence interval were excluded according to the distribution of the score factors. Besides, First Derivative (FD), Normalization (0, 1) and Standardization (0, 1) were used to preprocess the data. After that, we calculated the absorbance or KubelkaMunk function value F(R) for the samples at 12 channels. According to the multiple correlations between the independent variables of the Vis/NIR spectrum and the nonlinear correlations between the spectrum and SSCt, a PLS-BP neural network prediction model was developed for the grape SSC detection. The results showed that when β was 0.95, the preprocessing method was Standardization (0, 1), the Parameter of the spectrum was absorbance(A), and the prediction model worked best: R2p=0.93, RMSEP=0.181, Bias=-0.01, and RPD=3.78, which can be considered that the model has high accuracy and better adaptability to predict the SSC of table grapes. Finally, on the one hand, referring to the experimental results, a very interesting molecular scale principle analysis is obtained for the grape absorption spectrum: among numerous molecular vibration types, the 3x, 4x frequency of O—H bond (3x means stretching vibration at a triple fundamental frequency), the 3x+C, 4x+C frequency of O—H bond (3x+C means the combination of scissoring vibration at a fundamental frequency and stretching vibration at a triple fundamental frequency), the 8x, 9x of CO bond are the effective vibration frequencies for Vis/NIR spectral detection. On the other hand, the prototype provides a technical reference for future online quality inspection equipment that is high-precision, portable and low-cost.