Adaptive Query Decomposition via Learned Symbolic Primitives in Retrieval-Augmented Systems

Abstract

Retrieval-Augmented Generation systems have emerged as a promising paradigm for enhancing large language models with external knowledge retrieval capabilities. However, complex user queries often require sophisticated decomposition strategies to effectively retrieve relevant information from heterogeneous knowledge bases. This paper presents a novel framework for adaptive query decomposition that leverages learned symbolic primitives to intelligently break down complex queries into retrievable sub-components. Our approach combines neural representation learning with symbolic reasoning mechanisms to create a hybrid architecture that maintains interpretability while achieving robust performance across diverse query types. The proposed system employs a hierarchical decomposition strategy where queries are first analyzed to identify semantic components, which are then mapped to learned symbolic primitives representing fundamental information-seeking patterns through semantic pattern tree construction. These primitives guide the retrieval process through structured query reformulation and contextual expansion across distributed processing layers. Experimental evaluation demonstrates that our method achieves significant improvements in retrieval precision and answer quality compared to baseline neural ranking approaches, particularly for multi-hop reasoning tasks requiring integration of information from multiple sources.