RobustRAG: An Empirical Study of Corpus Poisoning and Query Noise in Retrieval-Augmented Question Answering

This study empirically evaluates the robustness of a dense Retrieval-Augmented Question Answering pipeline against corpus poisoning and query embedding noise. Using BGE embeddings, FAISS, and an extractive QA model on the SQuAD 1.1 validation set, we evaluate three corpus-poisoning strategies, Gaussian query noise, and a lightweight near-duplicate suppression defense. Results show that query noise causes substantially greater degradation in retrieval and answer quality than the studied 5% corpus-poisoning attacks. The suppression defense fails to recover retrieval performance and instead further reduces Recall@5, highlighting a structural limitation of greedy top-down filtering. The study discusses these findings, limitations, and directions for more robust retrieval defenses.