What is SmartQuery

SmartQuery is an enterprise NL2SQL platform built by Quanam for production use. At its core is an AI agent that reasons over the database schema and a set of tools to translate plain-language questions into SQL. Users get back the generated SQL, the execution results, a natural-language answer, and — when the result is suited for it — an automatically generated chart.

Key capabilities

• Natural-language to SQL: ****an agent reasons over the question and available tools to produce executable SQL against the user's connected database(s), validating and retrying as needed.
• Schema profiling: automatically profiles tables, columns, and primary/foreign key relationships to ground SQL generation in the real structure of each datasource.
• Per-datasource business rules: lets teams encode domain-specific conventions (e.g., how names, percentages, or numeric types should be handled) that guide how SQL gets generated for their data.
• Post-generation validation: checks generated queries before returning them to catch structural issues.
• Automatic chart generation: when a query result is suited for it, SmartQuery automatically picks a chart type and renders a visualization alongside the answer.
• Knowledge base (RAG): an optional retrieval-augmented layer for grounding answers in domain documents, on top of the database schema itself.

Security

• Role-based access control: per-datasource and per-user roles determine which users can query which data.
• Row-level security policies: fine-grained policies can restrict which rows a user's queries are allowed to return.
• Query validation before execution: every generated query is checked against access policies (unauthorized table/column access is blocked) and executed in read-only mode.

Model

SmartQuery is model-agnostic; the results below were produced using GPT-5.1.

BIRD Mini-Dev Benchmark Results (PostgreSQL)

To validate SQL generation quality against a public, reproducible benchmark, we evaluated SmartQuery on the BIRD Mini-Dev PostgreSQL split (500 questions) using the official evaluation_ex.py script from bird-bench/mini_dev.

For this evaluation, SQL generation used schema profiling and per-datasource business rules. The knowledge-base (RAG) layer was not used, and no restrictive per-user or row-level security policies were configured for the BIRD datasources (evaluation ran with unrestricted access). The BIRD evidence hints were included in the input question text.

Execution Accuracy (EX)