RedParrot: Accelerating NL-to-DSL for Business Analytics via Query Semantic Caching

Researchers from Xiaohongshu have published a paper introducing RedParrot, a novel framework that accelerates natural language to domain-specific language (NL-to-DSL) conversion for business analytics. The system addresses a critical bottleneck: existing multi-stage LLM pipelines for this task suffer from high latency, cost, and error propagation, making them unsuitable for real-time enterprise use at scale. RedParrot's core innovation is a semantic cache that stores 'query skeletons'—normalized structural patterns derived from frequent user queries. When a new request arrives, the system matches it against these skeletons and adapts their corresponding DSL outputs, bypassing the expensive LLM pipeline entirely.

RedParrot's technical contributions include an offline skeleton construction strategy, an entity-agnostic embedding model trained via contrastive learning for robust matching, and a heterogeneous Retrieval-Augmented Generation (RAG) method that integrates diverse knowledge sources to handle unseen entities. Tested on six real enterprise datasets from Xiaohongshu, RedParrot delivered an average 3.6x speedup and an 8.26% accuracy improvement over baseline pipelines. On public benchmarks adapted from Spider and BIRD, accuracy jumped 34.8%, substantially outperforming standard in-context learning baselines. The framework is designed for high-repetition, stable-pattern workloads common in e-commerce and advertising analytics, offering a practical path to low-latency, cost-effective business intelligence.