The Model Builder builds a semantic model from one approved design. You meet it in a chat with the Warehouse Builder: it proposes the design, works through it in rounds, reports each round back as a card, and calls on you when it needs a decision. To read each card it shows, see Model Builder cards. For the whole flow around the chat, from the request to the working model, see Design and hand over a semantic model.
Start a build
Open a chat with the Warehouse Builder Mind.
Type /star-schema in the message box. The command reads "Use the Star schema skill: design and build a semantic model", and it joins your message as a Star schema pill.
Send the message. The build loads the skill, and the transcript shows a chip reading Using: Star schema v1, with the tooltip "A DataMind playbook shipped with this platform version".
Answer the questions the builder asks in the chat, then review the design it proposes.
Approve the design
The design arrives as a card headed Semantic model design. It lays out the fact and the dimensions it will build, the objects the model will read, and the choices to make: the owner role, the data mode (Tokenized (recommended) or Clear values), and the history depth. For every section of the card and how to set the choices, see The semantic model design card.
Set the choices while the card waits for your decision.
Approve with the button built from your choices, for example Approve · owner role · data mode · history option. To send the design back instead, choose Reject with a note.
Once you decide, the card locks. The build records the decision as: Design approved (session <id>): owner role <role>, data mode <mode>, history "<option>".
Follow the rounds
The builder works through the design in rounds and reports each round back into your chat as a card. Between your decisions it works on its own, so you can leave the chat and come back when the card, the strip or a notice calls on you.
A strip above the message box keeps the session in view: Model Builder · model · round text · state, with a Needs you: n badge when something waits for you. Select the strip to open the session card; on a phone it opens as a bottom sheet.
The state starts at Getting ready, becomes Working or Model AI building, round n of m, pauses as Waiting for the Model AI chat or Needs you, and ends as Done, Stopped or Failed.
Rounds run in the model's Model AI chat, one after another. The brief for each round is posted there, and each round files a report back into your chat as a card. SQL never crosses between the chats: briefs describe the work in words, not query text.
A round that cannot start because the Model AI chat is busy parks, and starts by itself when the chat is free.
Messages you send while a round runs wait for it. The composer shows "Your message is delivered right after the Model Builder's round", and attachments are refused with "Attachments cannot be queued while the Model Builder works. Send the text now and the image after its round, or wait for the round to end." If someone else already has a message queued, the refusal reads "name already queued a message for after the Model Builder's round. Wait for the round to end, then send." The queued message itself and its Remove action are described on Model AI.
Answer questions and decide
A question from the build appears in the session card's Needs you section. Pick one of the choices, or write your own answer in the field; Send answer sends it, and Enter works too. An answer can hold up to 2000 characters.
If the build wants to change a table that existed before the design, it asks first, under The Model Builder needs a change on a shared table. The options are I will make it in the Warehouse Builder chat to do it yourself, Continue without it to let the build carry on and skip that change, or Answer in your own words. The hint reads "This table existed before this design, so the Model Builder changes it only with your word."
A decision on a change the Model AI wants appears as an approval row: Approve, or Reject with an optional note. The change itself is applied or dropped in the Model AI chat.
Grant access to the objects
Before the model is created, the build checks that its owner role can read every object the model needs. When something is missing for a business role, the chat shows a card titled Grant role access.
Choose Open permissions on an object that is not granted, add the role under Permission Groups in the Warehouse popup, and save it.
Choose I've granted them, check again. It sends "I granted the access. Check again." into the chat and runs the check again. The card counts the progress as "n of m granted", and each object reads Granted, Not granted, Columns missing, Not deployed or Not found.
Repeat until everything is granted. The admin role reaches every object, so its card finishes straight away.
The build creates the model empty in the requester's personal workspace: "Created empty in the requester's personal workspace; move it later with Move."
On the created card, Send the brief to Model AI sends the build brief into the model's Model AI chat and opens the model; Open Model AI chat opens the model without sending anything. When the Model Builder sends the brief itself, the card says "The Model Builder sends the brief and builds this model with its Model AI." and the send button is hidden.
The build records the creation as: Created the empty semantic model "<name>" in <requester>'s personal workspace (owner role <role>, data mode <mode>).
Note
The model is created empty. Its contents arrive in rounds, built through the model's Model AI from the build brief. The model itself is described in Semantic models and BI tables.
Stop, continue or start again
The session card's footer holds the actions.
Stop ends the build for now. Confirm with Stop in the dialog; it reads "The Model Builder stops working on this model. Anything it made stays; you can start it again."
Continue with 2 more rounds appears when the session stopped at its round limit. It adds two rounds, once per session, and only the person who started the build can use it. Everyone else sees Only {name} can give it 2 more rounds.
Start again begins a fresh session for the same design. It is for the person who started the build, and it is available once the current session has ended.
A notice confirms each action: Model Builder stopped, Model Builder sessions reset or Answer recorded. A failed action reads Stop failed, Continue failed or Start again failed.
When the build ends
The done card reports the outcome, with its scorecard and what still wants attention. Its headline reads Your semantic model is ready, or Ready, with charts that failed, Needs your attention, Stopped or Failed when the build ended differently. Model Builder cards shows how to read it.
Open items list what still wants attention, in this order: slow charts, questions, change requests, the full history, failed tests, model checks.
When the build ends on its own, a close notice says why in plain words. The endings you may see:
All rounds used: "Used all 8 rounds. Continue with 2 more rounds, or open the Model AI chat to finish by hand."
Time limit reached: "Reached its 12-hour limit. Start a new session from this chat, or finish in the Model AI chat."
Token budget used: "Used its 30M token budget. Start a new session from this chat."
A repeated request: "The Model AI asked for the same thing three times. Decide here, then start a new session."
Rounds not finishing: "Two rounds in a row did not finish. Start a new session."
Access or model gone: "The requester can no longer use this model's Model AI." or "The model was deleted."
Model not complete: "Finished without a complete model: see the open items."
Start or run failure: "The Model Builder could not start." or "The Model Builder failed while it worked."
The full history load
A design can ask for the full history of its fact tables.
The load runs once per design; a second full load needs a new design card.
Backfills run one at a time, with eco query settings, so they are gentle on the warehouse.
The card notes when every load has finished and the Model Builder is woken. It then checks each table's health and continues; findings come back as open items.
Limits
What
Limit
Rounds per session
8, plus 2 more once with Continue with 2 more rounds (10 at most)
Session time
12 hours
Token budget
30 million tokens
Sessions per model
2, and Start again resets the count
One build works on a model at a time. While a session is open, another one is refused with "A Model Builder is still working on this model."
Who can see and act
The build runs in a chat with the Warehouse Builder Mind. Admins reach it, and so do members of its permission groups who hold View Warehouse Objects and Create/Edit Warehouse Objects. See Warehouse Builder and Lake Builder.
You read a session with the same access that lets you view the chat.
Answering the build's questions and stopping it need the permission to write in the chat. A question that only writers can answer says Someone who can write in this chat answers this question.
Deciding the changes the build asks for, Continue with 2 more rounds and Start again belong to the person who started the build. When someone else tries to decide a change, the message reads: "Only name, who started this Model Builder, can decide its changes."
Where a viewer cannot use the model's Model AI, its text appears as a stub: "Only people who can use this model's Model AI can read it."
While you view the platform as another user, a session stays readable, but you cannot act on it: "Model Builder actions are not available while you are impersonating another user." See Impersonate a user.
Questions, answered
How is the Model Builder different from Model AI?
Model AI is the assistant panel where you ask for changes on a semantic model. The Model Builder is a build that runs by itself: it works from an approved design, in rounds, and reports back into your Warehouse Builder chat. You follow and steer it from the session card without opening the model. See Model AI.
Does approving the design build the model straight away?
No. Approval starts a build that works in rounds. The model is created empty in the requester's personal workspace, and the model's Model AI builds it from the build brief. The done card tells you when the build is finished.
How many rounds does the Model Builder get?
Eight rounds per session. When it uses them all, the person who started the build can add two more with Continue with 2 more rounds, once per session. A session also ends at its 12-hour limit or when it uses its 30 million token budget.
What is the difference between Continue and Start again?
Continue with 2 more rounds adds two rounds to the same session, when it stopped at its round limit, and works once per session. Start again begins a fresh session for the same design, and is for the person who started the build once the current session has ended.
What happens to the model if I stop the build?
Nothing is rolled back. The Model Builder stops working on this model and anything it made stays, and you can start it again.
Who can decide the changes the build asks for?
The person who started the build. When someone else tries to decide, the message reads: "Only name, who started this Model Builder, can decide its changes." Others can still read the session.
Can I answer the build's questions if I did not start it?
Yes, if you can write in the chat. Questions are answered in Needs you by anyone who can write there; the hint reads Someone who can write in this chat answers this question. The extra rounds and Start again stay with the person who started the build.
The build asked me about a table that existed before the design. What do the options mean?
I will make it in the Warehouse Builder chat keeps the change in your hands: the build skips it and you do it yourself. Continue without it lets the build carry on without that change, and Answer in your own words lets you reply in text. The build changes only the tables the design created or changed, so it asks before touching anything older.
Can I run a second build on the same model?
No. One build works on a model at a time. While a session is open, another one is refused with "A Model Builder is still working on this model."
Can I follow the build from the model instead of the chat?
Yes. While the build runs, the model's Model AI panel shows a banner above the chat, for example Model Builder is working · round n of m · started by name, with Stop where you can stop it. See Model AI.
Does the full history load by itself?
It runs once per design, one table at a time, and the card shows its progress. When every load has finished, the Model Builder is woken, checks each table's health and continues. A second full load needs a new design card.
Why can't I read a round's report?
Parts of a round's output come from the model's Model AI chat. If you cannot use that model's Model AI, they appear as stubs: "Only people who can use this model's Model AI can read it." You can still follow the session's states and decisions.