Improve chatbot accuracy from the source to the final answer.

Pick an answer that matters to the business and identify the passage that should support it. Confirm that the source is public, readable, current, and included in the site’s knowledge. Check whether other pages contradict it.

A missing exception or an outdated price can lead to a wrong answer even when the assistant accurately repeats retrieved text. Decide which page is authoritative and resolve inconsistencies there. Review confirmed facts and saved answers as well, since they can outlive an edit to the original page.

Sairo checks saved answers and retrieves site facts and page passages using keyword and semantic matching. Generated replies are instructed to answer from that context. Source exclusions, topic restrictions and website boundaries still apply; finding similar words does not make a passage appropriate evidence.

Answer checks look for unsupported claims and incomplete basic service replies. A service answer should retain its published description and available starting-price conditions. One bounded source-selection attempt can repair a weak reply; a limited source-based fallback is available for eligible basic service questions. If the permitted material cannot support the answer, the question remains unanswered.

Multipart questions receive a coverage check as well as factual checks. The supported parts can be returned with an explicit acknowledgement of what is missing. These are safeguards for known failure patterns, not proof that every date, policy, interpretation or future reply is correct.

An incomplete answer can have several causes. Check the source, the retrieved evidence and the final wording before changing a model or writing another saved answer. A correct service name with no explanation or available price is a completeness failure, even when the name itself is supported.

Build a small test set from actual visitor needs. Include direct questions, natural paraphrases, questions with an extra condition, and plausible questions whose answer is absent. Write down the expected behavior before running the test.

Test the boundary around a saved answer. A response approved for one service, location, or customer type should not be treated as a universal rule for a nearby but different question.

Save the question, the response, the source you expected, and the specific unsupported statement. Distinguish a retrieval failure from a content problem or an overbroad interpretation. Each points to a different intervention.

A response can be partly correct and still fail the test. Check conditions and scope, not just whether it includes the right phone number or a relevant citation. A linked page must actually support the claim being made.

Improve or exclude the source, update a confirmed fact, or write an approved answer. Ask the original question again after the source refresh finishes, then test nearby questions that should produce a different result.

If you change the model, repeat the regression set rather than assuming a more expensive model is always a better fit. Keep the source, configuration, and review criteria consistent enough to understand what changed.

For a specific Sairo reply, use the inbox’s “Review with Ask Sairo” action. Supply the confirmed answer, inspect the proposed change and check the original and nearby questions. This gives the correction a reviewable record without replacing source maintenance.

Some questions cannot be answered from a public website because they require an individual decision. Treat human handoff as a deliberate outcome: use the site’s live chat queue when enabled, or provide a usable follow-up route with a team that knows how to respond.

Continue sampling real conversations after launch. A test set provides a repeatable check for known situations; it cannot cover every future question, content change, or visitor expectation.