Multi-hop RAG · arXiv 2610.36991

Stop retrieving chunks.
Start retrieving context.

SpectRAG builds a weighted chunk graph from six evidence signals and uses its Laplacian spectrum to select a context set that is relevant, diverse and connected, all under a fixed token budget.

Zafer Sahinoglu · Abdulazeez Mohammed · Buse Yaren Kazangirler · Cahit Berkay Kazangirler

📄 Paper 💻 Code 🤗 Paper page 🎬 Watch the explainer
Animation: classic top-k retrieval picks six near-duplicate chunks and finds 1 of 3 gold passages; SpectRAG seeds, expands, diffuses and selects a diverse set that finds 3 of 3.
Same question, same 6-chunk budget. Top-k finds 1/3 of the gold evidence; SpectRAG finds 3/3. (Illustrative toy corpus.)

Top-k ranks chunks one at a time. Multi-hop questions need a set of chunks that works together as evidence.

How it works

An offline pass builds the graph once; a five-stage online pass selects context for each query.

1

A chunk graph built from six signals

Semantic similarity, document adjacency, section hierarchy, shared entities, LLM-generated utility questions and citations are fused into one weighted edge.

Animation: chunks from five documents are connected signal by signal into a weighted chunk graph.
2

Two geometries for every chunk

The dense embedding captures what a chunk says. The 32-d spectral coordinate from the normalized Laplacian captures where it lives in the corpus.

Animation: points morph from an overlapping embedding layout into well-separated spectral communities.
3

Relevance diffuses along the graph

Hybrid seeds (dense + BM25 + utility questions) are expanded two hops and re-scored with spectral closeness and Personalized PageRank, so bridge evidence surfaces.

Animation: relevance heat spreads from three seed nodes across a graph.
4

Token-budgeted selection

A protected core, a coverage-aware tail with a precision floor, and a backfill fill a ~12k-token window, adapting to 10–15 chunks per query.

Animation: candidate chunks fill a 12,000-token context window; redundant and low-precision chunks are skipped.
Full SpectRAG pipeline diagram: offline indexing and online query-time retrieval.
The full pipeline: offline corpus indexing and online query-time retrieval.

Explainer video

A 98-second narrated walkthrough. Captions are available in the player.

Results

MultiHopRAG (2,556 gold-answer questions) and four UltraDomain corpora, against NaiveRAG, Hybrid, HyDE and LightRAG. Every system uses the same generator.

91.8%overall gold-evidence recall (next best 76.1%)
95.1%exact match on inference questions
~6×smaller index than LightRAG (46 vs 274 MB)
>50%diversity win rate vs every baseline
Line chart: recall at 2, 5, 10 and overall for five systems; SpectRAG reaches 91.8% overall. Bar charts: index size 46 MB vs 274 MB and node counts per corpus, SpectRAG vs LightRAG. Dot plot: exact match by question type; SpectRAG leads on inference at 95.1%. Bar chart: SpectRAG diversity win rate above 50% against all four baselines in both configurations.
Honest trade-offs
  • Hybrid retrieval leads at the very top of the ranking (Recall@2, Recall@5, MRR); SpectRAG wins on breadth.
  • End-to-end accuracy gains are narrow (EM 71.2 vs 70.6 for the next best): better retrieval saturates with short multi-hop answers.
  • On unanswerable (null) questions SpectRAG is behind the vector baselines, and signal weights can need per-corpus tuning.

Citation

@misc{sahinoglu2026spectragweightedchunkgraphs,
      title={SpectRAG: Weighted Chunk Graphs for Token-Budgeted Context Selection in Multi-Hop Retrieval-Augmented Generation},
      author={Zafer Sahinoglu and Abdulazeez Mohammed and Buse Yaren Kazangirler and Cahit Berkay Kazangirler},
      year={2026},
      eprint={2610.36991},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.36991},
}