Enable natural language interaction over structured data (e.g., SQL databases, CSVs), with capabilities like charting, filtering, and drilldowns.
Translates a question into a query, executes it, and explains the result. Distinct from RAG in that the answer is computed from data rather than retrieved from text, which makes it both more reliable arithmetically and more dangerous when the query is wrong.
Structured data with a stable schema, where the useful questions are too varied to build reports for in advance.
Schemas that are large, poorly named or riddled with tribal knowledge about which table is authoritative. The model cannot know that one of three similar tables is the one anyone trusts.
Read-only credentials, always. A misinterpreted question must not be able to modify data, and this is a database permission rather than a prompt instruction.
A syntactically valid query answering a subtly different question, returning a confident number that is wrong in a way nobody can see. Far more dangerous than a query that errors.
A controller asking which entities have not certified their high-risk accounts — a question nobody built a report for, answered by a generated query against a read-only replica.