Training

Train is where everything you have added becomes something your twin can actually answer from. Adding a source doesn't change what your visitors get. A train does.

The Train stage tells you how many changes are waiting, lets you choose what the next one includes, and starts it. The sidebar carries the same count, so you find out there is something waiting without going looking for it.

With no material there is nothing to train, and the stage says so: it shows you four ways to add your first material, on that page, instead of a train that would build an empty index.

The three facts#

Everything below is detail on three things. If you read nothing else on this page, read these.

A train runs when you ask for one. What is waiting stays waiting until you press Train. If you would rather not come back for it, the Train overnight switch on the Train stage hands that job to the night — it is off unless you turn it on, on every agent — see Training overnight, if you switch it on for the hour and the two cases where no time is printed.

A finished train is live. Your agent answers from the new material the moment training ends. The Sandbox is where you check it afterwards, not something the train waits on — see Checking happens after it goes live.

It is counted in words and sources. That is what your plan holds, and it is what the panel quotes back to you before a train starts — see What a train counts against.

Changes wait for a train#

Adding, editing or removing a source no longer rebuilds your twin on the spot. The change is stored and counted straight away, and it joins what your twin knows at the next train.

Two reasons, and the first is the one that matters:

• Nothing changes what your visitors get without you. A twin you have shared keeps answering exactly as it did until you press Train — or, if you have switched overnight training on, until tonight's train runs.
• A rebuild is all-or-nothing anyway. Every train rewrites your whole index, so a day's worth of changes cost the same as one. Waiting is what makes them share the ride.

While a change is waiting, its row in Sources carries a Waiting chip, and asking your twin about it can still get an honest I don't know. That is the change waiting, not your twin failing.

The What it knows panel says which of those you are in. There is no knowledge base to read until a train has built one, and the panel now says why it is empty instead of assuming you have added nothing. With a source in and untrained it reads your sources are in — the knowledge base builds on the next train, and puts Train beside it. It says add a source first only when there is genuinely nothing there. If you have held everything back it says that instead, and offers no train — adding another source changes nothing while the hold stands, and a train would rebuild the same empty index.

One thing doesn't wait, because waiting would mean knowing nothing at all. Your twin's very first material is built immediately — a twin with no index answers nothing, and "your twin knows nothing until tonight" isn't a product working slowly.

A bulk import is not that thing, and it used to be. It trained once at the end, by itself, on the reasoning that "import this archive" is already a batch. Importing four hundred pages from a site turned out to be a different act from adding one: a lot of what a site holds is not what you want your agent answering from, and you cannot tell which until you can see it listed. So an import now reads everything, stores it, counts it — and stops. You choose what goes in, and press Train.

The import says so when it finishes — they are in your sources, and out of the index until a train folds them in — and Train is on the agent's Overview, on the panel that followed the import in. The confirmation used to promise a train "within a few minutes"; that sentence outlived the train it described, so the wait it asked for was for something nobody had started.

Nothing is lost if you walk away from that choice: what was imported is waiting like any other change, and the next train you run picks it up — tonight's, if you have overnight training switched on.

Train on these opens a panel over the page, and it is a panel rather than a page because nothing has been spent yet. It says how many sources are waiting and how many words that is — and, when your plan can be read at that moment, how many trains your month has left. When it can't be read, that row is left out and the panel tells you so, rather than printing a count it does not have. Train now starts it. Choose what to train takes you to the tick list instead, and Back backs out; so do Escape, the ✕ in the corner, and a tap anywhere outside the panel. While it is open the page behind it stays where you left it rather than scrolling under your finger, and backing out puts the keyboard straight back on the button you opened it from.

Training overnight, if you switch it on#

Train overnight is a switch on the Train stage, and it starts off. Turn it on and whatever is waiting is gathered into one rebuild during the night where you are, done by morning, without you coming back to this page. Leave it off and what is waiting stays waiting until you press Train.

It is off on every agent, including the ones that already existed. If your agent used to rebuild itself overnight, it no longer does until you switch this on. The Train stage says so where it matters — on an agent with changes waiting and the switch off, it tells you that what is waiting stays waiting until you train it or switch this on.

With it on, it runs at 3am, on your own clock. hiy takes your timezone from your browser when you open the dashboard, and follows you if you move — there is nothing to set and no schedule to pick. When a train is due, the Train stage prints the next one: "Next automatic train: tonight at 3:00 Europe/London". Until your browser has told us a zone, the same line reads "tonight at 3:00 UTC, until your browser tells us your zone" — that fixed UTC hour is genuinely what the nightly sweep would use, and it becomes your own night from the next time you open the dashboard.

With it on, it prints a time only when one is due, which is not quite the same as "whenever something is queued":

• Nothing queued, nothing printed — there is nothing to gather.
• Queued additions on a month whose trains are used up print no time either, and the panel says why instead — Learning is paused, with the date it resumes. Everything stays queued and the first train of next month picks it all up.
• A deletion is never paused. Taking material out does not wait for the month, so a queued removal is trained tonight and prints its time even when learning is paused.

Train is there, on any twin with material, whether or not you have switched the overnight one on — and with it off, it is what starts a train. It starts straight away and takes minutes rather than hours, and it is the right button whenever the difference matters now: you have just removed something, or you want to see the effect of a change before you tell anyone about it.

It is the same button whatever is waiting, and it always shows you the count first. It used to rename itself — Choose what to train when there was a list, Train now when there was not — so the control in the same place did two different things depending on a state you could not see. Now the press is always Train and always lands on the quote below; Choose what to train sits beside it, quietly, when there is a list to open.

Checking happens after it goes live#

This is the part worth reading twice.

The moment a train finishes, the new index is the one your twin answers from — for you in test chat and for every visitor, at the same time. Test is a check you run afterwards. It is not a gate in front, and nothing waits for your approval.

So the honest consequences:

• There is no draft index. A train that makes something worse is fixed by changing your material and training again, not by rolling back — the swap replaces the previous index rather than sitting beside it.
• For a twin you have already shared, treat a train as publishing. If you want to see a change before anyone else does, the moment to look is straight after the train, not before it.
• Your saved questions re-run themselves. Because a finished train is already answering visitors, hiy re-runs your whole saved set against the new index the moment training ends — automatically, on the server, whether or not this page is open. When it is done the Train stage says what it found: "2 of 12 saved questions no longer match what you expected." You are not charged for it: it is not a message off your allowance and not a train off your month.
• What "no longer match" means, exactly. Each saved question records one of two behaviours you expect — answered from your sources, or honestly declined — and the re-run records which one actually happened. A row that no longer matches did the other thing. It is not a comparison of wording: hiy never grades how an answer sounded.

A train that fails is the one case where nothing changes at all: the previous index keeps answering the whole time, because the swap only happens at the end of a build that worked. Sources explains what you see when one runs out of retries.

Choosing what a train includes#

Choose what to train opens the tick list: everything waiting, ticked, with its word count. Untick anything you would rather your twin did not answer from.

• Unticking holds it back, and holding back is sticky. A held source stays in your list, reads Held back, and stays out of every train until you tick it back on. It is not waiting for anything.
• Held material costs you nothing. It doesn't count towards the words and sources your plan holds, which is what makes unticking a real way back under a limit rather than a gesture.
• Ticking one back on queues it like any other change, so the next train picks it up.
• A source you deleted is listed but not tickable. It is on its way out and there is no keep-it option here: unticking a removal would hold it back, and a held source is left out of the build too, so it leaves the index either way. It is shown because you should see it going.

Ticking chooses what your index contains, not how much work runs. Every train rebuilds the whole thing — ticking six sources does not mean six sources' worth of work, and the panel says so before you start one. That is also why a train with nothing waiting is still offered: it is a rebuild of everything, and sometimes that is exactly what you want.

What the panel tells you before you start#

See what this changes replaces the tick list with a quote, in this order:

• what this train is adding — sources and words, including anything coming back from held
• what your twin will hold afterwards — sources, words, and roughly how many chunks that is
• that this rebuilds all of them, per the line above
• how many sources are held back, and which train of your month this is

Every figure is counted from material you have already added — your word counts are stored, and the rest is arithmetic over them. Words, sources and trains are the units on every training screen.

While a train is running#

The Train stage shows which phase the rebuild is in — reading through your material, working through it, then switching your twin over — and it says so in words as well as marks, so it reads the same whether you are looking at it or listening to it.

It does not show a percentage, and that is deliberate. A rebuild reports where it has got to in stages, not as a measurement: the point where your material has been read is the same point whether the next step takes four seconds or four minutes. A number there would look precise and be wrong, so you get the stage it is actually in instead.

Three things follow from training being a queued job rather than something this page runs:

• You can leave. Closing the tab does not stop a train. It finishes on its own, and the new material is live when it does. The running train says so itself now — it carries on without this page — you can close it — while it is working, rather than only surfacing after a long wait. A train is measured in tens of minutes, so the one sentence that frees you from the tab used to arrive long after the moment you needed it, or never, because you had already sat through the whole thing.
• A train the night started is followable too. Open the Train stage while one is running — whoever started it — and you see the same phases.
• A train that fails changes nothing. Your twin keeps answering from the material it already had, the stage says what went wrong, and the queue tries again. The Sources guide covers what a failed rebuild looks like on the rows themselves.
• Trying again goes through the same count. Try the rebuild again on that notice opens the quote — what is being added, what your twin will hold, what your month has left — rather than starting a train from under the notice. A retry spends one of your month's trains exactly as any other does, so it is asked for in the same words. On the Knowledge screen, which has no quote of its own, the same words take you to the Train stage, which has one. On a twin with nothing added the notice reports the failure and offers no retry, because there is nothing left to train on — adding something is the press that stage offers instead.

A new agent lands on this stage, straight from the last screen of the create flow, on the Train tab. Arriving starts nothing: what was imported is read and stored and waiting, and the train is still yours to confirm with the count in front of you.

What it learned#

The Train stage has two tabs. Train is everything above — what is waiting, what a train will include, and the press. What it learned is the read-only half: what the last train produced.

It answers two things and nothing else.

• What it is built from — how many sources the live index is drawing on and how many words those come to, and when it was last trained. Anything you have added since is deliberately not in that figure: it is waiting for the next train, and this tab is about the one that has already run.
• What it covers — the section headings of the document that train wrote, in the order they stand in it. Up to twelve, then a line saying how many more there are.

It is the outline, not the document. The whole thing already has a home: What it knows, on that agent's own Knowledge page, renders it in full. This tab is the shape of it — enough to see at a glance whether a train covered what you expected, without reading the index end to end to find out.

No model writes any of it. Every line is read off material hiy already holds: the sources and words are the same counts the rest of these screens quote, and the headings are lifted straight out of the document. So there is no second summary that can quietly drift away from the thing it claims to describe, and nothing on this tab is something new to take on trust.

Before a first train there is nothing to describe, and the tab says so plainly rather than showing you a row of zeroes — nothing has been trained yet, and where the answer will appear once one has run. A document that carries no headings gets the same honesty: it reads as one continuous piece, so there are no sections to list.

What a train counts against#

Two separate limits, and they fail in different ways.

When a month's trains are used up:

• Learning pauses, and hiy says so rather than going quiet. The Train stage states the pause and the date it lifts.
• Nothing is lost. Changes keep queueing, and the first train of next month picks up everything that accumulated.
• Removing is never paused. A source you deleted is still being answered from until a train takes it out, so that train is never refused — by the button, or by the overnight one if you have it on, whatever your month looks like. Taking your own material down is not something you can run out of.

Where to look#