I have faced some serious problem.
Now I’m developing chatbot with custom knowledge base in Pinecone.
Embedding and Retrieving works very well.
But when generating answer, in some cases text-davincci-003 generates better answer and in other cases, gpt-3.5-turbo is better.
Here, I have used completion and chatcompletion.
I want to use only one model, gpt-3.5-turbo.
Because text-davincii-003 will be deprecated.
With only one model, there is no way to solve this?
I hope your golden help.
Hello to you!
text-davinci-003 is a tool that is powerful in the right hands, but it takes the correct type of input to harness the capabilities. It is where “prompt engineering” could actually not be a laughable skill.
A davinci chatbot is more prone to safety and jailbreak concerns. It will generate whatever text looks best following the user input given the particular context. The chat-specific model gpt-xxx is better in that regard, but its public-facing disclaimers and denials and writing attitude can be an annoyance for private use.
If you do find particular cases for davinci (at over 10x the cost), perhaps the question to ask then is “to which AI do I submit question and answer from both engines, to then have the AI decide which is the best output?”
How is your prompt to get the answer looks like? I mean template?
Here’s one I wrote just for you (using an OpenAI API capable-account). It shows how we can quickly train a completion engine to “complete” a conversation.
text-davinci-003 is already trained to follow instructions, so it doesn’t actually take this much work, but you can use this on smaller engines like curie for more amusing unpredicable chat.
maybe this article will help you