IndexRAG: Bridging Facts for Cross-Document Reasoning at Index Time

TL;DR

IndexRAG introduces an offline indexing method for multi-hop question answering that improves retrieval efficiency and accuracy by generating and indexing bridging facts, outperforming existing graph-based approaches.

Contribution

It shifts cross-document reasoning to offline indexing, enabling single-pass retrieval and improving multi-hop QA performance without additional online processing.

Findings

IndexRAG improves F1 by 4.6 points over Naive RAG.

It requires only single-pass retrieval and one LLM call.

Outperforms graph-based baselines when combined with IRCoT.

Abstract

Multi-hop question answering (QA) requires reasoning across multiple documents, yet existing retrieval-augmented generation (RAG) approaches address this either through graph-based methods requiring additional online processing or iterative multi-step reasoning. We present IndexRAG, a novel approach that shifts cross-document reasoning from online inference to offline indexing. IndexRAG identifies bridge entities shared across documents and generates bridging facts as independently retrievable units, requiring no additional training or fine-tuning. Experiments on three widely-used multi-hop QA benchmarks (HotpotQA, 2WikiMultiHopQA, MuSiQue) show that IndexRAG improves F1 over Naive RAG by 4.6 points on average, while requiring only single-pass retrieval and a single LLM call at inference time. When combined with IRCoT, IndexRAG outperforms all graph-based baselines on average, including HippoRAG and FastGraphRAG, while relying solely on flat retrieval. Our code will be released upon acceptance.