Adaptive Chunking and Hybrid Retrieval for Enhanced Retrieval-Augmented Generation

In this part, we will look at two important techniques that may help improve the performance of RAG systems: one is adaptive chunking, which involves replacing simple chunking with a method that uses similarity to decide how to split text and adjust chunk size, as well as generating micro-headers. The other technique is the hybrid retrieval approach, which combines both sparse and dense models using RRF, and may also use neural re-ranking.The literature has plenty of examples of both techniques, and their limitations are often complementary.So, combining these might be one way to improve RAG systems in the future.

[1]     G. Kamradt, 5 levels of text splitting,” 2024. [Online]. Available: https://github.com/FullStackRetrieval-com/RetrievalTutorials [2]     T. Chen, H. Wu, S. Zhao, W. Zhang, Y. Duan, M. Wang, Y. Tian, and Z. Hao, “Dense X retrieval: What retrieval granularity should we use?” in Findings of EMNLP, 2023. [3]     P. Sarthi, S. Abdullah, A. Tuli, S. Khanna, A. Goldie, and C. D. Man-ning, “RAPTOR: Recursive abstractive processing for tree-organized retrieval,” in International Conference on Learning Representations (ICLR), 2024. [4]     Anthropic, “Introducing contextual retrieval,” 2024. [Online].Available: https://www.anthropic.com/news/contextual-retrieval [5]     “Optimizing chunking-method selection for RAG,” in Proceedings of LREC 2026, 2026. [6]        “Comparative evaluation of advanced chunking for retrieval-augmented generation in large language models for clinical decision support,” PMC, 2025. [7]        V. Karpukhin, B. Og˘uz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. Yih, “Dense passage retrieval for open-domain question answer-ing,” in Proceedings of EMNLP, 2020. [8]        S. Robertson and H. Zaragoza, “The probabilistic relevance framework: BM25 and beyond,” Foundations and Trends in Information Retrieval, vol. 3, no. 4, 2009. [9]        G. V. Cormack, C. L. A. Clarke, and S. Bu¨ttcher, “Reciprocal rank fusion outperforms Condorcet and individual rank learning methods,” in Proceedings of SIGIR, 2009. [10]    B. Strich et al., “From BM25 to corrective RAG: Benchmark-ing retrieval strategies for text-and-table documents,” arXiv preprint arXiv:2604.01733, 2026. [11]    T. Formal, C. Lassance, B. Piwowarski, and S. Clinchant, “SPLADE: Sparse lexical and expansion model for first stage ranking,” in Proceed-ings of SIGIR, 2021. [12]    O. Khattab and M. Zah