pAIring chat and channels

pAIring has two kinds of conversation:

• Chats are private. Only you can see or list them.
• Channels are shared. Every member sees the same conversation, and the AI can join it.

Most options below work the same in both. The differences are listed in Chats vs channels.

Sending a message

• Press Enter to send. Press Shift+Enter for a new line.
• In a chat, every message goes to the AI.
• In a channel, Send posts a normal message to the other members, with no AI reply. Press Ask AI to ask the AI. See Channels.
• You can stop a reply while it is streaming.
• Math in replies is rendered as formulas.

Chats get a title in the sidebar. An admin chooses how: Simple builds it instantly from your first message, and Smart asks the model to invent one, which reads better but costs a second generation.

Choosing the model

Each chat has its own model, and you can switch it at any time from the chat header. New chats start with the default model. If an admin disables a model, a chat that uses it asks you to pick another one.

Switching the model or engine in an existing chat

A model does not remember a conversation between messages. Every reply needs the whole chat: your notes, any knowledge-base passages, and every message so far. Local engines keep a processed copy of recent chats in memory, so the next message only has to read what is new. Matricxon also saves these copies to disk (see Prompt cache).

That copy belongs to one model on one engine. When the model changes, the new model starts from nothing:

• Switching a chat’s model makes the next reply read and process the whole conversation again. On a CPU-only server, the first reply in a long chat can take minutes before any text appears. Later replies are fast again.
• An admin switching the active engine (for example, from Matricxon to a cloud connector) does the same for every chat and channel: each one’s first reply after the switch processes its whole history again on the new engine.
• Switching back can be quick if the old model still holds the chat’s copy. Copies are dropped over time to free memory or disk, so this is not guaranteed.
• Cloud connectors, such as Cloudflare Workers AI, keep no copy in pAIring. Every message sends the whole conversation, so each reply’s input grows with the chat and counts toward your usage, whichever model you pick.
• A model with a shorter limit may see less of a long chat. When a conversation no longer fits the context window, or the new model’s own limit, the oldest messages are left out of what the model reads. They stay in the chat for you.

The same applies to channels, for everyone in them: when a manager or admin changes a channel’s model, the first reply after the change processes the whole channel history again.

Tip

To use another model for a long conversation, starting a new chat is often faster than switching, because the new model then has little to read.

Generation settings

You can tune how the model writes, for one chat or for all your chats. The panel asks you with a checkbox: “Only for this chat, leave unchecked to update this for all your chats”. Use default goes back to your defaults, which you also set under Settings → Behavior.

Setting	What it does	Range	Default
Temperature	Higher is more creative, lower is more predictable	0 to 2	0.8
Top-P	Limits choices to the most likely words that add up to this share	0 to 1	0.9
Top-K	Limits choices to the K most likely words	1 to 200	40
Repeat penalty	Discourages repeating itself	0 to 2	1.1
Reply length	Maximum tokens in a reply (-1 for no limit)	-1 to 8192	1024
Seed	A fixed number makes replies repeatable. -1 is random	any	-1
Document search results	How many passages the knowledge base adds	1 to 10	4

The context window (how much of the conversation the model can see) is set by an admin for everyone. See Admin settings.

Notes: persona, rules and skill

A note is a text that shapes how the model behaves. You write notes on the Notes page and pin each one to a chat, or for all chats, as one of three roles:

• Persona: who the model is.
• Rules: what it must follow.
• Skill: what it is helping with.

Pinned notes are added to the start of the conversation. You can edit them from the chat header. For one chat you can turn a role off, or go back to your default note with Use default.

Each user gets 3 default notes. Rule notes, at the top of the Notes page, sets whether your saved persona, rules and skill notes are added to new chats and channels automatically. It is off by default and only affects chats you start from then on.

Attachments

You can attach files to a single message with the paperclip. Each file shows as a chip before you send, and the × removes it.

Slow without a GPU

Working with documents and images on a Matricxon node with no GPU can be extremely slow. The model has to read every attached document or image before it can answer, and long files make replies take much longer. Use short files, or run the engine on a Matricxon node with a GPU.

	Documents	Image
Types	.txt, .md, .pdf, .docx, .odt, .rtf, .html, .csv, .json	.png, .jpg, .jpeg, .gif, .webp
Limit per message	5	1
Size limit	10 MB each	10 MB
How it is used	The text is added to that message’s prompt for the chat’s model	Sent to the default vision model for that one reply

• You can combine up to 5 documents and 1 image in one message.
• If any file is not allowed, the whole batch is rejected with a message that names the limit.
• An attachment belongs to its message. It is read once and is not added to the knowledge base.
• For the image, the chat’s own model is never changed. Only that reply uses the vision model. An admin sets it under Settings → System → Default models.
• Reading an image is much slower than text on a CPU-only Matricxon node, so image replies have their own, longer timeout.

Knowledge base (documents)

Documents in the knowledge base are searched automatically in every chat and channel. The AI answers using the passages it finds and shows Sources under the reply so you can see where an answer came from.

• Add your own documents under Settings → Knowledge. They are private to you.
• Admins can also share documents with everyone.
• An admin can switch the knowledge base off, or set the file size and storage limits.

Slow without a GPU

Working with the knowledge base on a Matricxon node with no GPU can be extremely slow. Adding a document means turning it into searchable data with the embedding model, and every search adds the passages it finds to the prompt. Large documents, and many of them, take a long time on a CPU. Start with a few small documents, and a GPU makes a big difference.

The difference from attachments: the knowledge base is permanent and searched on every message. An attachment applies to one message only.

How the search works, every setting that affects it, and how to make it faster or more accurate: Knowledge base (RAG).

Tools

Tools let the model take an action or look something up, such as searching or calling a service. They come from MCP tool servers that an admin connects.

• Turn on the Tools switch in the message box. A small panel lists the tools available.
• When the model uses a tool, you see each call in a block at the top of its reply, with the result.
• The model can use tools for several rounds in one reply. An admin sets the limit (5 by default, up to 20). After that, it must answer without tools.
• A very long tool result is cut before the model reads it. An admin sets that limit.
• A chat keeps the tools it started with, so an admin’s later changes do not make the model re-read the whole chat. Use Update tools to pick up the current list.
• If an admin removes a tool, a call to it is answered with “no longer available”.
• An admin can turn tools off for everyone.

Tools work with Matricxon (Qwen, Mistral, Llama 3.x and Gemma 4) and with llama.cpp models whose chat template supports tools. Only connect servers you trust, because whatever a tool returns is shown to the model.

Managing messages and chats

• Delete a message with its delete button. Delete chat removes the whole conversation.
• Share to channel: on a private chat message, you can share it to a channel. The channel gets a plain copy posted as your own message, with a note that it came from a private chat. There is no AI reply and no link back to your private chat.
• If you close the tab while a private reply is being written, the reply stops. If you reload the page while a reply is streaming, the chat reconnects and shows the rest.

Channels

A channel is a shared conversation, such as a team stand-up or a class. An admin creates it and chooses its members.

What each person can do

	Member	Manager	Admin
Read and send messages	Yes	Yes	Yes
Ask the AI	Yes	Yes	Yes
Attach files	Yes	Yes	Yes
Change the channel’s model, settings, notes and tools	No	Yes	Yes
Add or remove members, rename, reset or delete the channel	No	No	Yes

A channel needs at least one manager. Managers must be members and have the Channel manager role.

Normal messages and Ask AI. Press Send to talk to people. Press Ask AI to get a reply. The AI answers one question at a time, so Ask AI is greyed out while a reply is being written. Whether the AI also reads the normal messages when it answers a later question is an admin setting.

Seeing replies. An admin chooses how other members see a reply. Cheap checks every few seconds. Real pushes the reply to everyone live, token by token. Live delivery works reliably when pAIring runs as a single app instance.

Replies always finish. In a channel, a reply is saved even if the person who asked leaves the page.

Reset channel. An admin can wipe a channel’s history and set its notes fresh. The channel and its members stay. This cannot be undone.

Chats vs channels

	Chat	Channel
Who sees it	Only you	All members
Send	Always asks the AI	Send posts normally, Ask AI asks the AI
Model, settings, notes, tools	You decide	Admins and managers decide
If you leave mid-reply	The reply stops	The reply finishes
Created by	You	An admin