Particle.news
Download on the App Store

Study Finds Retrieval Essential for LLM SQL and API Generation

New arXiv results show large accuracy gains from retrieval for structured tasks, with CoRAG proving most robust in heterogeneous documentation.

Overview

  • An arXiv evaluation released Feb. 10 tests standard RAG, Self-RAG, and CoRAG on SQL, REST API calls, and a combined classification task using an SAP Transactional Banking dataset.
  • Without retrieval, exact-match accuracy was 0% across tasks, while retrieval raised execution accuracy to as high as 79.30% and component-match accuracy to 78.86%.
  • CoRAG led in hybrid documentation settings, reaching 10.29% exact match on the combined task versus 7.45% for standard RAG, driven by stronger SQL generation at 15.32% versus 11.56%.
  • The study identifies retrieval-policy design as pivotal, finding iterative query decomposition outperforms top‑k retrieval and binary relevance filtering under documentation heterogeneity.
  • A same-day DEV explainer outlines the RAG workflow, common use cases, and starter tooling such as vector stores and prompt augmentation for developers adopting the pattern.