Skip to content

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

Updated on 2026-10-04

How AI understands your data (grounding and context)

What the AI receives in order to answer — semantic model, PII-free catalog and snapshot — and how that keeps it faithful to your numbers.

On this page (7)

A conceptual page. All of ingestia.io's AI works better — and more safely — when it is grounded in your semantic model. Here we explain what the AI receives in order to answer, why that keeps it tied to your numbers, and how it explains a measure without making things up.

Note

The grounding logic and the egress controls are covered by automated tests. See the AI overview.

The problem grounding solves#

A "raw" LLM knows the language, but it does not know what revenue means at your company, nor that valor excludes returns. Without context, it guesses. Grounding means giving it the minimum it needs — the structure and the definitions of your model — so that it speaks your language and can be verified.

What the AI receives (and what it never receives)#

It depends on the capability, and everything goes inside a <dados> fence (treated as data, never as instructions):

CapabilityReceivesDoes not receive
Text-to-SQL, Assisted generationStructure: semantic model (ids of tables/relationships/measures) or the catalog schemaData rows — structure only
Q&A, InsightsDashboard snapshot (rows), subject to the egress policyMore than the person asking can see
NarrativeReport results (current and previous period)—

Never leaves, in any capability:

  • Columns flagged as PII — excluded at the source, before the prompt is built. The AI does not even know they exist.
  • Data from another workspace — a prefix guard discards any table outside your workspace.
  • In the capabilities that read rows, the AI sees the secure view: the member's RLS + masked PII. It never sees more than you see.

Details and kill switches are in Privacy, external sending, budget and shutdown.

Grounding in the central semantic model#

When there is a published semantic model, text-to-SQL and generation use the model's ids (tables, relationships, measures) instead of raw columns. This brings two benefits:

  1. Less ambiguity: if a term matches ≥2 measures/columns, the AI refuses before spending tokens and asks "which one did you mean?".
  2. Verifiability: if the answer cites an id that does not exist in the published model, the output is discarded — ungrounded SQL does not come out.

That is why a well-defined model (see Semantic model and Modeling concepts) directly improves the quality and safety of the AI.

Measure explanation#

"What does Average ticket mean here?" has two possible answers, and it is worth knowing both:

  • Measure description (deterministic, no LLM): each central measure can have a description written by you (a text field in the model, up to ~300 characters), in addition to folders and the certified badge. This description appears in the interface and does not involve any AI — it is the human source of truth. See Measures, formulas and KPIs.
  • Explanation by the AI (LLM): in Q&A, the AI can explain what a measure does because the model and the description are in the grounding — and every statement about numbers needs a verifiable citation. It describes the calculation; it does not claim cause and effect.

Note

Write clear descriptions for your central measures. They help humans and give the AI solid ground to explain from — reducing the chance of a vague answer.

Why this matters to you#

  • Fewer hallucinations: an AI tied to the model has less room to make things up — see Hallucinations and validation.
  • Privacy by design: PII and data from other workspaces do not get in; rows only when the capability needs them, and always in the secure view.
  • Better with a good model: investing in the semantic model pays off in AI quality.

Limits and caveats#

  • Grounding reduces, but does not eliminate, LLM error — always review (SQL is SELECT only, artifacts come as drafts, citations are checked).

Related#


Last reviewed: 2026-09-08.

Related links