Astronomical literature is expanding at an unprecedented rate, with thousands of papers added every month to preprint servers like arXiv.org and indexed by the NASA Astrophysics Data System (ADS). For academics and students, staying current with relevant work while keeping track of shifting trends therefore represents a critical challenge. I will talk about lessons learned from developing Pathfinder, a complement to systems like ADS that uses large language models combined with retrieval-augmented generation (RAG) to enable semantic search and question-answering across the astronomy literature. I will discuss some of the unique challenges of applying NLP and LLMs to scientific publications in astronomy, including grounding LLM responses in published literature to minimize hallucinations, and leveraging embeddings to create interpretable semantic spaces for literature exploration. Drawing from Pathfinder's deployment (pfdr.app) and user feedback from the astronomy community, I will highlight how interpretable intermediate representations such as semantic embeddings and citation graphs can lend interpretability and rigor to otherwise black-box models, and help their adoption in research pipelines.