Business Problem & Solution

Most guest questions are the same dozen questions — check-in time, the cancellation policy, whether pets are allowed, where to park, how loyalty tiers work. They arrive at every hour, in the booking flow and mid-stay, and each one that waits on a human is a worse experience and a support cost. But the obvious fix — a chatbot — fails in the way that matters most for a hospitality brand: it makes things up, and a confidently wrong answer about a refund or a pet policy is worse than no answer at all.

The Concierge & Knowledge Assistant is the 24×7 front line for those questions. It answers only from a known knowledge base, shows its sources and a confidence level, and — crucially — when it can't ground an answer it says so and offers to escalate to a human rather than inventing one.

It proves the grounded-answer contract: every reply is retrieved from the synthetic knowledge base, grounded on the top one or two articles, and rendered with its sources `[1][2]` and a confidence percentage. An ungrounded question never produces a guessed answer — it produces an escalation offer, and accepting it raises a real support ticket that shows up for the property team. That honest dead-end is the feature, not a gap.

It simplifies the retrieval: at this stage the knowledge base is a small synthetic corpus and matching is deterministic term-overlap rather than a vector store. The pipeline shape — retrieve → ground → answer → guard — is exactly what a production RAG system runs; only the corpus size and the retriever are downscaled, and the architecture doc says so plainly.

Synthetic knowledge only — invented policy/facility/loyalty articles, no real brand policies or PII. The escalation confirmation is a simulated message in the in-app sent viewer, not a real email. Confidence percentages and any figures shown are representative and labelled. The knowledge corpus is English-only at this stage; multi-language answering is scoped but not seeded.