One of my past projects is based neural network in 2020. The solution is quite simple: put the feature vectors into postgresql database create distanced based index I adopted the model and algrithm from face_recognition, and my team trained on sample faces from our internal system to improve the accuracy to reduce the human involvement and save costs. During development, we found that there is a limit of postgres: CUBE_MAX_DIM, my team customize the build of postgreql to work around the limit.