This simulator traces a query through the five-layer guardrail pipeline for four reference scenarios. Each scenario is scripted to demonstrate a specific guardrail firing condition. A free-text input mode is also available to run arbitrary queries against the live backend.
From voice input to cited response, the pipeline executes nine steps in sequence. The simulator above traces each step with timing. The steps below describe the mechanism at each stage.
One-handed. One button. Web Speech API captures audio from the device microphone, no install required.
Channel: Browser micRaw transcription ("uh E zero four error on the Haas") is cleaned and reformulated into a structured query by the LLM with a strict system prompt.
< 0.8sTop-1 similarity pre-check against the corpus. If nothing is remotely relevant, refuse immediately before spending retrieval budget.
< 0.3sTop-3 chunks by cosine similarity. nomic-embed-text embeds the query. Procedural sections returned with metadata: doc title, section, page range.
< 1.2sBest chunk similarity must be ≥ 0.70. Below threshold: refuse with explanation. Never generate a response the corpus cannot support.
< 0.1sScans retrieved chunks for LOTO, lockout, high voltage, pressure vessel, hazmat keywords. If triggered, mandatory safety prefix prepended to response.
< 0.2sStructured prompt enforces step-by-step format, maximum 5 steps, citation required. Response generated entirely locally — no external API call.
< 4.5sValidates source reference is present in output. If missing, one retry with stricter prompt. If still absent, response is blocked and refused.
< 0.2sCited, structured procedure appears on the technician's phone. Section number, page range, document title. Traceable to source.
Total: < 8s (target)