SightAndCite
Retrieval that sees the structure and cites the evidence.
Open research on retrieval-augmented generation (RAG), chunking, multi-hop reasoning and efficient LLM systems.
๐ญ Featured: SpectRAG
Stop retrieving chunks. Start retrieving context. SpectRAG replaces independent top-k ranking with context-set selection over a weighted chunk graph and its Laplacian spectrum. It returns evidence that is relevant, diverse and connected, all under a fixed token budget.
| ๐ Paper | arXiv:2610.36991 ยท ๐ค Paper page |
| ๐ฌ Visual walkthrough | sightandcite/SpectRAG |
| ๐ Evaluation results | sightandcite/SpectRAG-eval-results |
| ๐ป Code (MIT) | github.com/SightAndCite/SpectRAG |
Highlights on MultiHopRAG & UltraDomain
- 91.8% overall gold-evidence recall (next best 76.1%)
- 95.1% exact match on inference questions
- ~6ร smaller index than LightRAG (46 MB vs 274 MB)
- LLM judges prefer SpectRAG's answers for diversity against every baseline
๐งญ What we work on
- Retrieval-augmented generation: context-set selection, hybrid retrieval, graph-based RAG
- Chunking: structure-aware and semantic segmentation for long documents
- Multi-hop systems: recovering bridge evidence across chunks and documents
- Efficient LLM pipelines: distillation, lightweight indexes, fully local deployments
๐ฅ Team
Zafer Sahinoglu ยท Abdulazeez Mohammed ยท Buse Yaren Kazangirler ยท Cahit Berkay Kazangirler
๐ฌ Get in touch
Questions, ideas or a retriever that breaks in interesting ways? Open a discussion on the paper page or an issue on GitHub.