Retrieval-Augmented Generation (RAG), a method for integrating external knowledge into large language models (LLMs), is typically composed of a retriever, which searches for relevant information, and a generator, which produces responses based on the retrieved content. This study aims to examine whether RAG can function effectively when using a small language model (SLM) as the generator, assuming the retriever has sufficiently high performance. In our experiments, we conducted a question-answering (QA) task using the JQaRA dataset under an idealized condition where the retriever was assumed to always retrieve the correct document. The results indicate that even when the retriever demonstrates high performance, using a more capable model as the generator leads to an overall improvement in response accuracy. However, a certain number of test cases were observed where a lower-performing SLM produced the correct answer, while a higher-performing SLM failed. To address this, we investigated the potential of ensemble RAG, which combines multiple SLMs. The results suggest that incorporating multiple SLMs as an ensemble within the generator mechanism could significantly improve overall accuracy.