Our evaluation agent draws charts, but never writes the numbers

Description and use

Rhesis’s Architect now generates charts without reproducing raw numbers, providing the model with a query and a concise summary to avoid errors and misleading visualizations.

Problem

LLMs can choose appropriate chart types but struggle to accurately reproduce large numeric tables, leading to misleading or incorrect chart outputs.

Outcome

The new design separates chart generation from data provision, enabling accurate, non‑misleading visualizations while keeping the model focused on question formulation and narrative.

Workflow

• 1. Agent calls show_chart with a query JSON specifying entity, group_by, measures, order_by, limit, min_count, chart_type, title, x, y, target, and focus.
• 2. The tool validates the request against allowed dimensions and measures, rejecting invalid calls with guidance.
• 3. The tool runs the query through the insights API using the user’s credentials.
• 4. The tool builds a chart spec from up to 200 returned rows and attaches it to the reply.
• 5. The tool generates a short text summary of labels and values for the model to reference.
• 6. The frontend renders the chart spec with @mui/x-charts and provides toggles for exact numbers and export options.
• 7. The agent’s final reply includes the chart and a narrative that references the summary, avoiding tables of raw numbers.

Evidence

The model decides what to chart. It never supplies the numbers.

Source

The tool then: 1. Checks the request before running anything. 2. Runs the query through the same insights API behind the dashboard, using the user's own credentials. 3. Builds the chart spec from the rows that come back, capped at 200 rows, and attaches it to the reply. 4. Gives the model a short text summary.

Source

The frontend renders the spec with @mui/x-charts: line, area, bar, horizontal bar, stacked bar, scatter, radar, a shaded table for two-way breakdowns, and a single headline number.

Source