AGENTIX · ENGINEERING LOGARCHITECTURE / 2026
Architecture5 min read

Grounding beats cleverness

Why structure, not a bigger model, stops confidently-wrong ERP answers: VistaGPT grounds every query in 70,641 verified columns and refuses to guess.

By Agentix Tech · Engineering log

A plain-English question passes through deterministic grounding checks before any SQL is returned, or is refused

A generic language model pointed at a twenty-year-old construction ERP will answer almost anything you ask it. That is the problem. Viewpoint Vista has thousands of tables and no forgiving schema, and a model that has not been taught its traps will write a query that reads clean, runs without error, and returns the wrong number. Nobody gets an error message. The report just quietly lies. The instinct is to reach for a larger model, as if more parameters would somehow learn two decades of a specific ERP’s tribal knowledge. They will not. The fix is not a cleverer model. It is structure placed around a competent one.

That is the premise behind VistaGPT, our in-development conversational interface to Viewpoint Vista. It is engineered to refuse rather than guess.

70,641columns every query is checked against
19,249verified schema entries grounded
1,200+proven report queries reused
13deterministic stages before any SQL runs

The failure mode is confidence, not ignorance

An ERP question has a right answer. “What did we spend on this job last quarter” resolves to specific tables, specific joins, and a specific handling of soft-deleted and un-posted rows. Get the join wrong and the figure is not slightly off — it is a different number with no warning attached. A confidently-wrong report is worse than no report, because nobody knows to check it.

The danger is not that the model knows nothing. It is that it knows enough to be plausible and not enough to be right.

A bigger model narrows that gap; it does not close it, because the knowledge it is missing — how this ERP encodes posting paths, composite keys and purges — was never in its training data. It lives in the schema and in the queries people already trust.

Structure, not scale

VistaGPT puts deterministic machinery in front of the model and verifies its output against the real database before a human ever sees it. The grounding is done by rules, not by hope.

QUESTION Plain English "spend on this job" DETERMINISTIC GROUNDING SchemaRetriever ~19 zero-LLM rules HallucinationGate checked vs 70,641 columns Validated SQL copy and run Refuse rather than guess
The model drafts; deterministic rules decide. A query that cannot be verified against the real schema is refused, not returned.

Where the correctness actually comes from

None of the layers below involve a larger model. Each is a rule, a lookup, or a reuse of something already trusted.

Layer What it does Why it beats a bigger model
Schema grounding Grounds answers in 19,249 verified entries across 3,969 objects and 6,975 stored procedures The schema is fact, not inference
Deterministic rules ~19 zero-LLM checks: NOLOCK, composite-key joins, fan-out, soft-delete, posting-path NULLs Rules do not hallucinate
Column verification Every generated field checked against all 70,641 real columns A field that does not exist cannot ship
Report reuse Reaches for 1,200+ proven production Crystal Report queries before writing anything new An answer can cite a trusted report, not an invention

Behind the default interface, the CrewGPT engine runs nine agents across five crews, with a HallucinationGate that applies five checks and an auto-fix pass. The point of all of it is the same: the model proposes, and deterministic machinery disposes.

If the field is not in the database, the firewall drops it. No amount of fluency gets a non-existent column past a check that simply asks whether the column exists.

Honest about maturity

VistaGPT is in development and under internal validation only. We are not publishing accuracy scores we have not earned the right to publish, and the four engines behind the interface are selected in configuration rather than benchmarked against each other in public. What the architecture already gives you is a system whose correctness does not depend on the model being clever on the day — it depends on checks you can read, run and audit.

If your team waits on a DBA to answer questions your ERP already holds, the lever is not a larger model — it is the structure around it. See how the grounding, validators and engines fit together on the VistaGPT case page. To talk it through for your own stack, book a call at cal.com/agentix-tech or reach us via /contact.

Work with us

Building a system
that has to hold up?

Tell us what you're building. We'll tell you how we'd architect it, what the eval harness would cover, and what production deployment involves.

AGENTIX TECH · engineering log · Architecture