Three kinds of knowledge, with different jobs.
Sairo uses indexed pages for the breadth of your website, site facts for recurring details such as opening hours, and saved answers for questions that need approved wording. These sources belong to a particular site; adding a second client does not create one shared pool of business information.
An exact saved question is checked first, after trimming whitespace and ignoring letter case. It can return the saved answer without a model request. Similar questions can also match a saved answer through semantic retrieval. Other questions use relevant facts and passages from the indexed website.
| Information | Where to maintain it |
|---|---|
| A published service explanation | Update the website, then refresh and review its indexed page. |
| A phone number or opening hours | Check the original page and confirm the corresponding site detail. |
| An answer that needs exact wording | Save a question and approved answer, then test relevant variations. |
| An obsolete promotion | Remove or update the original and exclude its indexed source if needed. |
Reading the site is a controlled process.
A production website must pass ownership verification before a crawl can run. The public demo has a separate, limited sampling path. Crawling follows robots.txt, checks fetch destinations, and works within the workspace allowances.
Indexed text is split into smaller passages. Sairo combines keyword matches with semantic similarity to find material relevant to a question. This retrieval step supplies context for the answer; it does not retrain the underlying model on your entire website.
Review the evidence behind important answers.
Start with the pages visitors depend on: services, eligibility, location, contact information, pricing conditions, and policies. An attractive page can still be a poor source if its critical information is inside an image or appears only after a client-side interaction.
Test a question against its intended page. If the reply is wrong, first inspect whether the correct passage was indexed and whether another page conflicts with it. A new prompt or model cannot reliably repair missing or contradictory source material.
You can review this knowledge through Ask Sairo too. Ask about the source behind a reply, provide a confirmed correction and check the affected website before saving.
Teach Sairo a better answer. Save confirmed information and check the next reply.
Before you start: Use an account with permission to edit website knowledge. The workspace, visitors and figures shown are fictional examples.
Read the transcript
- A better answer starts with you.
A visitor asks about Sunday inspections. Here's how to give Sairo the confirmed answer your team wants to use.
- Open the library
Open Knowledge library, then choose Add saved answer. Select the website that this information belongs to.
- Be specific
Enter the exact question and a clear, verified answer. Here, the business confirms that Sunday inspections are unavailable and gives the visitor a useful next step.
- Save the answer
Select Save new answer. Wait for confirmation. Saved answers take priority when a visitor asks a matching question.
- Check coverage
In Check your knowledge, enter the same visitor question and select Check source coverage. This checks saved knowledge; it does not generate a live reply.
- Read the result
Look for A saved answer is ready. Read the returned wording and confirm that it answers the question completely.
- Try the next reply
Now ask the question in the website assistant. The visitor receives the saved answer, including the next step for arranging an inspection.
- Teach once. Help the next visitor.
Repeat this whenever a policy changes or a visitor finds a gap. Keep the wording specific and confirmed.
Example: changing weekend hours
Update the public contact page, refresh the website content, and check any confirmed hours or saved answers that still contain the old schedule. Ask both “When are you open?” and “Are you open on Sunday?” before considering the change complete.
Not every answer lives on a page.
A price list, a service menu or a policy is often a document rather than a web page. Upload a PDF, Word file, text, Markdown or CSV, up to 12 MB, and it is chunked, embedded and retrieved exactly like a crawled page. It appears in the library with the same source controls, and counts against your plan’s indexed-page allowance.
An uploaded document has no public web address, so a citation refers to it by name rather than linking a visitor to a page that does not exist. A scanned PDF with no text layer cannot be read; export a text-based version first.
A published help centre article is knowledge too. Writing one does both jobs at once: visitors can read the page, and the assistant can answer from it without you maintaining the same wording twice.
Assign an owner to the knowledge.
Agree who can approve service facts and who can change assistant configuration. A client who knows the business may contribute content within an assigned project, while the agency keeps broader controls.
Review unanswered questions regularly and treat recurring gaps as editorial work. Keep the public website useful in its own right: visitors should still be able to read the source without opening a chat.
Common questions
Does adding a website train a new AI model?
Sairo indexes and retrieves the website’s content. It supplies relevant material to the configured model during a response; this workflow is different from fine-tuning model weights.
Does excluding a page delete a confirmed fact?
Source exclusion affects retrieved page content. Explicitly confirmed facts and saved answers are separate knowledge records, so review them when a business detail changes.