The ingestia.io AI translates natural language into SQL over your datamart, answers with citations and provenance, and knows how to decline when it has no basis.
Availability#
- Plans: the ask-your-data feature exists on every plan; AI executive narrative and scheduled AI reports depend on the plan (Growth+ for narrative; Business+ for full reports).
- Protections: the path is covered by automated tests (anti-injection, provenance).
Permissions#
- Using the AI requires the admin/owner role (RBAC action AI features) and runs behind the kill switches for natural language query / AI insights.
- It's an expensive action — blocked when payment is overdue (402).
- Viewing data under RLS/PII: the AI respects row-level security and the member's
See PII data permission (
viewPII).
Setup (step by step)#
- Open the question wizard (Ask your data) in the app.
- Choose the scope: a Gold table or the semantic model (recommended — the AI stays "grounded" in the certified model).
- Write the question in plain language.
- The AI generates the SQL (with named parameters, anti-injection), runs it with a cost guard, and answers with the result + citations of the sources used.
Validation#
- The answer includes provenance: which tables/columns/measures were used.
- The generated SQL is visible for auditing.
- If the question has no basis in the model, the AI declines instead of making things up.
Example#
At Comércio Aurora, Ana asks: "Which branch sold the most in July?". The AI
generates the SQL over prd_aurora_gold.vendas_por_filial, runs it with
the per-query volume ceiling, and answers "Campinas, R$ 128k", citing the table and the
period used.
Cost#
- Each question consumes AI credits (LLM egress) and processing in the datalake.
- The budget uses an atomic ledger (reservation → settlement): you never "overrun" silently; when the credit runs out, the feature fails closed.
- There's a per-workspace ceiling and a fail-closed rate limit. Billed BRL = projection.
Security#
- Anti-injection for prompts and SQL (named parameters).
- Controlled egress: what goes out to the AI provider is mapped; egress is off by default where it requires opt-in.
- The AI respects RLS/ABAC — it doesn't return rows the person couldn't see.
- Answers come with provenance, not raw data without context.
Limits#
- The AI doesn't guarantee absolute accuracy; use the provenance to check.
- No promise of "full parity" with AI tools on the market.
Troubleshooting#
| Symptom | Likely cause | What to do |
|---|---|---|
| AI blocked (402) | subscription past due / no AI credit | settle it in Plans; top up credits |
| "AI feature unavailable" | natural language query/AI insights kill switch on | feature paused; try again later |
| Answer declined | question with no basis in the model | rephrase or ground it in the semantic model |
| Unexpected result | wrong scope/table | check the provenance and adjust the scope |
Next steps: Share the dashboard.