Run an AI skill

DeltaVault’s Ask AI panel runs published AI skills against your catalog: for example, recommending business domains and models from your connected sources. You’ve filled in a skill’s inputs and want to know what happens next, and what the skill will read.

One panel, many doors

Ask AI lives in the top navigation bar and is the same panel everywhere. Some screens add a shortcut that opens it with their context already loaded: the Ask AI button on a table’s detail header, the bulk-action bar in the knowledge base, the first-run buttons on the business modeling surface (the models grid or a model’s detail page), and the right-click Ask AI items on canvas nodes. Whichever door you use, you land in the same panel with the same skills.

Start a run

Fill in the skill’s inputs and click Run. The button switches to Starting… for a moment, then the panel flips to the live run view: you see the model’s progress stream in as it works (thinking, milestones, and the result as it forms). You’re never left staring at a frozen button.

How many results?

Some skills propose a variable number of items: business domains, entities, relationships, generated model elements, glossary terms. On those skills, the input form includes a dropdown labeled How many results? that scales how much the skill proposes in a single run. Skills with a fixed output don’t show it.

You have three levels:

• Minimal: only the essentials.
• Balanced (the default): a coherent, sensible set.
• Comprehensive: as many relevant items as your context genuinely supports.

Comprehensive never fabricates. Every proposed item must be grounded in the metadata, selection, or knowledge you provided; the model is told to prefer omission over invention. Asking for more results widens what it surfaces from your context. It never invents items to fill a quota.

This is a per-run choice that defaults to Balanced. Leave it untouched and the skill behaves exactly as before.

How hard should it think?

Most skills also show an Effort picker with four settings, ordered from lightest to heaviest: Quick, Standard, Deep, and Expert. Effort controls how much reasoning the skill puts into the run and which model runs it, so Quick is fast and cheap, Expert reasons hardest and costs most, and Standard and Deep sit between them. It is a different dial from How many results?, which controls how many items the skill proposes; Effort controls how hard it works on each one.

The picker is pre-set to the skill’s own default, which the person who published the skill chose to fit the work it does. You do not have to touch it. When you pick a different setting it applies to that one run only: it does not change your saved default, and the next run starts from the skill’s default again.

Your saved default (the setting a skill starts from when it has no published default of its own) lives in your AI settings, not here. To change the sticky default for every run, see Configure AI settings.

You do not pick a model here. Effort is the only per-run dial: the model follows from the level you choose and from whatever your organization has mapped that level to. See Configure AI settings for that mapping, and AI credits for what a credit is and how the cost estimate is built.

When a run fails

Every failure is visible on the run itself. If the run cannot start or stops partway (an exhausted AI budget, a credential problem, a skill configuration error), the run shows Run failed with the reason, right where the progress was streaming. Adjust and run again; nothing fails silently.

When the model returns no proposal

Skills that propose catalog changes normally end with a preview of actions to apply. If the model answers in prose without an actionable proposal, the panel says so explicitly and shows the model’s response text, so you can see what came back and try again, rather than staring at an empty summary.

Scope the reference knowledge

Skills that accept Reference knowledge show a searchable dropdown of your knowledge documents, grouped by where they live:

• Organization: documents uploaded for the whole organization. Always offered.
• Project: <name>, the current project’s documents. Offered when the panel is scoped to a project.

Documents from other projects are never offered, and only documents with status Ready that are not archived appear. Pick any combination; the closed dropdown shows a count of what you picked (“3 selected”).

Selected documents replace automatic knowledge retrieval for that run: the skill reads exactly what you picked.

How AI context compounds

What the field is. Rows across the catalog (projects, tables, columns, business domains, models, entities, and more) carry an AI context field: a short, operator-curated note (up to 600 characters) about what the row means, written for the model rather than for people. On projects, business domains, and business models you can also write or edit the stored value yourself at any time (see AI-assisted fields); for other rows, the accept step is where you shape it.

Capture on accept. When an AI skill proposes new or enriched rows, it also proposes an AI context for each row it understood. The accept step shows that proposal per row in an editable field labeled Proposed AI context.

You stay in control. Trim it, rewrite it, or blank it before accepting. Blank means the field is simply not set on that record. Accepting can never overwrite or clear context that is already stored on a record: what you have written always wins.

Inherit on read. When a skill that uses ancestor context runs against a row, the prompt automatically receives the context of that row’s ancestors: a column inherits from its table, a table from its project and the upstream tables it derives from, an entity from its model and domain. Long notes are kept to about 300 characters each when rendered into a prompt.

Why it compounds. Accepted context feeds the next run’s understanding, so each accepted suggestion makes the next one better grounded: you stop re-explaining your catalog to the model.