Train your bot

Knowledge sources

The three things Relay learns from: a crawled website, uploaded documents, and FAQ pairs.

Everything Relay answers with comes from a knowledge source. If it is not in one, the bot will say it does not know.

Website crawl

Point Relay at a URL and it follows links from there, extracting text as it goes. Crawls are capped at a configurable page limit and run as background jobs.

Re-crawling on a schedule is what keeps answers current: a bot trained once against a site that has since changed will confidently quote last quarter's pricing.

Documents

Upload PDFs, Word documents, and similar. Uploads go directly to storage and are then chunked and embedded by a background worker, so a large document does not block the request.

FAQs

Explicit question-and-answer pairs. These are the highest-precision source: use them for anything where the wording of the answer matters: refund policy, SLAs, anything legal.

Canned responses

Distinct from FAQs. A canned response is a reply your human agents can insert from the inbox, not something the bot serves on its own.

Retrieval

Sources are chunked with an overlap, embedded, and stored in a vector index. Each question retrieves the closest chunks above a similarity threshold; if nothing clears the threshold, Relay declines rather than answering from a weak match.

Reindexing

Knowledge statistics show what is indexed and how much. After a large content change you can force a reindex rather than waiting for the next scheduled crawl.

On this page