Using Casabase Cube in Snowflake Intelligence

After the Casabase Cube agent has been configured and access has been granted, users can interact with Casabase Cube through Snowflake Intelligence using natural language.

Instead of constructing SQL or a JSON Point of View (POV), users can ask business questions directly.

For example:

Show Actual revenue by region for FY26.

The agent can determine the required cube context, resolve dimensions and members, construct the appropriate Casabase Cube query, and return the governed result.

The basic interaction is:

Business Question
       │
       ▼
Snowflake Intelligence
       │
       ▼
Casabase Cube Agent
       │
       ▼
Casabase Cube
       │
       ▼
Governed Result

The agent provides the conversational interface. Casabase Cube remains responsible for evaluating the multidimensional result.

Before You Begin

Before using Casabase Cube in Snowflake Intelligence, verify that:

• The Casabase Cube agent has been configured.
• You have access to the agent.
• Your Snowflake default role has the required Cortex Agent access.
• You have a default warehouse and your default role has USAGE on it.
• You have the appropriate Casabase Cube application role.
• At least one cube is available to you.

For setup details, see Setting Up the Agent.

Open the Casabase Cube Agent

Open Snowflake Intelligence and select the Casabase Cube agent made available by your Snowflake administrator.

The exact navigation can change as Snowflake evolves the Snowflake Intelligence experience, but the agent should appear among the agents available to your Snowflake role.

Once selected, you can begin asking questions in natural language.

Start with Discovery

If you are unfamiliar with the available Casabase Cube models, begin by asking the agent what is available.

For example:

What Casabase Cube cubes can I access?

Then explore a cube:

What dimensions are in the Finance cube?

You can continue with questions such as:

What members are under North America?

What scenarios are available?

What years are in the cube?

Does the Account dimension contain Gross Margin?

Discovery is useful because the agent can learn the actual multidimensional structure before constructing an analytical query.

Ask a Business Question

Once the cube context is known, ask the question in business terms.

For example:

Show Actual revenue by region for FY26.

You do not need to provide:

• SQL
• Stored procedure syntax
• JSON POV syntax
• Fact-table names
• Join conditions

The agent translates the business request into the appropriate Casabase Cube operations.

Conceptually:

"Show Actual revenue by region for FY26"
                    │
                    ▼
              Understand Intent
                    │
                    ▼
              Resolve Cube
                    │
                    ▼
         Resolve Dimensions/Members
                    │
                    ▼
               Construct POV
                    │
                    ▼
          Execute Casabase Cube Query
                    │
                    ▼
                 Result

Be Specific When It Matters

Natural language does not mean that every question is automatically unambiguous.

Compare:

Show sales.

with:

Show Actual revenue by region for FY26.

The second request gives the agent considerably more analytical context.

Useful details can include:

• Cube name
• Account or measure
• Scenario
• Year
• Period
• Entity or geography
• Product
• Currency
• View
• Desired hierarchy level

You do not need to specify every dimension. Casabase Cube can use the cube’s configured defaults where appropriate.

However, providing important business context generally makes the requested analysis clearer.

You Can Name the Cube

If you have access to multiple cubes, explicitly naming the cube can eliminate ambiguity.

For example:

Using the Finance cube, show Actual revenue by region for FY26.

If you do not name a cube and the agent cannot determine which model applies, it may need to ask for clarification.

A good pattern is:

Using the <cube> cube, <business question>.

This is particularly useful when different cubes contain similarly named dimensions or members.

Refine the Analysis Conversationally

You do not need to restate the entire question every time.

For example:

Show Actual revenue by region for FY26.

Then:

Break North America down to the next level.

Then:

Compare it with Budget.

Then:

Show FY25 too.

Then:

Limit the result to the United States.

Cortex Agent threads can preserve conversation context across turns, allowing later questions to build on earlier ones.

Conceptually:

Turn 1
Actual Revenue by Region, FY26
          │
          ▼
Turn 2
Drill North America
          │
          ▼
Turn 3
Add Budget
          │
          ▼
Turn 4
Add FY25

This makes multidimensional analysis more conversational than repeatedly reconstructing a complete query.

Drill into a Hierarchy

Casabase Cube understands the hierarchies defined in each cube.

That makes hierarchy navigation a natural conversational workflow.

For example:

Show revenue for North America.

Then:

Break North America down to the next level.

Then:

Drill into United States.

The agent can translate these requests into appropriate hierarchy selections.

You can also ask more explicit questions:

Show the children of North America.

Show all descendants of Total Product.

Show only the leaf entities under Europe.

The agent uses the hierarchy defined in Casabase Cube rather than inferring organizational relationships from member names.

Change One Dimension at a Time

A useful conversational pattern is to begin with a known result and modify one part of the POV.

For example:

Show Actual operating income for FY26.

Then:

Use Budget instead of Actual.

Then:

Go back to Actual and use FY25.

Then:

Show it by region.

This is similar to interactively changing a POV in a traditional multidimensional analysis tool, but the selections are expressed conversationally.

Compare Members

You can ask for multiple members of the same dimension.

For example:

Compare Actual and Budget revenue by region.

Show FY25 and FY26 operating income.

Compare January, February, and March.

Show Gross Margin % for North America and Europe.

The agent can translate these comparisons into the corresponding multidimensional selections.

Ask for Calculated Members

Calculated members defined in Casabase Cube can be queried through the agent.

For example:

Show Gross Margin % by region for FY26.

or:

Compare Operating Margin for Actual and Budget.

If these members are defined by formulas in the cube, Casabase Cube evaluates those formulas.

The agent does not need to reconstruct the calculation.

This is important because a calculated result can depend on:

Formula
  +
Dependencies
  +
Solve Order
  +
Hierarchy Context
  +
Time Behavior
  +
Security

See Formula Overview.

Ask Time-Based Questions

You can express time-oriented analysis naturally.

For example:

Show revenue by month for FY26.

Compare January with February.

Show Actual operating income for Q1 and Q2.

Compare FY26 with FY25.

Where calculated members or formulas use Casabase Cube time-navigation behavior, those calculations remain governed by the cube.

Ask for a Chart

When the configured Snowflake agent includes Snowflake’s Data to Chart capability, you can ask for a visualization of query results.

For example:

Chart revenue by region.

Show Actual versus Budget as a bar chart.

Plot monthly revenue for FY26.

The flow is:

Natural-Language Request
          │
          ▼
Casabase Cube Query
          │
          ▼
Governed Tabular Result
          │
          ▼
Snowflake Data to Chart
          │
          ▼
Visualization

Casabase Cube produces the multidimensional data. Snowflake provides the visualization capability.

Whether charting is available depends on the tools enabled for the agent.

Ask Questions About the Result

After receiving a result, you can ask the agent to interpret it.

For example:

Which region had the highest revenue?

Which regions are below Budget?

Summarize the largest variances.

What stands out in this result?

The distinction is important:

Casabase Cube
     │
     └── Produces Governed Values

Agent
     │
     └── Explains the Returned Values

If answering a follow-up requires additional cube data, the agent can execute another Casabase Cube query.

Ask About Cube Structure

You can use the same conversation to learn about the model.

For example:

What is the parent of United States?

What members are under Total Product?

Is Gross Margin a calculated member?

What dimensions are in this cube?

These requests use Casabase Cube discovery capabilities rather than querying the fact data unnecessarily.

Ask Casabase Cube Product Questions

If documentation search is configured for the agent, you can also ask how to use Casabase Cube.

For example:

How do I add a dimension?

How do I rebuild a dimension?

What does solve order do?

How does row-level security work?

What does Health Check validate?

These are product questions rather than cube-data questions.

The agent can route them to the configured documentation capability.

Conceptually:

"What was Gross Margin in FY26?"
              │
              ▼
        Cube Query


"How is Gross Margin configured?"
              │
              ▼
     Metadata / Documentation

Ask Follow-Up Product Questions

Documentation conversations can also be iterative.

For example:

How do I rebuild a dimension?

Then:

When would I need to do that?

Then:

Do I need to rebuild after loading new fact data?

The conversational context allows the agent to understand that the follow-up questions still refer to dimension rebuilding.

Security Still Applies

Using natural language does not bypass Casabase Cube security.

If row-level security restricts a user to:

North America

then a request such as:

Show revenue for all entities.

can return only the data permitted to that user.

The effective flow remains:

User
  │
  ▼
Casabase Cube Query
  │
  ▼
Apply User Security
  │
  ▼
Aggregate / Calculate
  │
  ▼
Permitted Result

Two users can therefore ask the same question and receive different valid results based on their Casabase Cube security configuration.

See Governance and Security.

The Agent Uses Your Snowflake Identity

Cortex Agent access is evaluated using the querying user’s Snowflake default role.

This matters when troubleshooting a situation where:

I can access the cube in one interface,
but the agent says I do not have access.

The current session role alone does not determine Cortex Agent permissions.

Verify:

• Your Snowflake default role
• Cortex Agent access granted to that role
• Your default warehouse
• USAGE on that warehouse
• Access to the Casabase Cube agent
• Your Casabase Cube application role

See Setting Up the Agent.

When the Agent Asks for Clarification

Some questions can have multiple valid interpretations.

For example:

Show margin.

The cube might contain:

Gross Margin
Gross Margin %
Operating Margin
Operating Margin %

Rather than choosing arbitrarily, the agent may ask which member you mean.

Similarly:

Show sales for the US.

might be ambiguous if the model contains multiple members or aliases that could match “US.”

Clarification is preferable to silently selecting the wrong analytical context.

When a Member Cannot Be Found

If the requested business term does not match a member in the selected cube, the agent should not invent one.

For example:

Show Strategic Profit Index.

If that member does not exist, the appropriate response is to identify the mismatch and, where possible, help locate related members.

You can ask:

What Account members contain "Profit"?

or:

What calculated members are available in Account?

Then refine the request using the actual cube metadata.

When a Query Is Too Broad

Natural language can make it easy to accidentally request a very large analytical result.

For example:

Show every account for every entity, product, customer, month, scenario, and year.

A better approach is to request the grain required for the analysis.

For example:

Show Actual revenue by region for FY26.

Then drill into the relevant region or product branch.

This reduces unnecessary data retrieval and gives the agent a clearer analytical objective.

When Casabase Cube Refuses a Query

Some requests cannot be evaluated within supported Casabase Cube calculation semantics or scale.

When that occurs, Casabase Cube returns an explanatory error rather than an approximate result.

The agent should preserve that behavior.

Conceptually:

Request
   │
   ▼
Casabase Cube
   │
   ├── Supported ──► Result
   │
   └── Unsupported
          │
          ▼
     Explanatory Error
          │
          ▼
      Refine Request

If the agent explains that the query needs to be narrowed or changed, modify the request rather than repeatedly submitting the identical question.

For example:

Show all descendants of Total Entity for every calculated account.

might need to be narrowed to:

Show Gross Margin % by region.

or another smaller analytical grain.

Do Not Ask the Agent to Recreate the Cube

For governed analysis, avoid prompts that explicitly ask the language model to independently recreate calculations already defined in Casabase Cube.

For example, if Gross Margin % is a calculated cube member, prefer:

Show Gross Margin % by region.

rather than:

Pull revenue and cost and invent a Gross Margin % calculation yourself.

The first request uses the governed cube definition.

The second creates a parallel analytical definition that might not match the organization’s business logic.

Better Prompting Patterns

You do not need specialized prompt engineering, but a few patterns can improve analytical requests.

Identify the Measure

Instead of:

Show performance by region.

prefer:

Show Operating Income by region.

Identify Important Context

Instead of:

Show revenue.

prefer:

Show Actual revenue by region for FY26.

Ask for the Desired Grain

Instead of:

Show North America.

prefer:

Show revenue for the children of North America.

Make Comparisons Explicit

Instead of:

How are we doing?

prefer:

Compare Actual and Budget revenue by region for FY26.

Name the Cube When Ambiguous

Instead of:

Show revenue by entity.

prefer:

Using the Finance cube, show revenue by entity.

These patterns give the agent clearer intent while still allowing a natural conversational workflow.

Example Analysis Session

A complete conversation might look like this:

User:
What cubes can I access?

Agent:
[Returns available Casabase Cube models]

User:
What dimensions are in the Finance cube?

Agent:
[Returns the cube dimensions]

User:
Show Actual revenue by region for FY26.

Agent:
[Runs the Casabase Cube query and returns the result]

User:
Break North America down to the next level.

Agent:
[Changes the Entity hierarchy selection and reruns the query]

User:
Compare it with Budget.

Agent:
[Adds Budget to the Scenario selection]

User:
Show Gross Margin % instead.

Agent:
[Changes the Account selection to the calculated member]

User:
Chart the result.

Agent:
[Uses Snowflake visualization capability if configured]

Behind the conversation:

Conversation
     │
     ▼
Agent Orchestration
     │
     ├── Discovery
     ├── Query
     ├── Query Refinement
     ├── Query
     └── Visualization
             │
             ▼
      Casabase Cube remains
      the analytical authority

Tips for Effective Use

• Start with discovery if you do not know the cube structure.
• Name the cube when multiple cubes could answer the question.
• Use business terminology rather than trying to generate SQL.
• Include the important measure, scenario, time period, and desired grain.
• Refine results through follow-up questions instead of requesting everything at once.
• Use hierarchy language such as children, descendants, leaf members, or next level when useful.
• Ask for calculated members by their defined cube names.
• Ask for charts only when visualization is useful.
• Allow the agent to clarify ambiguous member names.
• Narrow a request when Casabase Cube reports that the query exceeds supported scale or semantics.
• Treat Casabase Cube results as the governed analytical values and the agent’s narrative as an interpretation of those results.

What You Do Not Need to Know

For normal conversational analysis, users generally do not need to know:

QUERY_CUBE procedure syntax
JSON POV construction
Snowflake SQL generation
Fact-table structure
Dimension-table joins
Formula implementation
Aggregation SQL

Those details remain available for programmatic and advanced use cases, but they are not required to ask a business question through Snowflake Intelligence.

That is the primary benefit of the integration:

Business Language
       │
       ▼
Governed Multidimensional Analysis

Related Topics

• Intelligence Overview
• How Casabase Cube Works with Snowflake Intelligence
• Setting Up the Agent
• What the Casabase Cube Agent Can Do
• Governance and Security
• Tips and Limitations
• Query Builder
• Point of View (POV)
• Formula Overview