In retrieval-augmented generation (RAG), the way documents are preprocessed—chunking, cleaning, formatting, and similar steps—can have a substantial effect on retrieval quality. Yet these choices are often made heuristically, and comparing them under consistent conditions is difficult.
To address this, the authors of a new arXiv preprint present RAGit, a framework that turns document preprocessing into an experimental variable. By making preprocessing a controlled factor, RAGit aims to allow researchers and practitioners to evaluate different preprocessing strategies in a more rigorous, reproducible way.
The abstract does not provide implementation details or experimental results, but the motivation is clear: without a standard way to vary preprocessing, conclusions about RAG performance may be confounded by arbitrary preprocessing decisions. RAGit appears intended to fill that gap.