Year: 2025
Type: conference-paper
Abstract: Retrieval Augmented Generation (RAG) combines chunks of documents with a prompt to be analyzed by an LLM to enhance the response's quality and reduce the risk of hallucinations.Existing chunking methods divides the text into chunks of predetermined length (fixed-size strategy), segments text at punctuation marks (recursive strategy), or aims at capturing the core meaning within each segment (semantic strategy).All these strategies limit context accessibility and encounter information loss, especially with larger documents and multi-layered information.They often require post-processing steps to avoid omitting crucial information, especially toward the documents' end.This paper introduces a paragraph-based chunking strategy that can benefit from larger context blocks.The proposed strategy leverages document structure metadata, such as font size and text formatting, to generate more semantically coherent chunks.The approach has been evaluated using both automatic and empirical approaches by exploiting standard metrics.The results demonstrate a promising enhancement for RAG applications by facilitating more effective document comprehension and improving context utilization in LLM-driven tasks when compared with existing strategies.
Source: Proceedings
DOI: 10.18293/dms2025-033
Institution: University of Sannio
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4415412958