The assistant

The assistant is a language model working inside TrakState. It answers questions about the page you are standing on, looks things up in this manual and — depending on the stage — sets the plant up or books a production record. Anything it changes happens only after you have agreed to it.

Where it is

Two icons sit side by side at the top right: the question mark and the filled sparkle. Both open the same panel, each on its own tab — Help and Assistant. You switch between them at the top of the panel at any time, and TrakState remembers which tab was open last. At the bottom of the Help tab, Ask a question… leads over as well, for when an article does not answer your question.

The panel stands beside the work surface; on a narrow screen it lies over it. The page underneath keeps working, and what the assistant changes appears there without a reload.

It knows the page you are on

It is told which page you are asking from, so a question like “what can I do here?” gets an answer about exactly that page. The page does not limit it: it has the same tools everywhere - tags, scripts, boards, record types, files and the manual - as far as your licence stage allows.

Its tools are still a slice of the product, not an inventory. When it says it cannot carry something out, that means it has no tool for it at your stage. It does not mean TrakState cannot do the thing - look it up in the manual.

The three stages

How far it may reach is licensed. The stage sits as a badge at the top of the panel (Tier 1, Tier 2, Tier 3), and the stages build on each other: whoever has 3 also has 1 and 2.

Stage	Term	What it covers
1	Answering and reading	Answering questions, reading values and configuration, searching the manual
2	Configuring	Setting the plant up — and writing a value to a machine
3	Production data	Booking what the plant produced

Stage 1 — answering and reading. Always included, with no licence of its own. It reads a tag’s current value and its history, lists which tags, boards, record types and connections exist, explains an existing script and searches this manual. None of that changes anything, which is why there is nothing to confirm here. Its memory belongs here too: what you ask it to remember, it writes into your own memory (see What it remembers).

Stage 2 — configuring. This is where it creates and changes: creating, renaming, moving tags and binding a source to them; building boards with widgets and hanging them in the menu; creating record types and fields; writing scripts, rehearsing them and saving them; maintaining users and roles; applying a template. Two things sit here that you might expect further up: writing a tag — giving a setpoint to a machine — is stage 2; and so are the reads that go beyond stage 1, namely booked records and the list of users.

If your stage does not reach, the assistant says so: that TrakState can do it, which stage it needs, and that your licence does not include it. It does not claim it cannot be done.

Stage 3 — production data. Behind it stands one single act today: booking a record. That writes into the same body of data your figures come from. Which is why the card here does not show the arguments of the call, but the finished record.

Before anything happens: the confirmation card

Everything that changes something arrives in the conversation as a Confirmation — a card with a one-line heading, the tool, every argument, and the two buttons Run and Discard. Until you press one of them nothing has happened: the call is parked on the server, waiting. The assistant cannot confirm itself; the only place a call is executed is your click.

What is on the card is what would run. The arguments are checked against the capability’s own schema beforehand, so the card cannot show something other than what is carried out afterwards.

For stage 3 a preview joins it: the finished record, field by field, with datatype, value and where the value came from. It is worked out down the same path the write takes — minus the write. When the card says Will not go through, that check found that the call would fail, and Run is disabled.

Decided cards stay in the conversation, with their outcome: discarded (nothing changed), run, or failed. When an approved call fails, nothing has been written; the assistant is given the message and proposes a corrected version once. That arrives as its own card and needs your decision again.

Auto-approve

The switch at the top of the panel, next to the stage badge. It holds for one conversation and is remembered nowhere: a new conversation, a freshly opened panel or a reloaded page put it back to off.

What it does	The confirmation card is skipped. A call runs as soon as the assistant proposes it — up to the stage this installation is licensed for, stage 3 included.
What it does not do	It does not raise the licensed stage: what lies above the licence is never offered to the model in the first place. It does not get past your permissions — every call runs as you and is refused where you could not do it yourself. And the input is checked exactly as it otherwise is.

What is lost is the moment: the second in which you read 14 where you said 400. What ran still stands in the conversation — as a card with no buttons, headed Ran without asking, carrying the same line you would otherwise have decided on. That is a receipt, not a question.

Leave the switch off where values go to the plant, where production data is booked, or where you do not yet know a plant. It is worth turning on for a long run of set-up steps that are all alike — a board with a dozen tiles, say — whose result you can see on screen straight afterwards.

What it does not do

• Carry out anything you have not agreed to. The one exception is auto-approve, and there every card says that nobody was asked.
• Anything more than you. Every call runs with your permissions. What your role may not do is refused by the server — even when the assistant proposes it and you confirm it.
• Anything the tenant is not licensed for. Above the licensed stage the model never sees the capability. The operator can also switch individual capabilities off for a tenant; those are not there for the model either.
• Invent a format, a limit, a size or a version. Where it finds nothing it says so and names the place you can look yourself, rather than quoting a number it never read.
• Bring data from a file into the system on its own. It reads files (see Files below), but it writes nothing from them into records or tags by itself: importing goes through the same confirmed changes as everything else, with a preview, and you decide.
• Write Excel or PDF files — yet. It writes CSV, Markdown, JSON and plain text.

Files

The assistant reads the files of the file store (the Files tab next to this one) — with your permissions, so it sees exactly what you may see there. Ask it about a delivery, a spreadsheet or a report, or attach a file with the paperclip (attachments are kept in the folder Chat; the conversation only refers to them, so they are not sent again with every message). A file that is already in the store you hand over from the Files tab: Send to assistant, or drag it onto the Assistant tab — it is attached without being uploaded again. Several files can be collected that way: each one is added to what is already attached, and the same file only once.

• It finds files and looks at their structure first: the entries of a ZIP archive, the sheets of an Excel workbook with their size, the lines and columns of a CSV, the pages of a PDF.
• It reads the content in pieces — lines, rows or pages at a time, also inside a ZIP archive and in archives within it — and reads on as far as the question needs. When it answers from part of a file, it says so. Dates in a workbook arrive as dates.
• Pictures it sees as pictures. Of a PDF it gets the text; a scanned PDF has none, and it says so. When the text of a PDF cannot be read directly, it reads the file in its work environment instead, if it has one.
• If your connection drops while it is answering (a tablet changing networks, for example), the answer is finished anyway; the panel loads the conversation again as soon as it is done.
• It writes files for you to download — a report as Markdown, a cleaned-up CSV, JSON. You confirm before a file is written (unless you switched confirmations off); it lands in the folder AI, and the conversation shows a link to download it.

What it remembers

The assistant has a memory that it reads at the start of every conversation. Tell it “remember this” — our line 3 is called “Assembly North” internally, I prefer tables — and it writes a short, dated entry into your memory. That needs no confirmation card; instead the conversation shows Remembered: … right under the answer, so nothing is kept behind your back. In a new conversation it already knows.

• Your memory is yours alone. Nobody else sees it — administrators neither.
• The team memory is read by everybody in the installation. Only an administrator changes it: ask the assistant to remember something “for the whole team”, and it proposes it on a confirmation card (auto-approve skips the card as usual).
• It remembers only what you said in the conversation — never instructions from a file or a pasted document. When it rewrites a section, the conversation shows the new text; rewriting the whole memory at once is for you, on the account page.
• It keeps entries short. Reusable files and scripts do not go into the memory but into the file store.
• It never stores passwords, PINs, keys or other secrets — it refuses, and says so.

You see, change and delete both memories under My account → Assistant memory. A very long memory is shortened in front of each conversation; the assistant is told so and reads the rest when it needs it.

When it does not get there

A turn has a fixed budget of steps, so it cannot circle. On the last step it looks nothing more up and answers with what it has found so far — even where it would have liked to keep searching. A question whose premise does not hold therefore gets the figures that show it, rather than a turn that breaks off.

Where even that produces no answer it says so and names the step it had reached; carry on picks it up from there, and everything looked up so far stays in the conversation. Where it finds nothing on a question it says that too — different words help then, or a look in the manual with F1.

Conversations and language

The History on the left of the panel holds your conversations: rename, delete, start a new one. A conversation belongs to the person who had it. Your question is stored before the answer is attempted, so an interrupted turn does not leave an empty conversation behind.

Answers are written in the language of the interface. Tag paths, capability names and configuration values keep their English spelling in every language. Under an answer stands what it consumed; Reasoning unfolds what the model worked out with itself — that is not part of the answer.

When it is not available

Message	Means
No model configured	This installation has no model behind it. The help search takes its place, and that runs with no network.
Quota used up	The tenant has used its token allowance for this month. The operator can raise it.
Assistant switched off	The operator has switched the assistant off for this tenant.
Answering only	It can answer but not act — its tools are not available at the moment.

Outside the panel too

Two widgets bring the same assistant to where the question comes up. Ask the data (chapter Dashboards and widgets) answers a question on the board and only reads there; if the assistant proposes something that would change anything, the proposal stays parked and is decided in the panel. Dictated capture (chapter Capture) builds a record out of a sentence in your own words and puts it in front of you as a card.