Chat

Scelo's AI is not one chatbot — it's a set of scoped assistants that appear where you're working, each framed for that context.

Where the chats are

Chat	Where	Scoped to
Stage chat	Bottom of each workstation (Soft / Tools / Hard)	That stage's job
Macro-node chat	On each card in the macro view	That stage, briefly
Column chat	Hover a column header in Soft Data	That one column
Model chat	On each model node in Tools	That model
Result chat	In a model's detail dashboard in Hard	That result

Each is framed so the assistant stays useful and on-task — the Soft chat won't jump ahead to model choice, the Hard chat won't re-open data collection, and so on.

Deterministic local commands

Some requests are handled client-side, instantly, with no provider call — so they work even fully offline. In the Soft Data chat:

• clean my data / do the initial cleaning — runs the cleaning plan.
• make the dates american (or european / iso) — reformats date columns.
• add 1000 rows through augmentation — bootstrap augmentation.

In a column chat:

• make this american, remove all non-dates, clean this column.

Anything that isn't one of these falls through to the normal AI chat.

Rendering

Replies render as markdown: inline code, lists, tables, and math (KaTeX). The Soft/Hard chats can also render embedded viz blocks — the assistant returns a small chart or stat-table spec that Scelo draws against your current dataset.

Provider

The chat uses whichever AI provider is active — Ollama by default. In the desktop app the request goes directly to the provider from the main process.