Skip to content
Help Center Back to AgenticObjects

Data models

The governed vocabulary your agents answer from — built or imported in Model Studio, reviewed and operated in the Central Management Console.

Console 8 min read Updated

A data model is the governed vocabulary your agents speak: the metrics they may measure, the entities those metrics belong to, and the connection the data comes through. A model is built in Model Studio — or imported there from an .axr model file — and appears here once it is live. This page is where you see what a live model contains and run it day to day.

What this screen is for

Every number an agent produces is read through a model. That makes the model the contract between your data team and everyone who asks questions: if a metric is not in the model, no agent can measure it, and if it is, every agent measures it the same way.

The work is split cleanly. The model carries the contract — metric definitions, entities, grain. It is made in Model Studio and is read-only here; this page holds what operates it — the query mode, row limit and timeout, row-level security and the query cache — and, until the model is made live from Model Studio, the name people see in chat, the default metric, the default period and the shared context. The glossary is read-only here: terms are edited in Model Studio. Data models is workspace-wide: every project in the workspace works from the same models.

What you see

The model list

  1. Data models with a workspace-wide badge — the page title.
  2. Model files are imported in Model Studio — a card that says where model files go: a file becomes a draft in Model Studio, checked before it goes live. A company administrator also gets Open Model Studio.
  3. The table — one row per model: the display name — the name people see in chat — as a link to its detail, with the technical name under it when the two differ (the list is sorted by display name), Type (whether the model reads a relational database or a cube, as recorded in the file), Version of the old file it was imported from (— for a model made in Model Studio), Trust (Governed, or Exploratory — cannot enter a brief), and Pending candidates, a link to the Promotion queue when candidates are waiting.

The model detail

  1. Header — the model's display name, technical name, source type, .axr version and language. On the right, Row security, Glossary, Flush cache and, on cube-based models, Re-sync in Model Studio (the re-sync with the cube is done in Model Studio now), shown when your role includes them. On a model made live from Model Studio a band at the top says This model is edited in Model Studio, with Open in Model Studio. Old-format backup (AXR 1.1) downloads what the agents read as an old-format file; on a model made in Model Studio it is replaced by Export in Model Studio, which opens the model's Versions page there.
  2. Freshness and counts — how far the data is complete and when that was measured; the number of metrics, dimensions and bound agents.
  3. Overview tab — the model's Connection; a card for each agent that uses the model (its status, its own description, who it writes for, how many of this model's metrics it may use); a summary of the model: categories, trust split, entities, dimensions carrying personal data, items without a description, the last import.
  4. Semantic content tab — everything that comes from the file, read-only: Metrics, Dimensions, Data dictionary and Entities. Below the entities sit the Relationships: one card per source table, listing the tables it joins, the columns it joins on and the kind of join. Search by name, description and synonyms; metrics are grouped by category. Opening an item shows its description, value type with a format sample, good direction, time axis, synonyms, the agents that may use it and its definition (SQL or DAX). Show all details opens every item in the list at once.
  5. Model context tab — the shared text every agent bound to this model receives in its prompt, its change history, and the agents' own notes for this model. On a model made live from Model Studio the text is read-only here and changes in Model Studio, with the model's next version.
  6. Query performance tab — over a 1-, 7- or 30-day window: attempts, the median, the 95th percentile and the longest; where the time went, times per metric, failed attempts and the slowest query shapes.
  7. Settings tab — two cards. Model content: Display name (with its own Save), Default metric and Default period — on a model made live from Model Studio these are read-only here, with Edit in Model Studio. Operational settings: Query mode, Row limit and Timeout — they take effect immediately, are not versioned and stay on this page for every model. Each row says what the setting does and which default applies when it is left empty.
  8. Imports tab — Archive: old file imports: every .axr file the old import door recorded for this model — when, what (import, export, re-sync), version, state, who, the file's fingerprint and when it was archived. The door is closed, so this list is history: nothing is deleted from it and nothing new is added.

How to …

Import a model file

  1. Open Model Studio — from the card on Data models, or from the menu.
  2. Choose New…, then Import a model file (.axr). Pick the file and the Source it will read through.
  3. Check the file shows what would happen, without changing anything: where it goes, what it cannot carry and what the quality checks find.
  4. Create draft opens the model as a draft. Nothing the agents read changes until that draft goes live from its Versions and deploy page.
  5. For a new version of a model made in Model Studio, open its Versions and deploy page and choose Import as a new draft….

Write the model's shared context

On a model made live from Model Studio, write it in Model Studio instead: open the model there, select its settings and edit Model context; it goes live with the next version.

  1. On the model detail, open the Model context tab.
  2. Write how this model is read correctly in this installation — for example which metric returns empty for which period. The counter shows the 3000-character limit and a rough token count.
  3. If an agent's note for this model holds for every agent, Add to shared context appends it to the text; the note stays on the agent.
  4. Press Save context. From the next run, the text reaches the prompt of every agent bound to this model, ahead of the agent's own note. Every save stays in the history as its own row.

Choose a default metric

  1. On the model detail, open the Settings tab and go to Default metric.
  2. Pick one of the model's metrics. With Not set — concept-word questions will ask which metric, a question that names a concept rather than a metric gets a clarifying question back.
  3. Press Save.

How to read it

  • Two owners. Semantic content comes from the model and changes only when a new version of it goes live from Model Studio. The Operational settings belong to this product and are yours to change at any time. The display name, default metric, default period and Model context move to Model Studio once the model is made live from there; until then they are changed here.
  • A missing description says so. When a metric or dimension has no description in the file, the screen states it and points to adding it in Model Studio.
  • Trust on the list is about the model as a whole. A Governed model is approved; an Exploratory — cannot enter a brief model can inform a conversation but never a delivered brief.
  • Pending candidates are proposals, not metrics. They wait in the Promotion queue until someone reviews them.
  • The freshness line always says where it stands. Measured, declared and awaiting its first measurement, or not declared — each is stated in its own words.
  • Query times separate waiting from working. A long connection wait means too many queries at once; a long run at the data source means the source itself is slow. The two are fixed differently.

Good to know

  • A model made live from Model Studio is retired there too (Retire model… on its Versions page): the model in Model Studio and the model the agents read are archived together; nothing is deleted.
  • The contract is authored in one place only: Model Studio. A model file is checked there before it becomes a draft, and a draft goes live only through its deploy window.
  • A metric that leaves the model with a new version is archived, not erased; its history stays readable.
  • A model belongs to one workspace and reads through one of that workspace's own connections, so the workspace stays the boundary for its data.