Teaching Veezoo (Ground Truths)

Sometimes an answer is technically valid but not what you meant: Veezoo picked nationality when you wanted country of residence, or built a KPI differently than your team defines it. When that happens, you can teach Veezoo the correct query - so the next time anyone asks a similar question, it gets answered your way.

What is a Ground Truth?

A Ground Truth is a pair of:

• a Question, phrased the way your users would ask it, and
• the Correct VQL Query that answers it.

Veezoo treats ground truths as authoritative examples. When a new question comes in, the AI looks at ground truths for similar questions and follows the query patterns they demonstrate. A handful of well-chosen ground truths can noticeably improve accuracy on the question types that matter to your business.

Teaching from an answer

Creators can save a ground truth directly from any answer:

1. Ask the question and look at the answer's query. If the query isn't what you wanted, adjust the question or the query until you get the correct result - or just proceed to fix the VQL by hand in the next step.
2. Click the Teach button below the answer. The Teach Veezoo dialog opens, prefilled with the answer's question and its VQL query.
3. Edit the Question and the Correct VQL Query as needed, then save.

That's it - the ground truth is stored in your Knowledge Graph and immediately available to the AI.

Managing ground truths in Studio

Ground truths live in the Knowledge Graph, so Creators can also create, review and edit them as VKL files in Studio - version controlled and available in development branches:

qb {
    ground_truth PreferNationalityOverCountryOfResidence {
        question.en: "Number of customers per country"
        vql:
            """
            var customer: kb.Customer
            # Prefer nationality over country of residence
            select(count(customer) by (customer.Nationality))
            """
    }
}

See the Ground Truths VKL reference for the full syntax.

Rephrase ground truths

Besides teaching full queries, you can also teach Veezoo how to interpret a phrasing: a rephrase ground truth maps a question to what it actually means - "when users ask X, they mean Y". For example, "how are we doing this quarter?" might mean "revenue and new customers in the current quarter compared to the previous quarter". These are defined in VKL alongside regular ground truths; see the VKL reference.

Tips for good ground truths

• Teach the questions people actually ask. Use realistic phrasing, not internal shorthand.
• Keep the VQL minimal. The cleaner the query, the clearer the pattern the AI learns from it.
• Prefer fixing the root cause first. If the confusion comes from an unclear name, a missing synonym or a missing measure, fix that - and use ground truths for genuine ambiguity and hard query patterns. See Improving Answer Accuracy.