Connections and context

In long-context work, the hard part is controlling what the model reads when it answers. On the canvas, connections are that control. A connected chat reads what flows into it, so a synthesis chat can read three sources while the side question next to it reads none of them.

It also keeps threads from polluting each other: connect a lecture PDF to three chats, one per exam topic, and none of them sees the others' conversations.

Handles

Every node has a handle on its right edge, and chat nodes also have one on their left. The right handle gives context; the left handle takes it. (Image nodes have a left handle too, but it only accepts other Image nodes.) Handles appear when you hover a node, and connected ones turn orange.

• Connect: drag from a node's right handle onto a chat node's left handle. Dropping on empty canvas instead creates a new, already connected chat (see Adding nodes).
• Remove: click the connection line to select it, then press Backspace or Delete. The toast that appears offers Restore.
• Move: drag either end of a connection onto a different handle.

How context flows

Context flows left to right and follows chains. If A connects to B and B connects to C, the chat in C reads both A and B. A node that reaches a chat by more than one path counts once. Context is gathered each time you send, so an edit to a connected note is picked up by your next message.

A branch is a new chat connected to the one you forked, so it reads the whole parent conversation. See Branching.

Node	What a connected chat reads
Chat node	Its full conversation
Text node	Its text
Web node	The captured page text
Post node	The transcript, when one is available
Document node	The document's text. A PDF is sent as the file itself until its conversion finishes.
Image node	The image itself
Scratch pad	Your drawing, as an image. An empty pad adds nothing.
Group	Every node inside it (see Groups as context)

In Image mode, connected Image nodes also become candidate reference images. See Image generation.

Large inputs

Nothing is summarized or trimmed: every connected node is sent in full with every message. Enough big sources can exceed a model's context window, and the request fails; disconnect what the chat doesn't need, or pick a model with a larger window.

Document text larger than 512 KB is left out (the AI is told it was skipped). An Office file that's still converting is skipped with a notice such as "'Report.docx' is still being processed and wasn't included in context."

Note: RAG mode, which sent only matching excerpts of connected content, has been retired. Connected content is now sent to the model directly.

Context modes

Every chat node has a context mode that decides what it reads beyond its own conversation.

• Summarized context (the default). A chat with no connections gets a short summary of every other node on the canvas: titles, previews, each chat's latest messages, and how nodes connect. In Agent mode, the AI can open the full content of any node that looks relevant. Once you connect something, the chat reads its connections in full instead.
• Isolated context. The chat reads only its own conversation: no canvas summary, no connections. Use it for a question unrelated to the canvas, or a critique you don't want anchored by your notes.

To switch, hover the chat's left handle and click the icon inside it (the lightning icon is Summarized context; the tooltip names the current mode). You can't switch while the chat is generating.

The switch only appears while the chat has no incoming connections. Once something is connected, the handle shows the number of direct connections instead, such as "2 incoming connections".

Note: An isolated chat keeps ignoring connections you add later, and a branch copies its parent's mode. If a chat's left handle stays gray after you connect something, it's still isolated: remove its incoming connections, switch it to Summarized context, then reconnect.

Mentions create connections

Type @ in a chat's message box and pick a node from the "Mention a node" menu. The chat connects to that node right away, as if you had dragged the connection, and the connection stays after you send. Deleting the mention from your message doesn't disconnect the node; remove the connection on the canvas.

Groups show their size in the menu, such as "group • 3 children", and mentioning one connects the whole group. See Chat node.

Add context from the navigator

The node navigator, the column of short lines on the left edge of the canvas, lets you wire in sources that are far away. Click a chat to make it active, then drag another node's line to the right ("Drag right to add as context"). Connected lines turn orange; drag one right again to disconnect it ("Drag right to remove context").

Groups are listed at the bottom with their child count. Click or drag a group's line to add or remove the whole group. See Navigating the canvas.

Groups as context

Drag from a group's right handle to a chat, or mention the group, and the chat reads everything inside it.

• Each node inside contributes as it would on its own, in reading order: top to bottom, then left to right.
• It stays live: nodes you add to or move out of the group are picked up by your next message.
• Only nodes inside the group count. Sources connected into a grouped node from outside aren't included.
• A node that arrives both directly and through the group counts once, and a chat inside the group never reads itself.

See what a chat reads

Click a chat node to highlight every connection feeding it in orange, all the way up the chain. While the chat generates, light travels along those same connections.

Edge styles

Choose how connection lines look under Edges in Canvas settings (the gear icon in the canvas header): Default, Straight, Step, Smooth step or Simple bezier. The style is purely visual and applies to every canvas.

Related

• Branching
• Group
• Agent mode and tools