Exploring Knowledge Filtering for Retrieval-Augmented Question Answering

Retrieval-Augmented Generation significantly enhances the performance of Large Language Models on knowledge-intensive tasks by incorporating external knowledge. However, retrieval is not always beneficial, as the presence of harmful knowledge in retrieved documents can degrade generation quality. While existing methods attempt to mitigate this by either assessing the necessity of retrieval or by compressing retrieved documents, they struggle to reject entirely harmful documents selectively. To address this, we propose a novel two-stage framework. First, we establish an annotation strategy to label the harmfulness of a document based on its impact on downstream task performance. Subsequently, our Knowledge Filtering module utilizes this annotated data to learn to identify and eliminate harmful knowledge. For this module, we develop and evaluate four distinct strategies, ranging from methods that leverage LLM self-evaluation and lightweight proxy models to those based on preference alignment. Experiments conducted on three diverse question answering datasets demonstrate that our approach effectively purges harmful knowledge, leading to more reliable and accurate responses.