Cannot Create or Edit Group Chat Prompts

[ZeriLeandra]

Describe the bug
It took me an embarrassingly long time to figure this out. I modify my prompts to enforce third person past tense which SOTA LLMs almost always respect right out of the gate. In singular chats it works fine but I was trying to figure out why group chats were all over the place with tense and POV at the start, despite the initial scenario text being 3rd person past, and having to edit every character's entry until the LLM finally got the pattern. In going to the logs, I was seeing the beginning of a prompt that wasn't the one I had setup as system default.

I could not find any prompt that matched that among the stock ones in the Prompt Editor.

I work in much older systems than Rust or Typescript (think like COBOL) but poking around the files, I found the group chat prompt under

src-tauri\src\chat_manager\prompt_engine.rs

I'm not going to pretend to know Rust, but from references like PromptType::GroupChatPrompt and PromptType::GroupChatRoleplayPrompt in different files, it's clear that group chats get a different system prompt than singular chats and I can't find a way to create or edit group chat prompts. Setting an override prompt on a character isn't respected within a group chat (makes sense).

I actually think the PromptType is a fantastic design. Group chats need extra handling that individual chats don't (they want to devolve into pandemonium). Not to mention the conversational versus roleplay variants on both group chats and reply helper have different needs. I think the Prompt Editor just needs to include the other types of prompts not displayed and that, maybe from a dropdown, the user can pick what PromptType a prompt they've created is supposed to be. That's probably more complicated to add than I'm thinking, but it's the kind of thing power users would expect to be able to do. Tweaking group chat prompts to get a particular model to behave better is par for the course with group chat shenanigans that some models handle better than others.

Version

• Pre-Release v0.1.139 (windows-cpu)

[MegalithOfficial]

The reason I didn't add the prompt is because it's fragile. It relies on a strict format that can easily break down if someone edits it. For example, to make sure AIs don't mistake each other, the prompt tells models to use this format:
[Character Name]: <text>

Due to that, I have to ignore the Character System prompt.

I had an idea to create a new type of prompt entry called GroupChatEntry. Basically, this type of entries activate when a character is used in a group chat, but the rest of the prompt will be ignored. So this will act like extra info to pass into LLM.

[ZeriLeandra]

The reason I didn't add the prompt is because it's fragile. It relies on a strict format that can easily break down if someone edits it. For example, to make sure AIs don't mistake each other, the prompt tells models to use this format: [Character Name]: <text>

Yep... Group chats love to devolve into chaos. For what it's worth, I know both Wyvern and Chub automatically pre-pend the character's name to the start of each of their messages in the chat history. So no matter what you set as the prompt instructions for pre-history or post-history, wherever the chat history is, each message is name prefixed basically like this. I snipped this from some debugger info between "Bob the Skull", "Lifebot" and my user persona as Regina.

    {
      "content": "Bob the Skull: \"Oooh! New chat, eh? Have we... met before?\" I blink. \"Regina. Sounds somewhat familiar. Oh well, doesn't matter anyway. What's up, Boss?\"",
      "role": "assistant"
    },
	 {
      "content": "Regina: \"Hey Bob, I wanted to introduce you to my buddy, Lifebot! She's got almost as much good information as you do. Almost,\" I say with a wink.",
      "role": "user"
    },
	{
      "content": "Lifebot: The small floating robot head swam through the air gracefully, its digital face sporting a gentle smile.\n\n\"Oh,\" Lifebot said, its chipper voice bright with interest. \"Hello, Bob! Regina has spoken highly of your... unique expertise.\"\n\nLifebot floated a little closer, its digital eyes scanning the skull with polite, simulated curiosity.\n\n\"I'm Lifebot unit #52846, issued by the Social Well-being Bureau to assist Regina with daily tasks and overall life improvement. It's a pleasure to be introduced.\"",
      "role": "assistant"
    }, 

SillyTavern makes it optional but gives a warning about it.

If you select Default or MessageContent, it does the same thing as the above where the first thing after "content": is the character's name and a colon. Only diff is whether or not it does it for the persona too (not in Default, yes in Message Content). But even without persona and colon, since those messages get the "user" role rather than "assistant" role, the LLMs for the mostpart seem to understand not to speak for them.

I had an idea to create a new type of prompt entry called GroupChatEntry. Basically, this type of entries activate when a character is used in a group chat, but the rest of the prompt will be ignored.

Hmm... I'm trying to visualize where that'd kick in. I feel like you don't necessarily have to ignore the entire prompt. Like looking at App Default, most of it would be fine. I feel like it'd just be a matter of pre-pending the character name into those assistant messages in a group chat. Which doesn't look to be an area the user can mess with in the modular prompt editor anyway.

[ZeriLeandra]

I decided to try that "Completion Object" option (never had before). That formats it like this. Doesn't change the content but makes an additional tag. Hmm... pardon me while I go see how that works. Probably just fine I suspect. But the GroupChatEntry were to essentially add the character name as another item sent like this, that'd probably work! And hopefully be less of a lift?

	{
        "role": "assistant",
        "content": "Lux pushed off from where he'd been leaning against the wall, a grin spreading across his face as he sauntered forward.",
        "name": "Lux"
    },
	{
        "role": "user",
        "content": "\"So is engineering,\" I replied to Tor, trying not to smile at the brothers' amusing 'grumpy meets sunshine' dynamic.",
		"name": "Regina"
    },
	{
        "role": "assistant",
        "content": "A flicker of genuine surprise crossed Tor’s features—quick as a spark before it was banked again—his pointed ears twitching involuntarily at the mention of *circuit panels* and *hazing*. \"Dwarves,\" he muttered, almost to himself.",
		"name": "Tor"
    },