FineFilter: A Fine-grained Noise Filtering Mechanism for Retrieval-Augmented Large Language Models

Retrieved documents containing noise will hinder Retrieval-Augmented Generation (RAG) from detecting answer clues, necessitating noise filtering mechanisms to enhance accuracy. Existing methods use reranking or summarization to identify the most relevant sentences, but directly and accurately locating answer clues from these large-scale and complex documents remains challenging. Unlike these document-level operations, we treat noise filtering as a sentence-level MinMax optimization problem: first identifying potential clues from multiple documents, then ranking them by relevance, and finally retaining the minimum number of clues through truncation. In this paper, we propose FineFilter, a novel fine-grained noise filtering mechanism for RAG, consisting of a clue extractor, a reranker, and a truncator. We optimize each module to tackle complex reasoning challenges: (1) The clue extractor first uses sentences containing the answer and similar ones as fine-tuning targets, aiming to extract sufficient potential clues; (2) The reranker is trained to prioritize effective clues based on the real feedback from the generation module, with clues capable of generating correct answers as positive samples and others as negative; (3) The truncator takes the minimum number of clues needed to answer the question (truncation point) as fine-tuning targets, and performs truncation on the reranked clues to achieve fine-grained noise filtering. Experiments on three QA datasets demonstrate that FineFilter significantly improves QA performance over baselines on both LLaMA3 and Mistral. Further analysis confirms its effectiveness in complex reasoning, robustness to unreliable retrieval, and generalization to different scenarios.