Enhancing Retrieval-Augmented Generation Accuracy with Dynamic Chunking and Optimized Vector Search

Abstract

Retrieval-Augmented Generation (RAG) architectures depend on the integration of efficient retrieval and ranking mechanisms to enhance response accuracy and relevance. This study investigates a novel approach to improving the response performance of RAG systems, leveraging dynamic chunking for contextual coherence, Sentence-Transformers (all-mpnet-base-v2) for high-quality embeddings, and cross-encoder-based re-ranking for retrieval refinement. Our evaluation utilizes RAGAS metrics to assess key performance metrics, including faithfulness, relevancy, correctness, and context precision. Empirical evaluations highlighted the significant impact of index choice on the performance. Our proposed approach integrates the FAISS HNSW index with re-ranking, resulting in a balanced architecture that improves response fidelity without compromising efficiency. These insights underscore the importance of advanced indexing and retrieval techniques in bridging the gap between large-scale language models and domain-specific information needs. The findings provide a robust framework for future research in optimizing RAG systems, particularly in scenarios requiring high-context preservation and precision.