Giving Multiple "System" Role Messages in Chat

I have created some code to answer questions better using the API that uses multiple AI instances in some programmatic components. I want to be able to feed all of the information into the final AI instances to take past findings into consideration. A lot of the data doesn’t follow the standard User<–>Assitant back-and-forth dynamic and causes issues when presented this way. Is there a way to give multiple System messages to give the AI instance background information? Does having multiple instances of System messages overwrite the previous one?

1 Like

You want to figure it out together like in a tomorrow evening meet session?
Do you store the communication in graphs?

1 Like

One approach we are about to test out is storing user queries and bot answers as linked data. The query and the answer get their own embedding and they are stored in redis. When a new query is made, the bot can get the k nearest data points via the embedding and inject it as context into the system card.

Although @jochenschultz’s communication graph is a new concept I have not heard of before. So that sounds kind of promising too.

I am not 100% sure but shouldn’t a graph db use the same approach in the background? I would say nlogn instead of n² is the only way to go.
Any math nerds here?

I want to let the communicating instances “live” in a computer science lab and let them work together on a better approach (let them have access to docker containers and code) and then move their memory onto the better world and repeat.

Let’s make it happen. Don’t be afraid.

1 Like

To many mad scientists in this world. We definitely all going to die.

But yeah, it is possible that the graph dbs handle it in the same way. I know Redis has a graph based system like that but I’m not really sure how they do it in the background. I tend to like creating my own methods and tweaking the experiment. As long as you have proper interfaces you can swap in and out methods pretty quick. Better than swapping out whole databases.

1 Like

I have not heard of the graph concept, they are just text strings in a list. It’s just that their format of them is not exactly a conversation. If I do put it in as a conversation rather than answering the prompt it will sometimes go back and make a comment about a previous conversation message. The intent is to give the information as context to aid in answering the question.

Also, when and where is the meeting session? Is this a discord thing?

Nah, I got to get some sleep now. Already awake for close to 48 hours. After that I will play around with it. I guess the experiment ends at the 120$ limit. So I’d go with gpt-3.5