I’d create an embedding center by averaging a dozen or so apology responses. If the output has an embedding too close to that cluster you can handle the excepti

I’d create an embedding center by averaging a dozen or so apology responses. If the output has an embedding too close to that cluster you can handle the exception appropriately.