Building a Governed Text-to-SQL Assistant for Enterprise Analytics

A new arXiv paper introduces semantic path compilation (SPC), a system that moves the stochastic boundary in text-to-SQL by using a multi-turn planner to ground phrases and select from governed options, while graph traversal, role predicates, grain lowering, SQL construction, and deterministic checks are implemented in code. The approach aims to reduce errors from wrong relationship roles or aggregation grains in enterprise schemas.

Social·Community

Can We Trust LLMs to Turn Natural Language into SQL?

A Reddit post argues that LLM text-to-SQL correctness differs from intent and authorization correctness: valid SQL can omit tenant-scope conditions and expose cross-tenant data. The author proposes an architecture where natural language is converted to a structured intermediate query that is validated against policy checks before execution, and links a tool, QueryForge, claiming it works in production.

A Reddit post argues that LLM text-to-SQL correctness differs from intent and authorization correctness: valid SQL can omit tenant-scope conditions and expose cross-tenant data. The author proposes an architecture where natural language is converted to a structured intermediate query that is validated against policy checks before execution, and links a tool, QueryForge, claiming it works in production.

A Python/Flask app that converts plain-English questions into SQL queries, with built-in guardrails to enforce read-only operations. It uses Telnyx AI Inference for LLM-based SQL generation and includes endpoints for querying, sampling, and validating SQL against a sample dataset. The app solves the problem of unsafe SQL generation by blocking write operations, making it suitable for users who need safe, natural-language database querying.

Researchers propose SPOC-SQL, a method that decomposes text-to-SQL generation into four sequential subtasks aligned with SQL execution logic, enabling stage-wise preference optimization and controllable intermediate generation. This addresses the limitation of existing single-step generation approaches that lack targeted feedback at key decision points.

An AI agent built for a Southeast Asian retail bank lets analysts ask plain-English questions against the data warehouse and get SQL-based answers. An LLM writes SQL from schema context, but the builders found guardrails, validation, and answer verification mattered more than the model itself. It solves unreliable analyst querying over a messy 10-table, 858-column schema.