Skip to content

Chat

Jump to bottom
overcatbe edited this page Nov 3, 2026 · 4 revisions

Chat

The Chat tab is AIcove's own chat, no browser involved. On the left your personas and the chats of the one you picked; on the right the chat.

  • Assistant is a plain persona with no prompt, knowledge or facts, on whatever model is loaded: for trying a model as it is. Delete it if you don't need it; it doesn't come back.
  • Balthazar the Unsupervised is an example of what a persona can be: an ancient, grumpy wizard who holds grudges. Change him on the Personas tab, or delete him; he doesn't come back either.
  • Enter sends, Shift+Enter makes a new line. Ctrl+N or New chat starts a new one.
  • At the top: the model the persona runs on. A persona names its models (models: in its file, set on the Personas tab); the list shows the ones this machine has. If another model is loaded, sending loads this one first (that stops the other). Load does it without sending.
  • Stop ends an answer early; what came so far is kept.
  • The box you write in can be made taller: drag the line above the Style menus (in the Writing tab, the line above the box). AIcove keeps the size.
  • Under each message: edit, delete, and for the last answer answer again. Editing your own message offers Save and answer again: everything after it goes, and the persona answers anew.
  • answer again keeps the answer it replaces: ◀ 2/3 ▶ under the last answer switches between them, and the one shown is what the chat goes on with. When you write your next message, the others are dropped.
  • again with… does the same on another model, picked from a menu (it loads it if needed). Each answer says which model wrote it.
  • 👍 / 👎 under an answer (the chosen one is shown bigger; click it again to take it back) are counted for the persona on the answer's model; the Personas tab shows the counts under Settings for:, to see which model suits it.
  • Picture… (or drop or paste a picture into the box) sends a picture with your message, for a model that can see: a launcher with an eyes file (mmproj), marked 👁 wherever models are listed. If the chosen model can't see, AIcove offers to load one that can; Send anyway tells the model a picture is there without showing it. Big pictures are made smaller first (1536 px at most). They are kept in the persona's chats/pictures/ folder; click one to open it.
  • ★ keep under an answer puts it into the persona's Best of… (the button at the top). To keep just a line, select it and use the right mouse button: ★ Add to … best of. It is a plain file, best-of.md in the persona's folder, with the date and chat of each entry; edit it in Best of… or any editor.
  • After each answer: how long it was and how fast it came, e.g. 312 tokens · 18.4 t/s (thinking counts too). Handy for comparing models and launcher settings.
  • ▸ Thinking opens what a thinking model considered before answering. It is kept with the answer but never sent back to the model.
  • While an answer comes in you can scroll up and read; the chat follows the answer only while you are at the end. If it finishes while you are further up, the persona's picture shows a ↓: click it to go to the start of the answer.

Every chat is a file in the persona's folder, ~/AIcove/personas/<persona>/chats/. Delete removes a chat for good. Rename… changes a chat's title; new chats are named after their first message.

  • A persona with file access shows its steps in the chat: ▸ Read /etc/fstab, ▸ Looked into its own folder and so on. Click one to see what it got back. Before it writes a file, AIcove asks you (see the topic Letting a model use your files).

Long chats. A model only takes in so much text (the launcher's context size). When a chat fills 70% of it, the older messages are summarised by the model itself, which takes a moment ("The chat is getting long — summarising…"), and from then on the persona gets that summary instead of them; the newest part stays word for word. The bar right of the menus, Context 5.2k of 32.8k (16%), shows how full it is (prompt, facts, knowledge and the chat; it turns orange from 70%). It shows while the persona's model is loaded. The chat on screen stays complete:

  • Summarise next to the bar does it now, before the 70%: the older half of the chat (at least) goes into the summary, the newer part stays. Handy before a long piece of work, or before switching to a model with a smaller context.
  • The summary shows as So far at the top of the chat. edit it to correct or add what matters; drop it and the persona sees everything again (and makes a new summary when needed).
  • A line marks where it starts: above this line … sees only the summary.
  • When the chat grows on, the summary is extended. Editing or deleting one of the summarised messages drops it, and the next message makes a new one.

What a persona remembers. Nothing carries over from one chat to the next by itself, except the persona's facts: short sentences in four sections, About you, Story so far, World and Style and rules. They go into every chat with that persona.

  • Remember… (top of the chat): the model reads this chat and proposes what is worth keeping. You tick, change or drop each one; only Keep the ticked ones saves anything.
  • Facts…: all of them, one tab per section, one fact per line, to edit or delete. Open in the text editor opens a section's file in your editor instead; Show files opens the folder with all four.
  • Each section is a file: ~/AIcove/personas/<persona>/facts/about-you.md, story-so-far.md, world.md, style-and-rules.md. Editing them by hand is fine.

Keep them short: every fact costs room in every chat.

Controls above the input set how this chat answers, over the persona's own settings. as set means the persona's (or its launcher's) settings. Each chat keeps its own; a new chat starts with as set.

  • Style: Factual (focused), Creative, Story (loosest) — the same presets as on the Launchers tab.
  • Thinking: on or off for thinking models (off answers quicker). Greyed out when the model's thinking can't be switched.
  • Length: the longest an answer may get: short, medium or long.
  • Keep for (shown when one of these three isn't as set): makes them the persona's settings on the picked model, on this machine, so every chat with it on that model gets them; the menus go back to as set. They show up on the Personas tab under Settings for:.

Knowledge is material a persona works from: notes, lore, books. Put the files into its knowledge/ folder (Knowledge… at the top of the chat opens it; the button shows how many files there are). Text, Markdown, PDF, EPUB, Word and LibreOffice files work; PDF, EPUB and Word are turned into text once and kept in knowledge/.text/. It works without file access.

  • Small knowledge (up to a quarter of the model's context) goes into every chat as it is.
  • Bigger knowledge, such as whole books, is too much for that. The persona then gets a list of the files and looks things up when a question touches them: ▸ Searched its knowledge for … and ▸ Read … in its knowledge show in the chat. This needs a model that can use tools; on one that can't, only the beginning of the knowledge is given.

Web above the input (off/on, per chat) lets the persona search the web and read pages: ▸ Searched the web for … and ▸ Read … show in the chat. It needs a model that can use tools; on one that can't, the menu is greyed out. Searching goes to DuckDuckGo, or to your own SearXNG if the web_search setting (Settings tab) is its address, e.g. http://127.0.0.1:8888. If that address is a tool of yours (Tools tab) that isn't running, the first search starts it. After many searches in a short time DuckDuckGo stops answering for a while; Wikipedia's search stands in until then. Pages on this computer or your home network can never be read, also not through a redirect, so a model can't reach llama-server or your router this way.

Voices. A persona with a voice (Personas tab, Voice) gets 🔊 listen under its answers: the answer is read aloud, without code and Markdown marks, and ■ stop ends it. The voice comes from a voice server (a tool on the Tools tab, started when needed, or an address; see the README for what it has to answer); its voices are the ones on offer (Get the voices on the Personas tab asks it). Making the audio can take a while, depending on the server and the machine; once made, it is kept (as MP3, if ffmpeg is there) in the persona's chats/audio/ and plays at once. Audio no message needs any more (the chat or message deleted, the answer edited, another voice) is deleted by itself. The voice setting names the tool, or the address of a voice server running elsewhere.

Pictures above the input (off/on, per chat) lets the persona make pictures: ask for one, or it decides a picture says more. ▸ Made a picture: … shows what it asked for (click to see the whole description), the picture comes under it; click the picture to see it full size. The image server (Tools tab, Pictures: stable-diffusion.cpp's sd-server) starts when needed, and a picture takes a while. Pictures are kept in the persona's chats/pictures/; deleting the chat or the message deletes them. The pictures setting names the tool, or the address of an image server running elsewhere (anything with the OpenAI image API). Needs a model that can use tools.

Its own face. A persona can look the same in every picture it is in. Train a LoRA on the Z-Image model with pictures of its face, put the .safetensors file in the LoRA folder (~/AIcove/loras/, Settings tab), and on the Personas tab pick it under LoRA with the word it was trained on (the trigger word). When the persona is in the picture itself, it describes itself as its instructions say and the picture is made with its LoRA; pictures without it stay as they were. The image tool must be named pictures and have --lora-model-dir "$AICOVE_LORAS" on its start line. The LoRA folder is inside the AIcove folder, so it moves and syncs with it.

Made from the guide inside AIcove (cove/data/guide.md); it opens with the ? button.

AIcove

Clone this wiki locally