AI concierge: an assistant that only answers from your own content

Grounded retrieval, then a constrained answer

A question is embedded and matched against your published content to retrieve candidates, which are assembled into a numbered context block containing each source's title, slug, excerpt and content. The system prompt then instructs the model to answer using only those excerpts and to support claims with verbatim quotes.

Retrieval is cached per session and query, protected by claim and release locks so two rapid messages in the same conversation cannot both pay for the same retrieval. Long conversations are compressed by a lighter model into a short summary, which keeps later answers coherent without resending the entire thread.

Citations are checked by code

Asking a model to cite its sources is not a control. Verifying the citation is. Every answer passes through a validator that checks whether the cited material supports the claim and repairs or retries once when it does not.

This is the single most important difference between an assistant you can put on a client-facing site and a demo. If the assistant cannot support an answer from your content, it must say so, and the interface shows which sources it read.

• Answers restricted to supplied excerpts by prompt contract
• Citations validated programmatically with one corrective retry
• Sources shown to the visitor alongside the answer
• Irrelevant-question and source-blocklist handling built in

Two different assistants, honestly labelled

The content assistant answers from your published material. On an ecommerce build there is also a storefront agent that can call real tools, searching products and listing stores, inside a multi-step loop with an explicit instruction never to invent product names, prices, stock numbers, store addresses or policy terms, and a bounded number of tool steps so it cannot loop indefinitely.

There is a third, internal variant worth mentioning because clients ask for it: a how-to assistant for staff that answers only from a hand-maintained knowledge base of admin procedures, returning the matching entries and deep links rather than improvising instructions.

Controls, cost and the off switch

Keyword search, AI search and AI chat are separate switches in settings, so a client can run search without chat, or turn chat off during a quiet period. Usage is metered through the credit wallet with pending counters, so conversational cost is visible rather than discovered on a bill.

That granularity is what makes the assistant deployable on a small business site. You can start with search only, watch what people ask, and enable chat once you know your content answers those questions.