Skip to content

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

Updated on 2026-10-04

AI, alerts and automations

What AI does in the product: questions about the dashboard, insights, natural language queries, forecasts, anomalies and alerts.

On this page (7)

This section documents, honestly and verifiably, everything ingestia.io does with "intelligence" — from purely mathematical (deterministic) calculations to the features that send data to an Anthropic large language model (LLM).

A single principle organizes every page:

The product's AI refuses and warns — it does not pretend. When there is no evidence, it says there is none. When the data is a sample, it warns you. When it cites a number, that number has to be in the data it received. Nothing here claims cause and effect that the data does not prove.


The most important thing: deterministic ≠ LLM#

The product has two families of "intelligence", and they are very different in privacy, cost and guarantees. Do not confuse them.

FamilyWhat it isSends data to an LLM?Token costHow it is proven
DeterministicPure calculations (statistics, regression, threshold comparison) running on our serverNo. Never.ZeroReference tests with identical results on every run; reproducible results
With LLMSends context to an Anthropic model and uses the responseYes (subject to the kill switch)Yes (metered and capped)Deterministic verification over the LLM output (citations, ids, filters)

Anomalies, forecasts and alerts are DETERMINISTIC — there is no LLM in them, no data leaves the product, and the result is reproducible bit for bit. An LLM never decides whether something is an anomaly, never projects a series, and never fires an alert.

Text-to-SQL, Q&A, insights, report narrative and assisted generation use an LLM — and that is why they go through the egress, budget and refusal controls described here.


Capability index#

Use an LLM (Anthropic)

Deterministic — no LLM, no egress

Cross-cutting


Required template for capability pages#

Every capability page follows the same template: the Engine line at the top and 9 fields. The Engine says whether the capability runs only on our server (deterministic) or sends data to an LLM.

The 9 fields

  1. What it does — in one sentence, the value to the user.
  2. Deterministic or LLM? — the engine, and which model/version when there is an LLM.
  3. Data sent and egress — exactly what leaves the product (and what never leaves).
  4. Privacy and retention — PII, RLS, where the record is kept, and for how long.
  5. Tokens, cost and budget — reservation→settlement and monthly cap (or "not applicable").
  6. How to turn it on, who can use it and how to turn it off.
  7. Refusals — when and how the capability declines to answer instead of making something up.
  8. Limitations and absence of causality — what it does NOT do; no unproven cause.
  9. How to validate — concrete steps to check each claim above.

Related links