Efficient Retrieval in RAG Systems

Abstract

This paper presents a comparative analysis of retrieval strategies for Retrieval Augmented Generation (RAG) systems. We evaluate dense, sparse, and hybrid retrieval methods across three benchmark datasets.

Introduction

The quality of a RAG system is fundamentally bounded by its retrieval component. Even the most capable LLM cannot generate accurate answers from irrelevant context.

Methods

We evaluated three retrieval strategies:

1. Dense Retrieval — bi-encoder models (e.g., text-embedding-ada-002)
2. Sparse Retrieval — BM25 with custom tokenization
3. Hybrid — weighted combination of dense + sparse scores

Results

Conclusion

Hybrid retrieval consistently outperforms either approach alone, with only modest latency overhead.