How the agent works

Five stages between a question and an answer

This page is the honest version, the same one we give technical buyers. No magic claimed, no steps skipped.

stage 1

Read

Your public pages are fetched within documented caps and split into passages, each tied to the page it came from. Nothing else ever enters the agent's world: no training data about your business, no third-party summaries, no guesses.

stage 2

Retrieve

For each visitor question, Sitora ranks your passages and selects the few that plausibly contain the answer. If nothing ranks well, that is already a signal: the question may be outside your site.

stage 3

Answer inside a contract

The selected passages and the question go to a language model with strict rules and a required JSON shape: the answer text, an action, and the sources used. A reply that breaks the shape, or cites pages it was never given, is rejected in code.

stage 4

Act

The action field is how the agent does more than chat. "qualify" asks one good question of a promising visitor. "book" proposes the exact next step your site offers, like a phone number or booking link. Both can capture the visitor's details into your lead inbox.

stage 5

Hand off

When retrieval is weak, when the model cannot ground an answer, or when the visitor simply asks for a person, the agent says it does not know and offers the human path. The handoff message is yours to write.

The contract

Every reply is a validated object

Free-form model output never reaches a customer. Each reply must parse into this shape, and the sources must point at passages Sitora actually retrieved from your site.

{
  "answer": "string, plain text",
  "action": "answer | qualify | book | handoff",
  "sources": [{ "title": "...", "url": "..." }],
  "suggestedReplies": ["..."]
}
Model-agnostic

The contract is the product, not the model

Sitora routes the same grounded prompt across model providers, currently OpenAI and Anthropic, chosen by configuration. Beneath them sits a deterministic extractive engine that answers with verbatim site passages, so a provider outage degrades quality, not honesty.

Provider downExtractive engine answers with verbatim passages, sources intact.
Model uncitesReply rejected, agent hands off instead.
Weak retrievalHandoff before the model is even asked to improvise.
Provider switchA configuration change. The contract, and your agent, stay identical.
Limits, stated plainly

What the agent will not do

  • It will not answer from general knowledge. If your site does not say it, your agent does not say it.
  • It will not invent or estimate prices, availability, hours, or policy, even when a confident guess would probably be right.
  • It will not pretend to be human. It is introduced as the site's assistant, and it hands off to real people.
  • It will not read pages behind logins, run site scripts, or submit forms while reading.
  • It will not keep data you delete. Removing an agent removes its ingested content, conversations, and leads.