Recent advancements in large language models (LLMs) demonstrate strong potentials for generating research ideas, yet such ideas often struggle with feasibility and novelty. In this paper, we investigate whether augmenting LLMs with relevant resources during the ideation process can improve idea quality. We made two different attempts: (1) incorporating data from related works as well as preliminary validation to guide models towards more feasible ideas; (2) bringing structure analogies from parallel domains to inspire novel research hypotheses. Both automatic and human evaluations show that our methods can not only improve the quality of the generated ideas, but also help human researchers propose better ideas. Our findings highlight the potential of LLMs in real-world academic settings.