Generating relevant and factual answers
GPT-3 is a language model – it predicts the statistical likelihood of the text that should follow the prompt text it was provided. It's not a knowledge base in the sense that it's concerned much with the factual accuracy of the responses it generates. That doesn't mean it won't generate factual answers; it just means you can't count on the answers being accurate all of the time. But the Answers endpoint can provide a lot of control over the accuracy or relevancy of the answers that will get generated.
As we discussed earlier in Introducing the Answers endpoint, answers will be generated from the documents we provide. At this point, we're providing documents as part of the endpoint request. Using that method, if the answer can't be derived from the documents, the engine defined by the model parameter will be used to generate an answer. You can find that set in the routes/answer.js file – we used...