Some main points are you can expand a max context range by using summarization and a algorithm for handling memories.
I have a library of about 100 expressive emotes labeled based on a keyword description of the emote. This algorithm leverages synonyms, antonyms, and nouns to get the likeliest emote (synonym example synonyms={“Hello”,“Hi”,“Hey”,“Greetings”}, antonyms={“goodbye”,“farewell”}) Scoring each entry by only looking for one example of each synonym group. Thus we can query a database with the sentence. “Hey there, nice to meet you. I’m very excited to go on a adventure.” and the npc would wave hello, then as it says the next sentence it would find the emotes with tags of excited, and adventure. Excited is a noun and so is adventure so adding noise to the algorithm would give it a 50/50 chance to execute either a emote tagged with adventure of excited.
Some main advice is that Good AI requires good data!
I’d be willing to share with you the library of emotes I have not open-sourced to speed up your project. I understand if you wouldn’t want to use my chat module library to leverage all those elements of the english language (synonyms, antonyms, reflections and nouns) to query such a database.
But the code is in the open sourced module it just doesn’t have the library if you were interested to see how it was implemented.
This was done by displaying each sentence at a time on a word by word basis (or character by character if FPS>30) then processing the sentence to determine the emote (and now actions based off my 170 action commands).
Here is a demonstration video of what I’m talking about. I also used a similar system to create a library of atmospheric particle effects and audio samples based on emotional tonality.
Here’s a demonstration video.
Also I just did an experiment where I inject the response as the starting string to Zephyr 7b and got this output.
[["Kahlani: Good day to you, traveler! I am Duchess Kahlani; it's my honor to make your acquaintance. How may I assist you on this fine morning?
ArtStudios: I am in search of a rare artifact, rumored to be hidden in this very place. Do you happen to know anything about it?
Kahlani: I'm afraid I'm not privy to such information, traveler. However, I do know that there are a few chests scattered around this area, some of which may contain items of interest. Would you care to join me in exploring this island?
ArtStudios: That would be most gracious of you, Duchess. I would be honored to accompany you on this quest.
Kahlani: Very well, let us set off then. But first, let us take a moment to orient ourselves. Based on my instincts, I believe we are currently near a Broadleaf tree to the southwest, and a chest is nearby. There are also a couple of locked chests in the vicinity, but I'm afraid I don't have the key to them. Shall we begin our search?
ArtStudios: Absolutely, Duchess. Lead the way!
Kahlani: As you wish, traveler. Let us proceed with caution and vigilance, for we never know what dangers may lie ahead. But with your skills and my intuition, I'm confident we'll find what we're looking for! "]]
In this example I have Zephyr acting as a story teller and start it with the response from your AI after it displays the response. Then it simulates a conversation between the player and the npc.
Then the conversation is quantized into a memory for Zephyr. and Perhaps it should also become a memory for the other model. So they could be seamlessly integrated.
Also one modification I would make is something like this in the perception module!
if primaryPart then
local pos=Vector3.new(math.floor(primaryPart.Position.X),math.floor(primaryPart.Position.Y),math.floor(primaryPart.Position.Z))
local description = name .. " at " .. tostring(pos)
table.insert(descriptions, description)
end
Should round down the floating points of each position to reduce token usage. That also may make it easier for the AI to call functions with those cordinates.