AI Assistant

A rehearsal / preparation chat: try prompts, see how each integration tool actually behaves, inspect what data about you the AI can read — without touching a real Contact conversation.

What can the AI see?

• Everything the app stores about you, except chat messages. Same content as the Show my data export (profile, linked platform accounts, mediator configs, enabled integrations, and so on). Chat messages are deliberately excluded — keeps the window responsive and private.
• All your connected integrations. Every action of every Composio toolkit you've linked. Unlike the mediator, there's no per-contact allow-list here — every tool is on by design.

What is NOT persisted?

• Assistant conversations never enter a Contact's chat. They live in a separate table (assistant_turns), visible only to you.
• Reactions you place on messages, model switches — session-local too.
• Clear history wipes everything.

Choosing a model

Two independent model pickers:

• Chat model — the model that produces replies. Switch any time; the next turn uses the new one. Doesn't affect your mediator's saved model.
• Compact model — used only for history summarization once the conversation grows long. Defaults to the same model as chat, but a cheaper one (Haiku, gpt-4o-mini) is usually more than enough.

Both use the same provider and same API key you configured in the AI Provider section above. No second key needed.

Compaction (history summarization)

If the conversation goes over 20 turns or 20 000 characters, the AI summarizes the older part into 4–8 bullet points and keeps only the last 6 turns raw in context. Automatic, always uses the compact model. The summary stays visible in the chat, marked as a "conversation summary" bubble, so you can see what the AI still knows about older turns.

Editing (edit-checkpoint)

Editing one of your own messages discards every message that came after from the visible history (technically marked with superseded_by_id, not deleted — the rows stay in the DB), and the AI re-answers your new text. This is the same checkpoint behaviour as ChatGPT's "edit message": use it to try a different prompt variant from the same point.

Reactions

If you react to an AI reply with an emoji, the next call sees a note that "the user reacted with X to your previous message". So instead of typing "try again differently", a 👎 does the same.

When does the AI start chatting?

Only when you send something first. No auto-greeting, no "try this" suggestions — the page starts blank so it's a clean slate every time.