Adaptive Retrieval-Augmented Generation (RAG) systems such as Self-aware Knowledge Retrieval (SeaKR) typically retrieve passages iteratively based on the Large Language Model (LLM)’s internal uncertainty, but use the full retrieved passages as external context without filtering irrelevant content which can reduce answer generation quality. In this work, we examine how sentence-level filtering methods applied to retrieved passages affect answer generation in uncertainty-based adaptive RAG for multi-hop question answering. We compare two filtering approaches: within-passage sentence filtering (WPSF), which filters sentences within each retrieved passage independently and passes all filtered passages to the LLM, and cross-passage sentence filtering (CPSF), which selects the most relevant sentences globally across all retrieved passages and passes a single fused passage to the LLM. Preliminary results on the HotpotQA and 2WikiMultiHopQA multi-hop question answering datasets show that WPSF is significantly more effective than CPSF, yet neither differs significantly from the unfiltered baseline (SeaKR). A qualitative analysis indicates that the results depend largely on how filtering is integrated into SeaKR: the filtered text is not kept for later reasoning steps, which leads to repeated searches and early stops, and CPSF, by selecting a fixed number of sentences across all passages, can leave out the sentence needed for the answer.