Rethinking Chunk Size For Long-Document Retrieval: A Multi-Dataset Analysis

TL;DR

The paper tackles how fixed-size chunking affects retrieval in retrieval-augmented generation across diverse long and short documents. It uses a LlamaIndex-based pipeline to systematically vary chunk size from 64 to 1024 tokens and compares two embedding models, Stella and Snowflake, on stitched documents from six QA datasets. Key findings show that small chunks suit concise, fact-based queries while large chunks help with long, dispersed, or technical content, with model-specific biases (Stella benefiting from larger context and Snowflake from smaller chunks). The results provide practical guidance for choosing chunk sizes in RAG and motivate future work on semantic chunk quality measures and richer datasets.

Abstract

Chunking is a crucial preprocessing step in retrieval-augmented generation (RAG) systems, significantly impacting retrieval effectiveness across diverse datasets. In this study, we systematically evaluate fixed-size chunking strategies and their influence on retrieval performance using multiple embedding models. Our experiments, conducted on both short-form and long-form datasets, reveal that chunk size plays a critical role in retrieval effectiveness -- smaller chunks (64-128 tokens) are optimal for datasets with concise, fact-based answers, whereas larger chunks (512-1024 tokens) improve retrieval in datasets requiring broader contextual understanding. We also analyze the impact of chunking on different embedding models, finding that they exhibit distinct chunking sensitivities. While models like Stella benefit from larger chunks, leveraging global context for long-range retrieval, Snowflake performs better with smaller chunks, excelling at fine-grained, entity-based matching. Our results underscore the trade-offs between chunk size, embedding models, and dataset characteristics, emphasizing the need for improved chunk quality measures, and more comprehensive datasets to advance chunk-based retrieval in long-document Information Retrieval (IR).