Text to SQL: Local, Secure, and Smarter (www.exasol.com)

Exasol announced an improved Text-to-SQL solution that can run entirely on-premises, keeping data and metadata inside enterprise infrastructure while using local LLM runtimes (they used LM‑Studio with Apple’s MLX format on an M4 Max). That makes natural-language SQL generation viable for security-conscious organizations and supports both ad-hoc querying by wider user groups and tighter control when you want LLM inference inside or outside the database. The team emphasizes trade-offs: embedding LLMs in the DB is possible but can consume database resources and may require larger models; for mostly fixed reports, external SQL generation is preferable. Technically, the solution depends heavily on prompt engineering: dynamically fetching current schema metadata (tables, columns, comments) and injecting it into the system prompt, instructing the model to return raw SQL (Exasol recommends Oracle SQL dialect), and handling dialect/case quirks (e.g., double-quoting mixed-case column names). They use a vector DB (Chroma) to find similar past NL→SQL pairs, store only validated SQL, and extract raw SQL from responses via regex before execution. Implementation notes: LM‑Studio/MLX for performant local inference, an API chat call pattern for generation, a Taipy-based dev bench for iterative prompt tuning, and plans to improve quality with semantic layers and RAG. Results are promising but require per-model and per-schema prompt work and don’t yet reach 100% accuracy.