Users ask questions in plain language — across a document collection or a transaction ledger — instead of learning a query syntax or opening every record.

Factual claims are tied back to source passages, and unsupported or unverified citations are visibly quarantined.

Counts, filters, and dashboard figures come from a verified index; the model is used for language understanding and synthesis.

The demo uses a limited built-in proxy by default, while still supporting user-provided Claude or open-source model endpoints.