Whet pixel markWhet

Whet user guide, chapter 9, 2026-10-09

The assistant

The assistant answers questions about your project and can propose changes to the part. It runs on a model you choose. By default that is a model on your own computer. Whet checks what the model says against your records and shows you what held and what did not.

The assistant does not draw numbers out of the air. Every figure in a result comes from the solver, the geometry kernel or a fixed rule. A model may quote a figure. It may not make one up and have it accepted.

Using the assistant needs a licence. The trial can look at the page but cannot ask. See Licence, trial and updates.

Open the page

Choose Window > Workspace > Assist. The page has three columns.

  • Route on the left. It says which model will answer and why.
  • Asks in the middle. Each question and its answer is one block. The box where you type sits under it.
  • Ground on the right. It lists what the next question will stand on. Acts sits under Ground.

With no project open, Asks and Ground read "nothing open". The status bar reads "AI connected" once a model answers. You can still ask a general question if a model is set up.

Set up a local model

A local model runs on your computer. Nothing you type leaves the machine. It costs nothing per question.

Whet works with Ollama, a free program that runs models. Whet does not install Ollama for you.

  1. Install Ollama from ollama.com/download and open it once. That starts its server.
  2. In Whet, open the Local Model step. It appears on the first-run page, as a strip above the ask box on Assist while no model is set up, and as a card on Settings > Whet > Models.
  3. Press Check Again. The step shows your memory, your free disk and whether Ollama answers.
  4. Press Set Up on the model your computer can hold.
  5. Watch the bar. It shows bytes, percent, speed and time left. Stop pauses it. Set Up again carries on from the bytes already on disk.

When it finishes, the step shows a line saying the model is registered. Ask is now live.

The four sizes

The window names four local sizes: S, M, L and Ultra.

SizeWhat it isMemoryDiskStatus
Ssmall checker model--retired, not used
Mmedium checker model--retired, not used
Lan 8 billion parameter model (qwen3:8b)10.0 GB more than Whet5.58 GBthe one to set up first
Ultraa 32 billion parameter model28.0 GB more than Whetabout 20 GBfor machines with a lot of memory

S and M are not offered, because they did not beat the plain rules at checking claims. L is the size that works today. Ultra has more room and is slower. On the one comparison made it was not shown to be better. The memory figures were measured on a Linux graphics card, so on a Mac treat them as a guide. 24 GB of memory in total holds Whet and L; Ultra needs 48 GB. If the memory does not hold a size, its Set Up button is greyed with the reason. If the disk is too small, the button stays live and Whet refuses in words when you press it.

The L download is about 5.6 GB. One download of it took about 100 seconds on a fast connection.

Use an API model

You can add a model from a cloud provider with your own key. Open Settings > Whet > Models. It lists providers, keys, pins, plans and, for each job, which model would be picked and why.

  • You pay the provider. Whet sends your question and the ground it needs to that provider, and nothing on its own. Keys stay on your machine.
  • Prices come from a dated list inside Whet. Where a price is unknown, the page says so. Money is shown in US dollars.
  • A ceiling limits metered spend per month. The call that crosses the line is not stopped, because the check runs before the call.
  • A plan is a subscription you already pay for. Whet only spends a plan whose own terms allow Whet. Other plans are shown and never spent.

Use a stronger computer of yours

If you have a second computer with more memory or a graphics card, the L and Ultra models can run there. Whet reaches it over ssh with your own key. Nothing is installed on that machine by Whet. Ollama must already run on it.

  • The setting takes the form ssh://user@host. Whet checks the form when you enter it.
  • Before it sends a question, Whet asks that machine for its Ollama version and the model's digest. If they do not match what you pinned, nothing is sent.
  • If the machine cannot be reached, you get the reason and nothing is sent. A failure on the chosen machine never sends your question to another model.
  • The answer names the machine and the model that ran it.
  • Whet does not measure the other machine's memory or disk. It cannot tell you whether the model will fit there.

Studies and renders can also go to that machine. See Your own server.

Ask a question

  1. Type the question in the ask box.
  2. Pick a model in the Model list, or leave it on Routed. Routed means Whet picks.
  3. Press Ask.
  4. While it runs, the block shows "asking" and the seconds waited. Ask is greyed until the answer comes. Stop ends the question and keeps the text so far.

An answer block shows:

  • your question and the ground it used, as chips;
  • the time, the model that answered, its tags and where it ran;
  • a state, checked in green when every claim held, and in amber "N claims not found in the records" when some were caught;
  • a cost row, a ground row and a check row.

If the model refuses or cannot be reached, the block carries the reason in the engine's own words.

Ground

Ground is what the model gets to read. It is made of:

  • your study and review records, with the count of findings in each;
  • files you attach;
  • a region you took from the viewport, described below.

Press Attach to add a file, or type a link into the question field. Whet keeps a copy in your project, so the answer still has its ground if the original moves.

Limits on ground:

  • A model is handed the first 6000 characters of one source. About three dense pages of a PDF fit. If a source is longer, the block says how much was cut. The checker still reads the whole file.
  • A picture is not read by a text model. The model gets the facts measured from the file, such as its size, and the block says it did not see the picture.
  • A file Whet cannot read, such as a locked PDF or a scan with no text, stops the question before anything is sent.

Each Ground row has a menu with Open, Drop and Ask About This. Drop removes the kept copy.

Take a region with Window Select

Turn on Window Select in the SELECT panel, then drag a rectangle over the viewport. The faces that lie wholly inside the rectangle and are visible are taken. Whet saves their names, areas and positions, plus a picture of the rectangle, and adds them to Ground. The next question gets both. Window Select needs a part open on the live kernel. Without one the button is greyed and says why.

Private

Tick Private to keep the question on machines you own. A cloud model is then not eligible for it.

Limits for one question

The Limits boxes set a spending limit in US dollars and a time limit in minutes for one question. A question may run up to 900 seconds.

Route

The Route column answers "who will answer this, and why". It shows the model and where it runs, the reason it was picked, Pin and Unpin to fix the model, Why (a table of every candidate with its price per million tokens), the budget (spent today, the cap and what remains) and the model's track record on your machine. Local models show as free. The Cost row shows the list price. The answer's own cost row says what the question actually spent.

Proposed steps

Some questions ask for a change to the part, such as "Shell it 2 mm and leave the top open" or "Make the hole 10 mm". For these the assistant proposes steps. It does not change anything by itself.

Each step appears under the answer as one line:

  • the step in plain words;
  • its state: proposed, applying, applied, not applied or refused;
  • an Apply button while it is proposed.

When more than one step is waiting, Apply All appears. It runs the steps in order and stops at the first one the part refuses.

To take a step back, use Undo. Each applied step is one undo entry.

After you apply, the block says "applied by you, measured by the kernel". The numbers in it, such as the new volume and the count of faces, are measured by the geometry kernel after the change. The checker then checks the written answer against those measurements.

What a step can do

Steps are limited to a fixed set of operations. Each is checked against your part before it is shown:

  • shell, fillet, chamfer;
  • resize a hole;
  • push or pull a face (move or offset);
  • pad and pocket from a circle or rectangle on a face;
  • add a box, cylinder, sphere, cone or torus;
  • set a field of the part, or the material;
  • add a Fixed Support, Force, Pressure, Temperature, Heat Flow or Convection, and solve the study.

A step that would not work is refused in words. Two examples of the wording: "a 50 mm circle does not fit on Face5 (top), which is 60 by 40 mm, so nothing was proposed" and "the fillet names Edge99, which is not an edge of this part (it has Edge1 to Edge20), so nothing was proposed."

A step that needs your choice

If a name fits more than one face, or a Force on a round face has no direction, the assistant does not guess. It shows a question that names the options. Answer it and ask again.

Limits of steps

  • Material names are matched to the library cards. If no card matches clearly, you get a question listing the cards.
  • Steps in one answer were checked against the part as it was when you asked. A later step that names a face an earlier step removed is refused, and the run stops there.
  • Boundary conditions are refused after a shape step in the same answer, because their faces were named on the old shape.
  • A strength run takes one Force or Pressure. A heat run takes one Heat Flow.
  • Whether every kind of step can be applied from the panel in every build is (to be checked).

What the checker checks

After an answer, Whet splits it into claims and compares each with your records and the files in Ground. The result is shown on the block as claims held and claims caught.

What it does:

  • It finds a number or statement in the answer that does not appear in the ground, and marks it.
  • It compares figures to the records, so a stress value that differs from the study record is caught.

What it does not do:

  • It does not prove that the geometry is right or that the physics is right. A study can be set up wrongly and still produce records that match the answer.
  • It does not catch everything. In our own tests with deliberately false claims, the plain rules missed more than half of them. No local model was shown to do better, so models do not raise flags.
  • It can flag a true claim. A caught claim means "not found in the ground". It does not mean "false".
  • It does not read pictures.

Treat a green mark as "every claim was found in your records". Treat an amber mark as "look at this one".

What a model may not do

  • A model cannot approve a plan. Approving is your press.
  • A model cannot pick between ranked options for you.
  • A model does not write numbers into a record.
  • A model does not run any command that writes a part on its own. It proposes. You apply.
  • A model can press only what you can press, in view, and the window refuses where it would refuse you.

Plans

Press Plan to have the assistant write a goal as rows of steps. The plan appears as a block with its rows, how many ask a question and how many were refused. A question only you can answer appears as a card beside the plan, and a row that depends on it waits. Answer the card with Set. Setting values needs an approved plan, and approving is yours.

Acts

The Acts column lists the last 50 things a model did in the window: the time, who, the action, where and the result. A refused act shows the reason. Open Log opens Edit > What this window did, filtered to the model.

Speed and cost

Be ready to wait.

  • A local 8B model on a second machine answered questions in the window in 52 to 180 seconds each. The first one is slowest, because the model must load into memory.
  • A local model on a computer without a graphics card runs on the processor. The speed is not measured yet.
  • Ollama may cut a long prompt to its own context size without saying so (to be checked in your version). If an answer ignores part of your ground, shorten the ground.
  • A local model costs nothing per question. An API model costs what its provider charges. Check the Route column before you ask.

Limits

  • The assistant is a helper, not an engineer. Read what it proposes before you apply it.
  • An 8B model can list every candidate when you wanted one answer, and it can do nothing when asked to act. Rephrase and try again.
  • The checker can flag a good answer.
  • A picture attached to a text model is not seen.
  • The assistant never changes the part until you press Apply.