What actually moved the needle on Genie

I recently set up a Genie space so a team could ask their sales/pipeline questions in plain English instead of waiting on someone to hand-write SQL every time. Sharing what actually mattered, because it wasn't what I expected going in. When you want to teach the model about your data you might think it is an idea to write a lot of instructions.. The truth is, that is not a very good way to do it. The box where you write those instructions is actually not very helpful. What works a lot better is to take each piece of information and put it in the place, where it makes the most sense. The model can understand your data a lot better that way. The model learns from the data when each fact is, in its place: -Table and column descriptions first -A handful of curated example question + SQL pairs -Declared join relationships, including cardinality -Free text instructions only for stuff with no structure Takeaways if you're doing this: 1. Treat the freetext instructions box as a last resort, not the main tool!! 2. Curated example SQL beats any freetext you can add 3. Build a tiny benchmark of question to expected-SQL mappong and test after every chang 4. Fix wrong answers at the metadata layer, not with prompt Hope this helps!