Posts by Mimoune Djouallah

LLMs + Semantic Models = Smarter SQL

Getting accurate Text-to-SQL results isn’t magic, you need two things: smarter language models and a clear semantic model that explains what your data actually means. LLMs are getting better at understanding natural language, but without structure, they still tend to guess or hallucinate. To explore this, I built a small experiment using a local DB setup, a semantic layer over a typical dataset , and a set of benchmark questions. I tested different models like o3-mini , Qwen3 to see how well they generate SQL when given proper context. The big takeaway? With the right metadata, even smaller local models can do surprisingly well, and running accurate Text-to-SQL on everyday hardware isn’t just possible, it’s only a matter of time.