<p>Retrieval-augmented generation (RAG) improves large language models by grounding generation in external evidence. However, in practice, retrieved passages often contain redundancy, noise, or conflicting information, while existing re-ranking methods still predominantly optimize query–passage relevance rather than generation utility, namely, the extent to which a passage contributes to generating an answer that is both correct and faithful to the provided evidence. Meanwhile, answer confidence in RAG is often estimated from model uncertainty alone, without sufficiently accounting for whether the answer is grounded in the retrieved evidence. To address these limitations, we propose Faith-Rank, a faithfulness-aware utility ranking framework for RAG. Faith-Rank reformulates evidence re-ranking as a utility-driven task, where passages are ranked by their contribution to generating answers that are both correct and faithful to the provided evidence. Concretely, we develop a Siamese utility ranker trained with a hybrid objective that combines margin-based ranking and binary classification, enabling both fine-grained evidence ordering and stable score calibration. We further introduce a joint reliability modeling module that integrates evidence-grounded faithfulness with generation uncertainty for answer assessment. Experiments on multiple open-domain QA benchmarks show that Faith-Rank consistently improves evidence quality and downstream answer performance across retrievers and LLM backbones. In particular, Faith-Rank improves Top-1 evidence accuracy by about 20% on average and Top-10 accuracy by about 13% on average over BM25, while also yielding more trustworthy RAG outputs.</p>