In Search of a Better Search: Using RAG on Cornell Virtual Workshop Content

Volume 17, Issue 2 (November 2026), pp. 34–38

Download PDF

@article{jocse-17-2-6,
  author={Zilu Wang and Susan Mehringer and Christopher J. Cameron},
  title={In Search of a Better Search: Using RAG on Cornell Virtual Workshop Content},
  journal={The Journal of Computational Science Education},
  year=2026,
  month=nov,
  volume=17,
  issue=2,
  pages={34--38},
  doi={https://doi.org/10.22369/issn.2153-4136/17/2/6}
}

The Cornell Virtual Workshop (CVW) is a continually updated HPC training platform with about 40 roadmaps and more than 2000 pages, organized in a three-level hierarchy of roadmaps, topics, and pages. The CVW's traditional keyword search works poorly for users who lack domain terminology or want navigational guidance rather than a specific page. We describe a RAG-based chat-bot that routes each query to a retriever matched to the appropriate level of the content hierarchy, ensuring that navigational queries return information about CVW structure and factual queries return specific page content. The chatbot is built with LangGraph and uses a hybrid BM25 and FAISS retrieval pipeline with cross-encoder reranking. A fixed-path pipeline provides prompt-injection defense, query rewriting, post-generation URL verification, and graceful handling of out-of-scope queries.