Reproduce “POQD: Performance-Oriented Query Decomposer for Multi-vector retrieval” on OpenResearch

About this paper

Yaoyang Liu, Junlin Li, Yinjun Wu, Zhen Chen

POQD proposes a performance-oriented query decomposer for Multi-Vector Retrieval that utilizes an LLM-based optimizer to adaptively generate sub-queries, enhancing the performance of Retrieval-Augmented Generation (RAG) systems. This framework achieved up to 12.61% higher QA accuracy and 5.28% higher retrieval accuracy on various datasets compared to existing methods, while maintaining computational efficiency.

Key insights

• Fixed token-level query decomposition in Multi-Vector Retrieval (MVR) is suboptimal for Retrieval-Augmented Generation (RAG) systems, indicating a need for more adaptable, performance-oriented decomposition.
• Large Language Models (LLMs) can effectively serve as meta-optimizers to iteratively search for optimal prompts for query decomposition, addressing the non-differentiability of the sub-query generation process.
• An alternating training algorithm that limits downstream RAG model training iterations while optimizing the query decomposition prompt provides an efficient and effective end-to-end solution, supported by theoretical analysis.

Summary by alphaXiv.