I have a software doing the same (use embeddings for semantic search and using completion davinci-003 for build a NL response once found the most relevant QA on db). I don’t use python, but PHP and i simply did a few curl lines to interact with API. So, it’s quite easy understand how does it work and what to change now.
In fact, i replaced the old endpoint by this one https://api.openai.com/v1/chat/completions and the model by this other gpt-3.5-turbo and also replaced the element prompt of the data send by this new messages array containing 2 arrays, one for role system with context and one for role user with user question/search string.
The response of this new endpoint chat also is different: only the element choices now contain a different array structure. But the difference is very minor.
Spoiler: there are a big DIFFERENCE in the responses. I’m still evaluating if it’s convenient for me. This new endpoint really SEEMS (talk) like the usual chatGPT. This means: their responses are easily 100% longer, but more comprehensive, i mean it give more details to the user… so, maybe is not so desirable for some uses… I’m talking even maintaining exactly equal the parameters (max_tokens, temperature, …).