Enterprise RAG Support Platform

Open the Assistant tab and pick a context - Portal user, Field inspector, or Internal admin. The selector is a cosmetic stand-in for the three real channels; it sets the asker's role and topic area and feeds session context to the pipeline. On a cold start you'll see suggested starter questions per topic; pick one or type your own.

A grounded answer streams back with inline citation chips - click any chip to open the cited source passage in a side panel - plus a confidence meter and the detected topic area. Every answer carries a "How this was answered" link that opens the inspector trace for exactly that turn.

Ask something deliberately vague and the assistant will ask a clarifying question before it searches, rather than guessing at your intent. Ask something the corpus barely covers and you'll get an honest "not confident" reply with two choices: Retry, or Escalate to the suggested team. Escalating creates an entry in the Console's escalation queue with the full thread, the retrieved evidence and the AI's draft attached - the human picks up exactly where the assistant left off.

Open "Behind the scenes" to see the full nine-stage RAG trace on the uniform trace-row contract - stage, provenance, model, tokens, cost and confidence - with deterministic ($0 in both modes) and metered-LLM stages clearly flagged. A Cloud/OSS toggle and a cost meter sit here too; the meter only moves on the generate and verify stages, and at the budget cap it fails closed to the $0 OSS path. The inspector also shows recorded Cloud-vs-OSS divergences, including one honest "not captured" case that is never fabricated.

Everything in this drawer is honestly labelled: synthetic data, simulated integrations, and whether the run was Cloud (live) or OSS (recorded). None of this instrumentation leaks into the answer surface itself.

The Console is one unified, role-gated operator surface. Use the View-as switch (Owner / Viewer / Guest) to see the gating. Overview shows CRM-style metrics - questions asked, deflection and escalation rates, average confidence, citation coverage, feedback - plus breakdowns by topic, team and confidence. The Escalation queue is where a human overrides the AI: answer, reassign, resolve or dismiss, all logged. Knowledge gaps cluster the low-confidence and negative-feedback questions to show where the corpus is weak. The owner-only Configure section tunes the engine; every change is audited. The walkthrough video on the project page narrates the whole tour.