smspolt.blogg.se - Chatgpt 3 api

Instructions (about how it should answer, format, etc).

So how can we use this to do question answering? Let's think back to the information we need to pass in. Presumably, the model is trained to treat the user messages as human messages, system messages as some system level configuration, and assistant messages as previous chat responses from the assistant. The content field can be anything, while the role field should be one of user, system, or assistant. These messages are essentially dicts that have two fields: content and role. What do I mean by that? The ChatGPT API differs from the GPT-3 API in that it takes in a list of messages (rather than a single string) and returns a message. Another would be to try to use the ChatGPT API more natively.

One would be to use a wrapper that and treat ChatGPT has just another LLM. There would be two potential ways to do this. What models/prompts are we comparing?įirst up, we've got the standard text-davinci-003 model, with the standard VectorDBQAChain prompts. We will hold the "retrieval" step constant, so we are just evaluating the "generation" step of the chain. This chain takes a query, does a "retrieval" step to look up relevant documents in a vector store, and then does a "generation" step to pass them, along with the original query, to a model to get back an answer. In this article we will evaluate the performance of a chain on question answering over a particular dataset. Because I'm lazy, I also enrolled the help of the ChatGPT API itself to help do this evaluation.

But how does it actually compare to the existing API? It will take some time before there is a definitive answer, but here are some initial thoughts. OpenAI released a new ChatGPT API yesterday. 5 min read Photo by Tingey Injury Law Firm / Unsplash.