I'm trying to figure out the new chat completion API...
What's the 'system' role for? Can you use it during the conversation between the user and the AI and have your application be the "system" that listens in on the conversation and chimes in? Or should those guidances use the 'user' role?
#what's the system role for?
buoyant wigeon
lyric cargo
Check this out https://platform.openai.com/docs/guides/chat/introduction
As quoted from the documentation
The system message helps set the behavior of the assistant. In the example above, the assistant was instructed with “You are a helpful assistant.”
The user messages help instruct the assistant. They can be generated by the end users of an application, or set by a developer as an instruction.
The assistant messages help store prior responses. They can also be written by a developer to help give examples of desired behavior.
buoyant wigeon
I've read that.
This suggests that man in the middle "steering" should still use the 'user' role...
So 'system' is like setting the style
shrewd rune
Hi, is there a way to configure the employee playground so that the employees will not have access to it?. In that case the system would work from the Administrator playground
stoic geyser
I also am still trying to figure it out but as they say context I'm guessing this is where you might tell it "you are not a nlpai, you are a doctor and respond as a doctor." Or something like this. I will probably play a little bit with this in the next day or two if we don't see a better answer here I might be able to contribute by then.
buoyant wigeon
It would make sense to me if we had 3 actors:
- your application that uses the API
system - the end user
user - the AI model
assistant
...but based on the documentation that's not how it works...
crystal pike
I'm also wondering this. I wonder if you add an additional "system" message, does that replace the previous one, or are they all considered?
stoic geyser
here on openai docs it says this "If adding instructions in the system message doesn't work, you can also try putting them into a user message. (In the near future, we will train our models to be much more steerable via the system message. But to date, we have trained only on a few system messages, so the models pay much more attention to user example"
They show very simple direction for the system here > https://github.com/openai/openai-python/blob/main/chatml.md
buoyant wigeon
I'm not using the Python API, I'm using the "higher level API"
So far system seems to work for me in early testing.
buoyant wigeon
My tests yesterday suggest that user role messages seem to have more effect than followup system messages for driving behavior of the assistant. Is anyone else finding this? It would match what the documentation says and also likely change i.e. improve in later releases...
opaque harness
That's correct, user role messages have more weight than system role messages, their documentation mentions that system messages are deweighted
@buoyant wigeon
buoyant wigeon
I'm going to try adding an internal hint role and for the time being just replace "hint" with "user" when calling the API
crystal pike
I wonder if there's actually any reason to use "system" instead of "user" like, is it considered in a different way than user messages.
glacial turtle
The different roles have different weights, they make system more significant than user in the future
buoyant wigeon
I wonder if there's actually any reason to use "system" instead of "user" like, ...
it would make sense to me... it wasn't the "user" who said it... but the "system"
crystal pike
The different roles have different weights, they make `system` more significant ...
I mean now. There seems to be no information other than “it has less weight” which seems to make it useless unless it’s somehow considered differently