Chat features¶
Everything you can do in a conversation beyond sending a message.
Streaming responses¶
Responses stream token-by-token. The status line narrates the pipeline — "Gathering perspectives…", "Reading knowledge…", "Actively learning…", "Composing…" — so you can see the machinery work. Messages you send appear live in other tabs viewing the same conversation, the connection reconnects automatically if it drops, and it is always safe to leave or reload mid-turn: the turn is queued and resumes when you're back. Failed turns show a Retry button on your last message.
Perspectives on demand¶
By default each perspective's short take streams above the synthesized response. Two toggles control the display:
- "Hide perspectives (reduces usage, same quality)" — skips generating per-message perspective evaluations. Your choice is remembered per browser.
- Show perspectives (per response) — retroactively generates the perspectives for a response that lacks them.
Each response also carries a contribution bar (how much each perspective weighted this answer) and a 3D subjective shape you can rotate, download as a GLB file, or open in a shareable standalone viewer. A Hide message shapes toggle is available if you prefer text only, and a Show thinking toggle exposes the reasoning trace.
Fork and revert¶
- Fork (on any agent response) — creates a new conversation containing a copy of everything up to and including that response. The original is untouched, and messages reverted in the source aren't copied. Explore a branch without polluting the main thread.
- Revert (on any agent response except the newest) — hides everything after that response, rewinding the conversation to that point. Reverting is idempotent, and reverting to the newest response is a no-op.
Ask mode¶
Ask (in the navigation) is a one-off question page. It answers against your project's accumulated knowledge — read-only: nothing it touches is written back to the knowledge repository, and nothing is stored in any conversation. It keeps the last few turns of its own context so short follow-ups work, but refresh clears it. No perspectives are generated, and usage is still metered against your plan. Ask is the only generative capability available to viewers.
Attachments¶
Attach a file to any message with the upload button:
- Accepted types:
.pdf,.txt,.md,.csv,.json,.xml,.html,.htm,.docx,.xlsx,.xls,.pptx - Max size: 10 MB per file
- Text is extracted from the file and prepended to your message as
[Attached: <filename>]context. A message may consist of just a file.
Conversation settings¶
Each conversation has a Project knowledge setting, changeable any time from the chat page:
| Setting | Behavior |
|---|---|
| Full | Reads knowledge before responding and learns from the conversation afterward. |
| Read-only | Reads knowledge, never writes back. |
| None | No knowledge access for this conversation. |
New messages in Full conversations also enqueue catch-up ingestion, so past messages are eventually learned even if an update failed.
Publish and share¶
Publish (on the newest agent response) makes the conversation readable at a public link — anyone with the link can read it, no login required. The message then shows a Public link button and an Unpublish button; continuing the conversation lets you publish again at the new end. See projects and teams for who can do what.
Export¶
Export (on the newest agent response) downloads the conversation as Markdown. Copy on any message copies the perspectives and response body to your clipboard.
Deleting conversations¶
The conversation list has a per-conversation delete. Deleting removes it from the product permanently; billing records are unaffected.