Skip to content

Chat

Everything here works in both running modes unless a page says otherwise. Anything an admin can share or lock is noted on the page for that feature.

Text, vision and reasoning models, with replies streamed as they are written. Answers are rendered as Markdown: syntax highlighting on code, KaTeX for maths, and a copy button on every code block.

A message you sent can be edited, and one you received can be retried. Both rewrite the conversation from that point rather than appending, so a wrong turn is removed instead of buried.

Copying works at three sizes: a single message, a single code block, or the whole conversation as JSON or as Markdown.

One menu in the composer, Add context, attaches everything: knowledge, images and documents. Each attachment becomes the same removable pill, on the home screen and halfway down a conversation alike.

Pictures have two shortcuts past the menu: paste one, or drag it onto the composer from the desktop. PNG and JPEG only, and anything else in the same drop is counted back to you rather than silently dropped.

What each kind of attachment does once it is in there is on its own page: Knowledge for text you keep, Documents for a PDF or a spreadsheet read in the browser.

System prompts apply globally, per model, or to one conversation, and the most specific one wins. Prompts covers the resolution in full, including what an admin can impose.

A conversation names itself after the first reply, using whichever model you point at the job. It is a small, cheap request to a model of your choosing rather than the one you are talking to.

Commands and shortcuts for what the composer and the keyboard understand. Search for finding a passage across every conversation. Compaction for when a conversation stops fitting. Generation for where a reply is actually written, which decides whether reloading the page costs you one.