Enterprise RAG Support Platform

A large operations business runs on knowledge that is scattered across policy manuals, standard operating procedures, portal help articles and the accumulated answers of its support team. The people who need that knowledge - a portal user mid-task, a field inspector on site, an internal admin handling an exception - rarely know which document holds the answer, or whether the version they found is still current.

The cost of getting it wrong is real: a confidently wrong answer about a refund rule, a safety step or a tax treatment wastes time at best and creates compliance exposure at worst. A plain chatbot makes this worse, not better - it answers fluently from nothing, inventing policies that were never written. The job is the opposite of fluency for its own sake: answer only from approved sources, show the evidence, say how confident the answer is, and when the evidence is thin, hand the question to the right human rather than guess.

Four properties make the difference between a demo and something an operations team would trust. It must retrieve from an approved, current corpus rather than the open web. It must cite - every claim traceable to the passage it came from, so a reader can verify it in one click. It must score its own confidence honestly, and route low-confidence questions to a clarification or an escalation instead of bluffing. And it must control cost intelligently, because answering tens of thousands of questions a week with a large model on every one is neither necessary nor affordable.

Retrieval-augmented generation is the shape of the answer: find the few best passages, then write an answer constrained to them. The engineering is in the guardrails around that core - query understanding, an evidence-sufficiency gate, a verification pass that claim-checks the draft against its own citations, and a routing layer that decides between answering, clarifying and escalating.

This is a faithful, downscaled reimplementation on a synthetic approved-knowledge corpus - never the production system. It runs the full nine-stage RAG pipeline end to end: grounded, cited, confidence-scored answers; a clarify path for vague questions; an honest low-confidence path that escalates to a named team with the full conversation and evidence attached; and a transparent cost story you can watch in the inspector.

It deliberately simplifies several things, all labelled in the Architecture doc: production query understanding used a light language model - here intent, clarify and rewrite are deterministic stages; multi-tier model routing is downscaled to one cloud model plus a recorded open-source model; document ingestion, ticketing and SLA routing are simulated; and the multilingual surface is narrowed to English, Dutch and German. The client is not named; all data is synthetic.