Most chat agents are built to sound smart. Cadey is built to be right or say it does not know. Here is the exact pipeline behind every answer.
Each one can refuse. None of them can fabricate.
We rewrite the visitor’s message into a precise search query with a fast, cheap model, using the conversation so far.
We search only your published knowledge, your site, docs, and house answers. Below a confidence threshold there is nothing to answer from, and we say so.
The answer is written strictly from the retrieved sources and your rules. No outside facts, prices, availability, or policy.
A separate grader checks every claim against the sources. If it is not grounded, the draft is discarded. We would rather decline than guess.
Grounded? The visitor gets the answer with sources. Not grounded, or they ask for a person? Human takeover, a booked meeting, or a ticket. Never a dead end.
THE REAL PIPELINE IN PACKAGES/AGENT, THE PAGE CANNOT DRIFT FROM THE CODE
No, by construction. The agent answers only from your published content and tool results; a separate validation step rejects anything ungrounded, and below the confidence threshold it refuses and offers a human, a meeting, or a ticket.
A grounded retrieval pipeline backed by a commercial foundation model for answers, a fast model for refine and validation, and Voyage embeddings for retrieval. We name our real stack rather than claim a proprietary model, and that transparency is the point.
It hands off honestly: a human can take over mid-conversation, the visitor can book a meeting on a real calendar, or a support ticket is filed with the full transcript attached.
Point the crawler at your site, review and publish the drafted articles, drop in one script tag. Minutes, not weeks.
Crawl your content, review the drafts, drop in one script tag. The grounded pipeline runs from minute one.