view article
Figure 5
Seven-layer NLP action pipeline. User input in natural language (Korean, English, or Japanese) is pre-processed in Layer 1 (element/edge detection, intent classification, language auto-detection, knowledge-query routing) and sent to the selected LLM backend (Layer 2). The raw output then passes through five post-processing layers (Layers 3–7) that perform fuzzy function-name matching against 47 registered beamline functions, empty-response retry, energy-range validation, domain rule enforcement (function-signature checks, motor-group constraints, scan-motor compatibility), and a sample-preparation gate that blocks measurement actions until the user confirms that a physical sample is mounted and aligned. Knowledge-query inputs are routed instead to a retrieval-augmented generation branch, bypassing the command pipeline. The entire pipeline is deterministic and stateless so that each layer can be debugged and extended independently.

Journal logoJOURNAL OF
SYNCHROTRON
RADIATION
ISSN: 1600-5775
Volume 33| Part 5| September 2026|
https://doi.org/10.1107/S1600577526006788
Follow J. Synchrotron Rad.
Sign up for e-alerts
E-alerts
Follow J. Synchrotron Rad. on Twitter
Twitter
Follow us on facebook
Facebook
Sign up for RSS feeds
RSS