Intelligent Retrieval Systems for Accelerator Physics: Matching Retrieval Approaches to Use Cases

Accelerator facilities generate diverse documentation, from technical reports to structured wikis and semi-structured logbooks, which complicates efficient knowledge access. While Retrieval-Augmented Generation (RAG) offers a path toward intelligent operator assistants, no single method is universally optimal. We present three use cases from PSI: for technical documentation, naive dense retrieval with summarization provides fast and interpretable access; for the AcceleratorWiki, a graph-augmented approach improves reasoning over hierarchies and cross-references; and for ELOG, an agentic pipeline with specialized agents supports multimodal interpretation, temporal reasoning, and iterative refinement. Together, these case studies illustrate how matching retrieval paradigms to data types enables reliable, context-aware assistance in accelerator operations.