Reference implementation for the CAIL challenge. Translates plain-English requests into HPE NonStop commands, and refuses to guess when it doesn't have enough to go on.
How it works. Each request is matched against a small knowledge base of NonStop commands. Every entry declares its required parameters, why each one matters, and any caveats that apply. Missing a required parameter blocks the command and produces a question instead. A caveat produces a warning that sits above the command.
Where the LLM goes. This prototype uses deterministic matching so it demos reliably offline. In a real build, a model does intent classification and parameter extraction, with retrieval over the manuals supplying the parameter requirements and caveats. The structure of the response stays exactly as it is here: the model proposes, the knowledge base constrains, the operator approves.
Honesty about syntax. Some entries in this knowledge base are marked low-confidence, and the assistant says so rather than presenting a guess as fact. That's the behaviour being demonstrated, not a limitation to work around.