Text2SQL is dead – long live text2SQL (www.exasol.com)

Exasol’s team demonstrates a full on‑premises Text-to-SQL pipeline to answer the privacy critique that “Text‑to‑SQL is dead” when cloud LLMs leak metadata or data. Their solution combines a local LLM server (Ollama or LM‑Studio with GPU support) with an MCP (Model Context Protocol) gateway that integrates the LLM and the Exasol database, an in‑house Text‑to‑SQL processor built on LangGraph/LangChain, and an AI desktop (Open‑WebUI) front end connected via a small proxy. Key safeguards—what they call “Governed SQL”—limit operations to read‑only, inject database schema and metadata into prompts, retry and validate generated SQL up to three times, and log successful NL→SQL pairs in ChromaDB for similarity‑based hinting. Technically, the stack supports separate LLMs for translation and rendering, uses the OpenAI API protocol to talk to local models, and preserves control by keeping metadata, prompts, and result rendering inside the enterprise. Exasol’s MCP Server and a Text‑to‑SQL extension are available on GitHub, and the authors report good results with models like Qwen3‑coder‑30B while warning LLM hallucinations are possible—so user verification remains essential. The approach is significant because it makes natural‑language querying viable for compliance‑sensitive environments, broadens access to analytics without exposing data, and establishes an auditable, governed workflow for production Text‑to‑SQL.