As a kind of RAG, how should you add conversational knowledge?
OpenAI said below, but is it still effective? I tried and didn’t find so.
To help clarify that the example messages are not part of a real conversation, and shouldn’t be referred back to by the model, you can try setting the name field of system messages to example_user and example_assistant . ( How to format inputs to ChatGPT models from OpenAI Cookbook)
Edit: I need clarify.
I’m just finding (or feeling) example_user and example_assistant methodology isn’t working well compared to just putting like conversation history. messages are fed as <|start|>{role/name}\n{content}<|end|>\n. They use name if available, instead of role. When you use example_assistant and example_user, does GPT recognize it is something that an assistant CAN refer back to? Is it trained like that?
What the quote describes is what is called multi-shot training.
Traditional completions AI quickly learns a pattern, and thus a previous pattern that shows it “chatting”, or processing your data in the manner intended, can be beneficial.
Chat models have enough training that they don’t really take to or learn from those examples. An AI that only barks like a dog in your examples is going to resume chatbot behaviors despite showing it past conversation.
You don’t refer to them. Past output patterns train future output patterns.
(This is not retrieval-augmented generation, which is injection of relevant information for satisfying the latest query by search)
Using a “name” and then departing from that name in real chat gives a bit more assurance that the user doesn’t ask “what did I just say” and get the examples confusingly played back for them.
The technique is more useful for demonstrating and training structured output or multi-part responses than for any kind of answering about previous sample inputs as “knowledge”.