Reproduce “Influence Factors on RAG Poisoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pedro Pereira, Eva Maia, Isabel Praça, Adrien Bécue
A systematic experimental study quantified how various architectural and configuration choices within Retrieval-Augmented Generation (RAG) systems influence their susceptibility to poisoning attacks. The research reveals that vulnerability is a complex interplay of retrieval strategy, generative model, and knowledge base composition, affecting both document retrieval and final answer generation.
Key insights
- RAG system poisoning vulnerability is an emergent property resulting from complex interactions between retrieval, generation, and knowledge base components, not isolated factors.
- Advanced retriever architectures like dense BGE and graph-based methods consistently enhance robustness against poisoning compared to lexical approaches such as BM25.
- The choice of generator LLM and the configuration of the knowledge base, including the number of poisoned databases, critically influence the ultimate success rate of poisoning attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.