Hallucination-Free Answers: VQL & Verification

"Why did revenue go up?" is a question no single chart can answer. This lecture shows how Veezoo plans and answers such open questions, and why the results are hallucination-free. Even a model that writes correct SQL 99 percent of the time is too risky at the scale of thousands of users, and the VQL layer removes that risk.

The lecture also shows the verification workflow in the interface. The bold parts of the title and filter section are clickable, which opens a traditional edit interface where filters can be adjusted.

Key points

• For open questions, Veezoo builds a plan, runs the sub-questions itself, and summarizes the findings.
• Such a plan can check order counts, average order value, customer counts, and revenue per brand.
• Every answer block carries a title and filter section that states exactly what was computed.
• That section is generated from the VQL and not written by the AI, so it always matches the numbers shown.
• The AI only writes the VQL query, a language built for the AI and aimed at the Knowledge Graph.
• The translation from VQL to SQL is deterministic, so no filter can silently appear or disappear.
• Business logic and row-level security rules are applied in that translation step, so they hold every time.

Check your understanding