Title: Optimizing RAG Through Intent-Driven Retrieval and Semantics-Focused Chunking

The growing requirement for large language models (LLMs) to follow instructions and reason effectively has accelerated the evolution of retrieval-augmented generation (RAG). RAG systems enhance LLM outputs by pulling relevant supplementary knowledge fragments from external databases that match the user's query. However, standard RAG approaches often fall short due to limited information availability, a problem stemming from two main issues: retrieval that ignores user intent and the fragmentation of information.

To overcome these limitations, we introduce InSemRAG, a new RAG framework designed to tackle these challenges through an iterative retrieve-and-check process. This framework relies on two core components: an Intention-Aware Retriever (IAR) and Semantics-Preserving Chunking (SPC). The IAR utilizes a dynamic hybrid retrieval strategy that adjusts the weight of different retrieval channels according to the specific intent of the query. Meanwhile, the SPC module identifies and repairs damaged evidence chunks to ensure semantic integrity is maintained.

To mitigate the computational delays associated with this iterative approach, we employ small language models (SLMs). Our extensive testing across various benchmark datasets shows that our method competes effectively with the latest state-of-the-art RAG techniques. Specifically, it delivers substantial improvements on tasks requiring multi-hop reasoning and evidence sensitivity, achieving a 2.65-point rise in F1 score on HotPotQA and a 1.5-point boost in accuracy on FEVER. Furthermore, by using SLMs, our approach matches the performance of Multi-Hop RAG while operating with 4.32 times lower latency.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC