Working with messages
Each message carries a row of actions. Replies can be copied, rated, read aloud, and regenerated; your own messages can be edited or branched from.
Actions on a reply
| Action | What it does |
|---|---|
| Copy | Copies the reply text to your clipboard. The icon becomes a checkmark on success, or a cross if the browser blocked clipboard access. |
| Good response | Marks the reply as good. This sends the exchange to Rootcomputer — see below. |
| Bad response | Opens a short form where you can optionally say what should have been better. |
| Read aloud | Speaks the reply using the selected voice. See Voice and audio. |
| Retry | Discards the reply and generates a new one from the same prompt. Only available on the most recent reply. |
| More | Additional options for that message. |
Retry is deliberately limited to the latest reply. Regenerating something in the middle would invalidate every turn after it, so to redo an earlier point in the conversation, branch from it instead.
Actions on your own messages
- Copy — copies what you wrote.
- Edit message — rewrites your message in place and regenerates the reply from the corrected version. Everything after that point is replaced.
- Branch into new chat — copies the conversation up to that message into a new chat. See Branching.
If a reply misses because your question was ambiguous, editing the original message usually works better than asking again. A follow-up leaves the confused exchange in the history where the model can keep tripping over it; an edit removes it.
Long messages
A long message of your own is shortened in the transcript with a Show more link rather than filling the screen. Expanding it shows the whole thing, and Show less collapses it again. This is display only — the model always received the full text.
Thinking
On some replies a small Thought for… line appears above the answer. Expanding it shows the model's reasoning before it committed to a response.
Whether a model thinks first depends on the question and on your settings. See Response controls to turn it on for every reply.
Sources
Epos can consult a document collection, or search the web when you ask it to, before answering. When it does, a disclosure line appears above the reply and expands to show what it read.
Most replies show no line at all, which is normal — lookups only happen when there is something worth looking up. Full detail is on Retrieval and web search.
Rating replies
The thumbs on each reply are how Rootcomputer finds out what the models get wrong. Ratings directly inform training.
Unlike your conversations, which stay in your browser, feedback is sent to Rootcomputer. When you rate a reply, the surrounding conversation turns are transmitted and stored so the rating can be interpreted — a rating with no context is not useful for training.
This only happens when you press a thumb or answer the comparison prompt. Simply chatting uploads nothing beyond what is needed to generate the reply. Do not rate replies containing anything you would not want stored. Full detail is on Your data and privacy.
Comparing two replies
Occasionally — roughly one eligible turn in six, and never at the very start of a conversation — ChatHaiku generates a second reply and asks which you prefer.
This is the single most useful signal for improving the models, because it captures which of two plausible answers a real person actually preferred. Like the thumbs, answering it uploads the exchange.
Formatting and length
Messages are limited by word count, and the limit depends on the model — 2,048 words on the Epos models, 200 on Haiku H2. Going over does not truncate silently: the message is refused with a note giving the limit and your word count. Pasting a long block lifts it into an attachment chip rather than filling the text box; see Pasting long text.
Reply length is governed by your response settings rather than by a fixed limit — see Response controls. If replies are consistently shorter or longer than you want, that is the page to visit.