What you are looking at: an AI agent answering business questions from a sales database (modelled on Microsoft's Wide World Importers sample). Without an ontology it only sees table and column names, so it picks whichever column looks right. With an ontology it uses certified business definitions and an exact join path.
Try this: keep the ontology off and press Ask again three times. Every answer is valid SQL, and every answer is different. Now switch the ontology on and ask again. Then try the ad spend question in both modes.
What you are looking at: the five steps Auto Ontology's agent takes for every question: classify it, retrieve matching terms, write SQL, check and repair it against the real data, then run it and explain. The highlighted graph shows the part of the ontology each step uses.
Try this: tick add a typo ("Pensylvania") and step through again. Watch step 4 catch that no row matches, and repair the value before the query runs.
What you are looking at: Auto Ontology generates a draft automatically. A draft is a guess. People who know the business review each definition and certify the correct one. Answers that rely on uncertified definitions are flagged.
Try this: for each draft, pick the definition you think is right. Wrong picks explain why. Watch the questions on the right turn green as their definitions are certified.
How to read this lab: the first question, its SQL and its answer ($588.80 from 3 invoice lines) are taken from the Auto Ontology quickstart on Wide World Importers. All other numbers, wrong answers and similarity scores are illustrative, chosen to show the behaviour, not measured. Auto Ontology itself is open source at github.com/NVIDIA/auto-ontology.