Meet SpectRAG: set-based RAG retrieval hits 91.8% gold-evidence recall on multi-hop QA
An author of SpectRAG announced the method on Reddit, presenting it as a fix for multi-hop question retrieval. Instead of ranking chunks individually against the query, SpectRAG selects a set of chunks that work together as evidence, capturing bridge passages top-k retrieval misses. It reports 91.8% gold-evidence recall on MultiHopRAG versus a 76.1% next-best baseline, with a 6x smaller index than LightRAG. Code and paper are public, and the author is soliciting comparison requests.
- SpectRAG selects a set of chunks working together as evidence instead of ranking chunks individually, targeting multi-hop QA where top-k retrieval misses bridge passages
- Scores 91.8% gold-evidence recall on MultiHopRAG and UltraDomain versus 76.1% for the next-best baseline, and 95.1% exact match on inference questions
- Runs a 6x smaller index than LightRAG at 46 MB versus 274 MB; LLM judges preferred its answer diversity over every tested baseline
- Paper posted on arXiv (2610.36991); MIT-licensed code runs fully local
Related
SpectRAG: Weighted Chunk Graphs for Token-Budgeted Context Selection in Multi-Hop Retrieval-Augmented Generation
A new arxiv paper proposes SpectRAG, a retrieval-augmented generation framework that builds weighted chunk graphs and uses spectral relevance propagation to select token-budgeted context for multi-hop reasoning. It outperforms dense, hybrid, HyDE, LightRAG, and NaiveRAG baselines on MultiHopRAG and UltraDomain benchmarks. The approach matters to practitioners building RAG pipelines that need cross-document, multi-hop evidence rather than independently ranked passages.
A new arxiv paper proposes SpectRAG, a retrieval-augmented generation framework that builds weighted chunk graphs and uses spectral relevance propagation to select token-budgeted context for multi-hop reasoning. It outperforms dense, hybrid, HyDE, LightRAG, and NaiveRAG baselines on MultiHopRAG and UltraDomain benchmarks. The approach matters to practitioners building RAG pipelines that need cross-document, multi-hop evidence rather than independently ranked passages.
SAG2: Embedding-Free Multi-Hop Retrieval with BM25 and JEV
An arxiv paper introduces SAG2, a multi-hop retrieval architecture that drops dense embeddings, LLM entity extraction, and dedicated database servers in favor of a single SQLite database with BM25 lexical indexing. It reports top Recall@5 and Recall@10 on multi-hop QA subsets, with indexing of 11,656 passages taking 2.36 seconds and no model calls. The result makes strong multi-hop retrieval deployable with lightweight, fast-to-prepare infrastructure.
An arxiv paper introduces SAG2, a multi-hop retrieval architecture that drops dense embeddings, LLM entity extraction, and dedicated database servers in favor of a single SQLite database with BM25 lexical indexing. It reports top Recall@5 and Recall@10 on multi-hop QA subsets, with indexing of 11,656 passages taking 2.36 seconds and no model calls. The result makes strong multi-hop retrieval deployable with lightweight, fast-to-prepare infrastructure.
MatRAG: Matryoshka Hierarchical RAG Cuts Indexing and Query Costs for Multi-Hop QA
An arXiv paper presents MatRAG, a hierarchical RAG framework combining RAG with Matryoshka Representation Learning for efficient multi-hop question answering. It cuts both indexing and query-time costs while maintaining retrieval quality.
An arXiv paper presents MatRAG, a hierarchical RAG framework combining RAG with Matryoshka Representation Learning for efficient multi-hop question answering. It cuts both indexing and query-time costs while maintaining retrieval quality.
LiteRAG: Cost-Efficient Graph-Based Retrieval-Augmented Generation
A new arXiv paper introduces LiteRAG, a graph-based retrieval-augmented generation method targeting multi-hop question answering. It replaces retrieval-time LLM control with query-conditioned algorithmic exploration and builds contexts from reasoning chains. The approach aims to cut query-time costs and avoid the diffuse, oversized contexts that hurt generation efficiency in existing graph-based RAG systems.
A new arXiv paper introduces LiteRAG, a graph-based retrieval-augmented generation method targeting multi-hop question answering. It replaces retrieval-time LLM control with query-conditioned algorithmic exploration and builds contexts from reasoning chains. The approach aims to cut query-time costs and avoid the diffuse, oversized contexts that hurt generation efficiency in existing graph-based RAG systems.
QUBO Formulation for Evidence Selection in RAG
A new arXiv paper proposes formulating evidence selection in retrieval-augmented question answering as a Quadratic Unconstrained Binary Optimization (QUBO) problem, enabling the use of unconventional solvers like quantum annealers. This approach aims to overcome the limitations of top-k ranking and costly LLM-based selectors for multi-hop questions.
A new arXiv paper proposes formulating evidence selection in retrieval-augmented question answering as a Quadratic Unconstrained Binary Optimization (QUBO) problem, enabling the use of unconventional solvers like quantum annealers. This approach aims to overcome the limitations of top-k ranking and costly LLM-based selectors for multi-hop questions.
