Notes about your tables

Where the server turns them on, with the schema finder, the DataSquares AI assistant keeps short notes about the tables of your data models: what it learned while it found the tables a question needed. A later question on the same tables starts from them, so the assistant finds what it needs sooner and makes fewer of the same mistakes. Nothing changes until the server turns them on.

What a note says

About a table: what one row means (one order, one order line), its key, the date column to use, the filters it usually needs, and what to watch for, such as includes cancelled orders.
About a column: what each of its codes means, such as O = open, X = cancelled, for the codes value hints keep.
About a join: what a check of your data found: the join holds, it repeats rows (so amounts are added up before the join), or it can’t be relied on.
About a question: a question like yours, with its values taken out, and the tables, columns, joins and measures that answered it.

A note the AI learned says where it came from: checked against your data, learned from an answer, or written by you.

When a note is kept

A note is kept only after an answer you were happy with:

you rate it with a thumbs-up,
you save it to a dashboard or a report, or
you carry on with your next question without correcting it.

The notes are made after the answer, from what the schema finder did for it: the answer never waits for them. Without one of these, nothing is kept. A correction drops what that answer would have taught; a thumbs-down drops it too, and puts in doubt the notes the answer used, and any it had already taught: they are not used again until they are learned again.

What a note never holds

No values from your data, except what a code means, for the codes value hints keep. When value hints no longer keep a code (they are turned off, or the field is labelled sensitive, hidden, left out of the AI scope or covered by row-level security), a note about it is no longer used, and turning value hints off deletes it.
Nothing outside the AI scope. A note is used only on tables and fields you may see, and one that names a field labelled sensitive reaches only the people who may read labelled values in clear.
Nothing the rule for notes turns away. Each piece is checked as every note is: text with personal or secret details, a data value or an instruction to the AI is not kept, never cut down to fit.

Who uses a note

Yours alone, at first. What the AI learned from your answers is used for your questions only. It applies to others only once the data model’s owner (or a workspace admin) approves it in Prep for AI; an owner approves only while their role can edit data models. They see what the note says and its table, never whose answer it came from; they may approve it as it is or in their own words. If they reject it, no one uses it any more, you included.
Facts a check proved are shared. That a key is unique, that a join repeats rows or holds: these are shared with your workspace at once. They hold no value, and everyone sees them only on tables they may see.
A note you rewrite is yours, in your words: the AI never replaces it with one of its own. It waits for the model owner’s review as the AI’s notes do.
Explore this model in Prep for AI has the AI learn a whole model’s tables at once, on a model editor’s click; the notes it infers are that editor’s until the model’s owner approves them (details).
While you pause your memory, your own notes aren’t used (the workspace’s shared ones still are) and nothing new is kept; an incognito chat uses them, but keeps nothing and records no use.

How notes stay true

A change to a table (its columns, keys, relationships or what the AI scope leaves out of it) puts its notes in doubt: they are not used until they are learned again. Removing a table or a field removes its notes.
Unused notes fade. A note counts for less the longer nobody uses it (by half every 90 days), and one unused for 180 days is retired, never your own words. A retired note is deleted 30 days later, and a note no answer was happy with, after 7 days.
Facts are checked again. A fact a check proved that answers use often is checked against your data again now and then, within the checks the answer’s effort allows; when the data no longer agrees, the note is put in doubt.

The first question on new tables

When the tables a question points to have no notes yet, or a thumbs-down put theirs in doubt, the schema finder looks further the first time: two more steps, more of the tables, and more checks of your data (2 at Quick, 5 at the other efforts), only as far as the question needs. You see Learning this model’s tables while it does. It is decided for each part of a model: a model well known for sales can be new for inventory. The answer never waits for the learning, and the extra work is part of the answer’s one charge, in AI credits.

See which notes an answer used

Under an answer that used notes, Used 2 notes about your tables opens the list: each note, where it came from, and whether it is yours or shared with your workspace.

Read, rewrite and delete your notes

Your notes are on the memory page, under Settings → Memory → Table notes (Settings → AI notes opens it there), the most recently used or changed first: what each says, its data model and table, where it came from, whether it may be out of date, when an answer last used it, and what the data model’s owner or a workspace admin decided:

Waiting for review: only you use it. Edit rewrites it in your words: text the rule for notes turns away is refused whole, with the reason, and the note stays as it was. Delete removes it, and the AI stops using it at once.
Approved: shared with your workspace. It is the workspace’s now, reviewed in Prep for AI, so you no longer edit or delete it here.
Not approved: no one uses it any more. Delete removes it now; otherwise it is deleted 30 days after the decision.
Removed or Replaced: it was shared, then the model’s owner or a workspace admin took it out of use, or approved a newer note about the same thing in its place. No one uses it any more; it is deleted 30 days later.

When the model’s owner or a workspace admin reviews your notes, you also get one notification with the counts, saying which of them decided.

The notes Explore this model drafted when you started it are listed here while they wait for review; once decided, they are followed in Prep for AI, with no notification. One you rewrite is yours, in your words: you see here what was decided about it, and are told.

The facts shared with your workspace are not listed there. AI notes (or the Table notes tab) shows while the AI keeps notes, or while you still have some.

Turn them off

Notes are kept and used only where the schema finder runs: a company administrator who turns AI search off under Administration → AI turns them off too. Turning Value hints off stops the checks of your data and deletes the notes about codes. Your own notes are erased with your AI data (removing what is kept).

AI search on large models — the schema finder and value hints.
How the AI uses your data — what the AI is sent and what DataSquares keeps.
The AI scope — choose the tables and fields the AI may use.
AI controls — turn AI features on or off.