Task-Aware Reduction for Scalable LLM-Database Systems