Skip to content

Automatic translation — the Portuguese version is the reference. View in Portuguese

Updated on 2026-09-08

Hallucinations and validation

What an LLM hallucination is, how ingestia.io detects and refuses it, and how you can check.

On this page (8)

A hallucination is when the LLM states something confidently, but without grounding in the data — a number that does not exist, a made-up id, a total over a sample. This page explains how ingestia.io reduces and detects hallucinations and how you can check. The house principle is a single one:

The product's AI refuses and warns — it does not pretend. When there is no evidence, it says there is none.

Always review

The defenses described here are deterministic and covered by tests — but no defense eliminates 100% of the risk. Always review.

The core idea: the LLM writes, the code validates#

The point that changes everything: what decides whether the answer is grounded is pure code, not the LLM. The model writes the text; deterministic validators check that text against what actually traveled in the prompt. It is verification over the LLM output.

How each capability defends itself#

CapabilityDeterministic defense
Q&AEach widget/row gets an id ([W1], [W1.r3]); the LLM must cite these ids, and verificarCitacoes checks each citation against the index of what traveled — not against the model's text. A forged citation does not pass.
Text-to-SQLGenerates only SELECT; on the semantic path, nonexistent id → SQL discarded; ambiguity → refusal before spending tokens; a real dry-run on the datalake validates bytes/cost.
Assisted generationThe artifact goes through the same validators as the editor (measure, formula typecheck, model compilation, visual allowlist, SQL validation). Failed → self-correction once; still failing → comes out as a draft with the errors shown.
NarrativeNon-negotiable rule in the prompt: use only numbers present in the data. Without a previous period, the change is null and the narrative states "no basis for comparison" — it never makes things up.
InsightsThe "anomalies" and "recommendations" are the LLM's reading, labeled as hypotheses — separate from the deterministic statistical detector (see Anomalies).

The forms of refusal (an ungrounded answer does not come out as fact)#

  • fora_do_contexto — the question touches no term of the dashboard: refused before spending tokens.
  • sem_evidencia — the LLM answered, but no citation checks out: it becomes a refusal (the real cost is settled honestly).
  • evidenciaFraca — most citations are bad: the answer comes out with a warning, and only the citations that check out count as evidence.
  • ia_indisponivel — no provider/error: honest fallback, empty citations, never fake evidence.

Golden rule: no unproven causality#

No AI answer claims cause and effect that the data does not prove. The narrative uses only numbers that are present; Q&A only states what it can cite; insights are hypotheses; anomalies/forecasts describe statistical deviation and trend, never "X caused Y". A correlation seen on a dashboard does not become a cause.

How you check (practical validation)#

  1. Test the forged citation

    In Q&A, put a text like "[W9.r9]" in a cell and confirm that it does NOT become a valid citation.

  2. Force a refusal

    Ask a question the dashboard does not support and confirm that it comes out as a refusal, not as fact.

  3. Review the SQL

    In text-to-SQL, read the SQL and the explanation before running it; check column/table names.

  4. Check the honest draft

    In assisted generation, ask for something impossible and confirm rascunho: true with the errors listed.

  5. Check the provenance

    Every answer includes proveniencia (model, prompt version, cost in R$, opId) — use it to audit.

Warning signs (when to be suspicious)#

  • A number that is too "round" or that you cannot find on the dashboard.
  • A claim about a total over a visual flagged as a sample.
  • Causal language ("dropped because of…") — the AI should not claim that.
  • An answer without citations in Q&A.

Limits and caveats#

  • Deterministic validation covers structure, ids, citations and SQL — it does not guarantee that the business interpretation is right. The human decides.

Related#


Last reviewed: 2026-09-08.

Related links