ONNX runs the ML model that is f(string) -> vector. The similarity search is done using those vectors, and needs to return the original strings.