How-to

Generate and weave with hosted AI

Two different operations that sit next to each other. Generate writes new nodes; Weave connects the ones you already have and never adds any.

Before you start

Hosted AI switched on— account menu → AI & Privacy. Off by default.
A plan with units. The free plan has none, so these decline rather than charge. Generation says it isn't configured if you are hitting that.

If what you want is bulk nodes and you would rather not pay twice, the free lane does the same job with a model you already subscribe to — Bring your own AI. This page is for when you would rather the app made the call.

Never confuse the two


✦ Generate	You give it notes, a conversation, or a request in prose. It writes new nodes into the lens. Costs scale with how much you put in.
⟲ Weave	You give it nothing. It reads nodes you already have and proposes connections between them. It never creates a node. Costs scale with how many nodes it looks at.

The practical test: if your lens feels thin, Generate. If it feels like a pile of unconnected notes, Weave.

Generate

Paste in the raw material — notes, a transcript, a chunk of reading — or just describe what you want. Before it runs you pick a mode, and the two differ in behaviour rather than in warnings:

Mode	What it does
Grounded	Leaves a claim out rather than guessing at it. Fewer nodes, all of them ones the model will stand behind. The default, and the right one for anything you intend to learn from.
Exploratory	Sketches more of the territory. Useful for mapping a subject you are still circling — and its unverified claims cannot become flashcards.

If you have rejected nodes from a previous run, the dialog says so and avoids those titles. That is a hint worth reading: five rejections usually mean the request was wrong rather than that the model was.

Weave

Weave runs on one node or across a whole lens. The lens version has a plan step before anything is spent, and it is not a confirmation ritual — it tells you three things you would otherwise only learn afterwards:

What it will look at, by name.
How much that costs. One unit per node.
What is excluded, and why.

The plan is a pure calculation over one query, so it is free to look at. There is no reason not to.

Restricted nodes are never included in a lens weave, and they are not being skipped as a nicety — they can never be sent to a hosted model at all. If you want one woven, open it and weave it individually, which sends it deliberately rather than as part of a batch you did not inspect. What leaves this app.

A lens weave runs in batches and can be stopped part-way; what has already run is kept, and what has not is still eligible. There is an undo on the whole run.

What a unit is

A unit is a measure of input, not of output — because input is the thing you can see and estimate before clicking, and output length is not under your control.

Operation	Cost
Generate	One unit per 1,000 words of input. Minimum one.
Weave, one node	One unit.
Weave, a lens	One unit per node it looks at. Priced honestly — it is that many calls, so a flat batch discount would mean the meter stopped predicting spend exactly where spend gets large.
Asking a question of your graph	One unit.
Cards for one node	One unit.
Drafting a reading of an entry	One unit.

Rounding is up. A 1,001-word input costs two. That is deliberate and it is stated rather than buried: rounding down would let the meter under-report real spend, and rounding up makes the number on the button an upper bound — so the charge can never exceed what you were shown.

The estimate and the charge are the same function call, not two calculations that ought to agree. A button that said “~40 units” and then charged 60 would destroy the point of showing an estimate at all, so the two cannot disagree by construction.

When a call fails

If the model errors, or returns nothing usable, nothing is charged and the message says so. Your allowance is only reduced by work that produced something.

Everything lands as a draft

Whatever comes back arrives grey, rounded, and barred from becoming a flashcard until you have opened it. There is no setting for this and no way to answer your way out of it — when the app made the call, no claim about authorship could make the text yours.

Woven connections carry the same posture one level down: an edge a model proposed is marked as provisional rather than silently mixed in with the ones you drew.

Drafts, and where a node came from is the full model.

Bring your own AI — the free lane, which does the Generate job for nothing.
Typed relations — what a weave is trying to produce, and how to judge whether it got it right.
What leaves this app — scope, the floor, and what is kept.