Screencast
Five walkthroughs of a fully on-prem CS chatbot running on an NVIDIA DGX Spark. Every answer is retrieved, cited, and validated before a customer sees it, and anything touching money or policy is held until a human approves it. One confidently wrong refund promise costs more trust than the productivity gain buys.
The bot answers from the knowledge base and cites the document it used. If retrieval finds nothing relevant, it generates no answer at all — it would rather say nothing than guess.
Money and policy questions are Tier 3: the draft is written but withheld. The customer sees "an agent is reviewing this answer", never an unapproved promise about their refund.
A prompt injection is caught by the input rails and never reaches the model. PII is scrubbed before anything is written to a log.
Phase 2 tool calling. The bot collects the missing order number, calls the order system, and the tool's result is cited and groundedness-checked exactly like a knowledge-base chunk.
The agent workspace, and the point of the whole system. An agent approves, edits, discards, or flags the draft. An edit produces a (context, draft, human answer) triple — which is precisely a training pair.