[Idea]: Generating AI generated notes

[FlorianHeigl]

What are you trying to do?

successively adding more papers on a topic as new research is published

What feels difficult or missing today?

The AI based findings stay the same, even if the information base shifts

What outcome would help?

Being able to have trigger a re-run of the same query with an updated number of sources / (probably also useful to be able to do this with an updated toggle switch between full/insight based generation)

How do you handle this today?

I'll copy my original query into a new notebook search, and later probably delete the old one.
I strongly suspect a historic perspective is very valuable, so I should not really delete them at all.
"Archive"?

Additional context, examples, or possible directions

No response

How would you like to participate?

• I can add use cases or join the product/design discussion.
• I can help explore the technical approach.
• I can help implement an approved work item.
• I can test or validate a proposed solution.

Before posting

• I searched existing Discussions and Issues for related proposals.

[lfnovo]

Let me play this back with a concrete example — correct me where I get it wrong:

In March, your notebook has 12 papers. You ask "what does the evidence say about Y?" and save the answer as a note. By June, 7 new papers are in. The March note is frozen — it doesn't know the base shifted, and the question that generated it isn't even stored with it. So today you copy the question by hand, run it again, and (against your own instinct) delete the old one.

If that's right, I read two coupled needs:

1. Re-run: an AI-generated output should remember its recipe — the question, the scope it ran against, the generation settings (full vs. insights) — so "run again with today's sources" is one action, not copy-paste archaeology.
2. History: each re-run becomes a new version and the old ones stay. You called it yourself — "what did the evidence say in March, with 12 papers" vs. "in June, with 19" is itself research data. Overwriting would destroy exactly what makes successive re-runs interesting.

This aligns with where the platform is heading: outputs that know what they were generated from (provenance) is a direction we're actively investing in — and once an output knows its inputs, staleness becomes detectable ("generated from 12 sources; the notebook now has 19") and regeneration becomes a first-class action instead of a workaround.

No commitment on timing yet, but to sharpen the design:

• Is the unit a saved chat/search answer, a note, or "anything AI-generated"?
• Should the system tell you when an output has gone stale (N new sources since generation), or is a manual trigger enough?
• When you re-run, how do you want to consume the history — side-by-side compare, an archived timeline, a diff? (You raised comparing in [Idea]: manual annotations #1215 too — curious whether it's the same mental model.)

Status: exploring — core need framed as re-runnable, history-preserving generated outputs.

[FlorianHeigl]

• unit: a note created from chat/search answer, I would not include the document search from the non-notebook search options. it's very intuitive that those search the currently available dataset. maybe automation users will differ on this, but i would say start with the most focussed and most valuable example. not just valuable to me but to the system design itself - notes are the one object that gains persistence.
• manual trigger is enough, I suppose it would be best working IF a 'last run at' is shown and sortable.
• i can't answer on the best presentation/consumption, at least as of now. I looked at [Idea]: manual annotations #1215 again and that is probably an argument pro a side-by-side views.

Note I'm not sure if I can supply a better answer for the history representation ever, I got some cognitive trouble that heavily impacts visual thinking. I'm gonna flip a few UX design books and try to come to a good example. If I don't post back on that item though, don't wait for me. It might turn out impossible.

[lfnovo]

That's enough to write it down properly — recording the spec here:

• Unit: a note created from a chat or search answer. Agreed on the reasoning — notes are the one object that gains persistence, so they're where "remembering the recipe" pays off first. Document search stays live-only.
• Trigger: manual, with last run at visible and sortable. No staleness nudges in v1.
• History: side-by-side as the working assumption, informed by [Idea]: manual annotations #1215.

On the history representation: you've given more than enough. We'll take it from here and won't wait — if a design book turns something up, it's welcome, not owed.

Timing: this sits inside a broader direction — generated outputs that remember what they were generated from — which I'll be writing up publicly soon. When that lands I'll link it here, and this thread stays the home for the note-specific spec until it's scheduled.

Status: incubating — spec settled, scheduling open.

[lfnovo]

As promised: the broader direction is now public — #1315. Your case is in it twice: generated outputs that keep the evidence they were built from (which is what makes "re-run with today's sources" a first-class action), and question 6 on which output proves the infrastructure first. This thread stays the home for the note-specific spec; status unchanged (incubating).