Customize charts in Snowflake Intelligence¶

Overview¶

Customization works at two levels:

• Agent level: Applies to all charts across every semantic view attached to the agent. Use this for global defaults such as brand colors and fonts.

• Semantic view level: Applies only to charts generated from that specific semantic view. Use this for column-specific rules and domain-specific chart type preferences.

At each level, two mechanisms are available:

• vega_template: A partial Vega-Lite (https://vega.github.io/vega-lite/) JSON spec that is deterministically merged into every generated chart. Use this for anything that must always apply.

• Free-text instructions: Natural language guidance injected into the chart generation prompt. The LLM makes a best effort to follow these, but they aren’t guaranteed.

Agent-level customization¶

Add a <chart_customization> block inside instructions.orchestration in your agent configuration. You can combine font theming and a global default palette:

<chart_customization>
Prefer horizontal bar charts for ranked data.
vega_template:
{
  "background": "antiquewhite",
  "config": {
    "title":  { "font": "monospace", "fontStyle": "italic", "fontSize": 20, "fontWeight": "lighter" },
    "axis":   { "labelFont": "monospace", "titleFont": "monospace", "titleFontSize": 15, "labelFontSize": 10 },
    "header": { "labelFont": "monospace", "titleFont": "monospace", "labelFontSize": 10 },
    "legend": { "labelFont": "monospace", "titleFont": "monospace", "titleFontSize": 18, "labelFontSize": 15 },
    "mark":   { "font": "monospace" }
  }
}
</chart_customization>

Use the agent level for:

• Brand color palette

• Default visual theme (fonts, background)

• Cross-domain style preferences

• Number or currency formatting defaults

Avoid column-specific color mappings at the agent level. Column names differ between semantic views and are silently ignored when not found.

Semantic-view-level customization¶

Add a <chart_customization> block inside the module_custom_instructions.sql_generation field of your semantic view YAML. This field takes precedence over the legacy custom_instructions field when both are set.

CREATE OR REPLACE SEMANTIC VIEW my_db.my_schema.my_view
  FROM @my_stage/semantic_view.yaml;

# semantic_view.yaml
name: my_view
module_custom_instructions:
  sql_generation: |
    <chart_customization>
    Always use a line chart for time series data.
    vega_template:
    {
      "transform": [
        {
          "calculate": "datum.CATEGORY === 'Furniture' ? '#4e79a7' : datum.CATEGORY === 'Technology' ? '#f28e2b' : datum.CATEGORY === 'Office Supplies' ? '#e15759' : ''",
          "as": "_color"
        }
      ],
      "encoding": {
        "color": { "field": "CATEGORY", "type": "nominal", "scale": { "range": { "field": "_color" } } }
      }
    }
    </chart_customization>
tables:
  ...

Use the semantic view level for:

• Per-column color mappings

• Domain-specific chart type rules

• Metric-specific formatting

• Overriding agent-level defaults

Use with caution: templates affect every chart¶

vega_template is merged into every chart generated at that level. There’s no per-question or per-chart-type filtering. If you add an encoding.y override at the agent level, it applies to bar charts, line charts, scatter plots, and pie charts alike.

Before adding a template, consider:

• Scope: Agent-level templates affect all charts across all semantic views. Prefer the semantic view level when a rule is specific to one domain or dataset.

• Wildcard encodings: A template encoding that omits field (for example, "y": {"axis": {"format": "..."}}) applies to every chart’s y axis regardless of what column is plotted. Use field to pin it to a specific column when the semantic view is known.

• Mark overrides: Setting "mark": "line" at the agent level forces every chart to a line, including ones where the LLM would correctly choose a bar or pie. Only override mark at the semantic view level where you have domain knowledge about the data.

• Transform arrays: A calculate transform in the template (for example, _color) is injected into every chart’s transform array. If the data doesn’t contain the referenced column, Vega-Lite silently produces null values for the calculated field.

When in doubt, start at the semantic view level and promote to the agent level only after confirming the rule is safe for all charts.

To validate a template before deploying it, paste a representative chart spec (with your vega_template already merged in) into the Vega Editor (https://vega.github.io/editor). The editor shows live warnings and errors in the console. A valid template should produce no warnings. Common things to catch this way: invalid property names, type mismatches, unreachable calculate expressions, and scale configuration errors.

Fonts¶

Font settings are controlled through the config block in vega_template. All font properties are applied globally to the chart and affect every chart generated, regardless of data.

{
  "config": {
    "title":  { "font": "serif", "fontSize": 20, "fontWeight": "bold", "fontStyle": "italic" },
    "axis":   { "labelFont": "monospace", "titleFont": "monospace", "labelFontSize": 11, "titleFontSize": 13 },
    "header": { "labelFont": "serif", "titleFont": "serif", "labelFontSize": 11 },
    "legend": { "labelFont": "serif", "titleFont": "serif", "labelFontSize": 12, "titleFontSize": 13 },
    "mark":   { "font": "serif" }
  }
}

Common config font properties:

You can also set a global background color alongside fonts:

{
  "background": "#f9f9f9",
  "config": {
    "title":  { "font": "monospace", "fontStyle": "italic", "fontSize": 20, "fontWeight": "lighter" },
    "axis":   { "labelFont": "monospace", "titleFont": "monospace", "titleFontSize": 15, "labelFontSize": 10 },
    "header": { "labelFont": "monospace", "titleFont": "monospace", "labelFontSize": 10 },
    "legend": { "labelFont": "monospace", "titleFont": "monospace", "titleFontSize": 18, "labelFontSize": 15 },
    "mark":   { "font": "monospace" }
  }
}

Colors¶

<chart_customization>
Color Active status green, Inactive status red, and Pending status yellow.
</chart_customization>

{
  "transform": [
    {
      "calculate": "datum.STATUS === 'Active' ? '#22c55e' : datum.STATUS === 'Inactive' ? '#ef4444' : datum.STATUS === 'Pending' ? '#eab308' : ''",
      "as": "_color"
    }
  ],
  "encoding": {
    "color": {
      "field": "STATUS",
      "type": "nominal",
      "scale": { "range": { "field": "_color" } }
    }
  }
}

{
  "encoding": {
    "color": {
      "scale": {
        "domain": ["Furniture", "Technology", "Office Supplies"],
        "range":  ["#4e79a7", "#f28e2b", "#e15759"],
        "scheme": "tableau10"
      }
    }
  },
  "usermeta": { "merge": "extend" }
}

Disabling Snowsight styling¶

By default, Snowflake Intelligence applies Snowsight UI theme adjustments on top of the generated chart. To opt out and render the chart exactly as specified in your vega_template, set ui-merge to "none" in usermeta:

{
  "usermeta": { "ui-merge": "none" }
}

This is useful when you want full control over the visual output, for example, when applying a custom brand theme and you don’t want Snowsight to override colors, fonts, or backgrounds.

Number and currency formatting (experimental)¶

Axis and legend labels can be formatted using D3 format strings (https://d3js.org/d3-format) through vega_template. This is useful for enforcing consistent currency symbols, decimal places, or SI suffixes across all charts.

Set axis.format for quantitative axes (x, y) and legend.format for color/size legends:

{
  "encoding": {
    "y": { "axis": { "format": "$,.0f" } }
  }
}

{
  "encoding": {
    "y": { "type": "quantitative", "axis": { "format": "$,.0f" } }
  }
}

Common D3 format strings:

To apply formatting to all quantitative channels at the agent level (without knowing the specific column name):

{
  "encoding": {
    "y": { "axis": { "format": "$,.0f" } },
    "x": { "axis": { "format": "$,.0f" } },
    "color": { "legend": { "format": "$,.0f" } }
  },
  "usermeta": { "merge": "extend" }
}

Use "merge": "extend" so the format is added only to channels the LLM already populated, without overwriting their field or type settings.

Merge modes¶

Control how vega_template interacts with the LLM-generated chart by setting "usermeta": {"merge": "<mode>"} inside the template.

Rules that apply to both modes:

• The data block is never overwritten.

• Encoding overrides apply only when the template’s field matches the chart’s field, or the template omits field.

• After merging, domain entries not present in the actual data are automatically removed.

Example: force a line chart

{
  "mark": "line",
  "usermeta": { "merge": "override" }
}