Elamur Research⌕
☰
home / papers / 2026 / One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG

One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG

Paper · 2026 · Shivam Ratnakar, Yixuan Zhu, Cecilia Cheng et al.

Summary

Retrieval-augmented generation is usually tuned to answer one question at a time. Enterprise support users, however, tend to arrive with a whole episode of connected needs, and the documents relevant to that episode often sit far apart in embedding space. The authors frame this as a session problem and propose two metrics: the share of a session's documents returned by a single retrieval call, and the number of calls needed to reach a target coverage τ.

Their method reorganizes the knowledge base offline. Document IDs are treated as tokens, and sequences of documents needed together are used to train a Word2Vec CBOW model. These sequences come from expert-labeled article groups, random walks with jumps over a similarity graph, and synthetic QA trajectories. The resulting vectors are then grouped with agglomerative clustering. At query time, a few top hits are expanded with the other members of their clusters, and the combined pool is re-ranked.

On WixQA, a single call now covers 58% of a session instead of 41%. Fewer follow-up calls are needed, and the effective KB becomes much smaller. The authors argue that enterprise RAG should be judged by session coverage, not single-query recall.

Key points

  • Evaluated on WixQA (6,221 enterprise support articles). Session coverage at k=8 rises from 41% to 58%, a gain of +17 points (95% CI: [14.1, 20.4]). Gains of +17 to 21 points hold across four embedding models and in all six functional domains.
  • Retrieval calls needed to reach 70% coverage drop by 34%. Clustering into ⌈N/5⌉ groups shrinks the effective KB to about 20% of its original size.
  • Co-occurrence embeddings are 100-dimensional, trained with window 2 and 10 negative samples. Ground-truth groups are upsampled 10× and all sequences are augmented 3×. Online, the method takes the top-3 vanilla hits, adds roughly 12 cluster candidates, and re-ranks to a final top-8.
  • Trade-off: Hits@8 falls slightly, from 96.0% to 93.0%. On HotpotQA, where documents are not reused across queries, there is no benefit (−1.4%).
  • Limitations:
    • Co-occurrence signals are largely synthetic, since no real navigation logs were used.
    • Sessions are derived from the clusters themselves.
    • The quality of downstream generated answers is not measured.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of Southern CaliforniaLos Angeles · United States175 works · 37,047 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksInformation Systems · 1,730 works

🧭 Topics 2Research subjects and languages the work is about.Information retrieval & search · 552 worksKnowledge graphs & ontologies · 211 works

Cite

Cite as Shivam Ratnakar, Yixuan Zhu, Cecilia Cheng and Chaya Vijayakumar. 2026. One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG. Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.knowfm-1.14

@inproceedings{ratnakar2026one,
  title = {{One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG}},
  author = {Shivam Ratnakar, Yixuan Zhu, Cecilia Cheng and Chaya Vijayakumar},
  booktitle = {Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.knowfm-1.14},
  url = {https://doi.org/10.18653/v1/2026.knowfm-1.14}
}

References

14 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Orion Weller, Michael Boratko, Iftekhar Naim, and Jinhyuk Lee. 2026. On the theoretical limitations of embedding-based retrieval. Preprint, arXiv:2508.21038. — [arXiv:2508.21038](https://arxiv.org/abs/2508.21038) find
  2. Simon Akesson and Frances A. Santos. 2024. Clustered retrieved augmented generation (CRAG). Preprint, arXiv:2406.00029. — [arXiv:2406.00029](https://arxiv.org/abs/2406.00029) find
  3. Oren Barkan and Noam Koenigstein. 2017. Item2vec: Neural item embedding for collaborative filtering. Preprint, arXiv:1603.04259. — [arXiv:1603.04259](https://arxiv.org/abs/1603.04259) find
  4. Dvir Cohen, Lin Burg, Sviatoslav Pykhnivskyi, Hagit Gur, Stanislav Kovynov, Olga Atzmon, and Gilad Barkan. 2025. WixQA: A multi-dataset benchmark for enterprise retrieval-augmented generation. Preprint, arXiv:2505.08643. — [arXiv:2505.08643](https://arxiv.org/abs/2505.08643) find
  5. Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. 2024. RAGAS: Automated evaluation of retrieval augmented generation. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 150–158, St. Julians,…
  6. Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762–1777, Toronto, Canada. Association for Computational Li…
  7. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-augmented generation for knowledge-intensive NLP tasks. Preprint, arXiv:2005.11401.
  8. Weitao Li, Xiangyu Zhang, Kaiming Liu, Xuanyu Lei, Weizhi Ma, and Yang Liu. 2025. Efficient dynamic clustering-based document compression for retrieval-augmented-generation. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 9833–9849, Suzhou, China. Association for Comp… find
  9. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173.
  10. Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2023. Fine-tuning LLaMA for multi-stage text retrieval. Preprint, arXiv:2310.08319.
  11. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. Preprint, arXiv:1301.3781.
  12. Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D. Manning. 2024. RAPTOR: Recursive abstractive processing for tree-organized retrieval. Preprint, arXiv:2401.18059.
  13. Yifan Xu, Vipul Gupta, Rohit Aggarwal, Varsha Mahadevan, and Bhaskar Krishnamachari. 2025. Cluster-based adaptive retrieval: Dynamic context selection for RAG applications. Preprint, arXiv:2511.14769. — [arXiv:2511.14769](https://arxiv.org/abs/2511.14769) find
  14. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ratnakar-2026-one-retrieval-cover-them-all/