When I'm chatting with the AI, I've worked to train the AI to get the kinds of responses I want. For example, through editing previous AI responses into a format I want to see. When TV creates a summary, it then hides those messages and relies on the summary to fill in the details going forward. This sounds good on paper, but it's causing a couple of problems for me:
-
As an example, after TV hides the most recent messages, the AI completely forgets the formatting it was using in its responses. This might be fine for some people who prefer variety, but for me it's very distracting.
-
Another example, even if a summary has just been created for the most recent messages, relevant details can easily be lost to the AI. I can see those details in the hidden messages, but obviously the AI cannot. So I have the AI asking things like "So what did you do that night?", when in the message or two previously I've clearly stated it is daytime. Later in the chat, these details might not matter, but if TV has suddenly hidden every previous message, it's very jarring to have completely incorrect responses.
It's immersion breaking.
So, I'd like to propose a feature that would help with this.
I'd like to see TV create summaries beyond the number defined in the settings, then hide them, leaving the messages up to that number unhidden.
So if you have the setting as 'Create summaries every 10 messages', then the following would happen:
Do create summaries for these messages, then do hide them:
20
19
18
17
16
15
14
13
12
11
Don't create summaries for these messages and don't hide them:
Doing it this way would give the AI the most recent 10 full and complete messages to look at when forming its responses, plus it can pull little details that matter without breaking immersion.
Then TV can create summaries for the ten (or whatever the user has set) messages that are older than that, when the little details that mattered more immediately before might now have become less relevant as the chat goes on.
Hopefully my explanation makes sense. If so, hopefully it's something that can be implemented on a technical level.
When I'm chatting with the AI, I've worked to train the AI to get the kinds of responses I want. For example, through editing previous AI responses into a format I want to see. When TV creates a summary, it then hides those messages and relies on the summary to fill in the details going forward. This sounds good on paper, but it's causing a couple of problems for me:
As an example, after TV hides the most recent messages, the AI completely forgets the formatting it was using in its responses. This might be fine for some people who prefer variety, but for me it's very distracting.
Another example, even if a summary has just been created for the most recent messages, relevant details can easily be lost to the AI. I can see those details in the hidden messages, but obviously the AI cannot. So I have the AI asking things like "So what did you do that night?", when in the message or two previously I've clearly stated it is daytime. Later in the chat, these details might not matter, but if TV has suddenly hidden every previous message, it's very jarring to have completely incorrect responses.
It's immersion breaking.
So, I'd like to propose a feature that would help with this.
I'd like to see TV create summaries beyond the number defined in the settings, then hide them, leaving the messages up to that number unhidden.
So if you have the setting as 'Create summaries every 10 messages', then the following would happen:
Do create summaries for these messages, then do hide them:
Don't create summaries for these messages and don't hide them:
Doing it this way would give the AI the most recent 10 full and complete messages to look at when forming its responses, plus it can pull little details that matter without breaking immersion.
Then TV can create summaries for the ten (or whatever the user has set) messages that are older than that, when the little details that mattered more immediately before might now have become less relevant as the chat goes on.
Hopefully my explanation makes sense. If so, hopefully it's something that can be implemented on a technical level.