Can we build high-accuracy, low-latency text-to-SQL systems without runtime retrieval? We show the answer is yes! 🔍 We introduce Iterative Prompt Optimization (IPO) — a framework where LLMs refine instructions and exemplars offline to create a static, database-specific prompt for production use, without requiring retrieval at inference time. 💡 Key results: ✅Outperforms strong retrieval-based baselines (e.g., MIPROv2, RES) ✅ Achieves 59.2% execution accuracy on BIRD ✅ Reduces prompt size by ~70% (from 23K to 6.5K tokens) ✅ Supports multi-objective tuning via latency-augmented BIRD-MULTI By eliminating runtime retrieval, IPO delivers faster, simpler, and more reliable NL2SQL pipelines — better suited for real-world deployment. 📄 Read the paper: megagon.ai/publications/effe…, accepted at SIGMOD’s NOVAS Workshop taking place this Sunday, June 22nd. @eserkandogan @subZero_saj #NL2SQL #LLMs #PromptOptimization #SIGMOD2025 #birdmulti #DatabaseAI #TextToSQL #AIResearch #RAG #NLP #NLG2SQL #AI #MachineLearning