SiReRAG: Indexing Similar and Related Information for Multihop Reasoning

TL;DR

SiReRAG tackles the bottleneck of relying on a single information perspective in RAG indexing by jointly modeling similarity and relatedness. It builds two trees—one from recursive summaries of semantically similar chunks and one from aggregated, entity-centered propositions linked by shared entities—and flattens them into a unified retrieval pool. Across MuSiQue, 2WikiMultiHopQA, and HotpotQA, SiReRAG yields consistent improvements over state-of-the-art baselines, with an average F1 gain around 1.9% and notable gains when integrated with other reranking methods. The approach is efficient enough for practical use and demonstrates broad applicability across retrieval backends, strengthening multihop QA capabilities with comprehensive knowledge integration.

Abstract

Indexing is an important step towards strong performance in retrieval-augmented generation (RAG) systems. However, existing methods organize data based on either semantic similarity (similarity) or related information (relatedness), but do not cover both perspectives comprehensively. Our analysis reveals that modeling only one perspective results in insufficient knowledge synthesis, leading to suboptimal performance on complex tasks requiring multihop reasoning. In this paper, we propose SiReRAG, a novel RAG indexing approach that explicitly considers both similar and related information. On the similarity side, we follow existing work and explore some variances to construct a similarity tree based on recursive summarization. On the relatedness side, SiReRAG extracts propositions and entities from texts, groups propositions via shared entities, and generates recursive summaries to construct a relatedness tree. We index and flatten both similarity and relatedness trees into a unified retrieval pool. Our experiments demonstrate that SiReRAG consistently outperforms state-of-the-art indexing methods on three multihop datasets (MuSiQue, 2WikiMultiHopQA, and HotpotQA), with an average 1.9% improvement in F1 scores. As a reasonably efficient solution, SiReRAG enhances existing reranking methods significantly, with up to 7.8% improvement in average F1 scores. Our code is available at https://github.com/SalesforceAIResearch/SiReRAG .

Figures (8)

• Figure 1: Challenges of existing RAG indexing methods for multihop reasoning. Entity 1 and 2 chunks contain similar information while entity 2 and 3 chunks contain related contents. Since synthesizing information only based on entity 1 and 2 (or entity 2 and 3) will lead to a higher probability of a wrong answer, an indexing method that considers both similarity and relatedness is needed to maximize retrieving relevant knowledge for multihop questions.
• Figure 2: SiReRAG Tree. We adopt RAPTOR sarthi2024raptor to construct the similarity tree (left). On the right, we construct the relatedness tree by clustering the propositions based on their entities to get proposition aggregates and having recursive summaries on top. Note that propositions are not included in the relatedness tree, so their connections to proposition aggregates are marked with dashed lines.
• Figure 3: Examples of propositions with and without associated entities.
• Figure 4: Relevant part of the SiReRAG tree for the question: "who is the father of the artist who painted Head I?".
• Figure 5: Prompt of identifying a two-level hierarchy for all candidate chunks.