Enhancing Retrieval-Augmented Generation via Dual-Granularity Document Indexing
R. Uma Mahesh, Kumar Abhishek, G M Pradeepkumar · 2025
This paper presents a novel framework that enhances large language models (LLMs) by integrating multigranular text segmentation with ensemble retrieval techniques. By leveraging dual chunking strategies-a broad context with larger overlapping segments and a narrow context with more precise, smaller segments-we generate complementary embedding representations that are stored in separate FAISS indices. At query time, our system retrieves semantically relevant chunks from both indices and fuses them to create a dual-layered context for the LLM, thus bridging the trade-off between recall and precision. In this paper, we provide detailed mathematical formulations, algorithmic complexity analysis, and extensive experimental results. Our framework not only demonstrates improvements in retrieval metrics but also offers new insights into context augmentation for LLMs.