📥 Feeding It Your Own Documentation, And The Search That Answers

Six ways to add what you know:

Open the Knowledge Base tab. Under Teach The Desk What You Know there are six tabs, and you can use all of them.

• Upload files. Drop in as many as you like at once: Word documents (.docx), text, Markdown, saved web pages (.html), spreadsheets saved as .csv, and one JSON file that can cover every product you sell (see Many Products? Teach The Desk Everything With One JSON File), up to 20 MB each. Got a PDF? Save it as Word or plain text first.
• A web address. Paste the address of your FAQ page, your refund policy, or a documentation file that lives online. Leave Read it again every day ticked and the desk keeps itself current. If the address stops answering, the last good version stays in use and the row tells you why. Some sites refuse automated readers; the desk says so, and you can save the page and upload it instead.
• Your blog. Paste the blog address and the desk finds the feed it already publishes, tells you the address it found before it reads anything, then reads every post, following the archive back as far as the blog will serve it. It reads again every day, so a post you write in the morning can be answering customers by the afternoon. Each post keeps its own web address, so an answer sends someone straight to the post. Any blog with a feed works.
• Paste text. Give it a name and paste. The answers you type to customers again and again are the best thing to put here.
• Pages on this site. Tick Pages, Posts or any other content type and the desk reads what you have already written, where it is. Edit a page and the desk catches up within a couple of minutes. Password-protected pages, and the desk's own form and ticket pages, are never read.
• Your plugins. Every ResultsAiX plugin on your site carries its own readme and its own Training screen, written in plain English for the person using it. Tick this and the desk reads them where they sit, keeps every answer under the name of the plugin it came from, and reads them again every day, so a plugin update brings its new wording into your answers with it. Only plugins that say they are ours are read, never another author's, and release notes are left out. Untick it and everything it added goes with it.

Most of your real answers are already written down somewhere that is not a Knowledge Base article. A licensing FAQ in Markdown. An install guide exported as HTML. A product list as JSON. A spreadsheet of questions as CSV. Upload those files and they join your articles in one index.

That index then does three jobs at once. It answers visitors on your knowledge base page. It suggests articles while somebody is typing a ticket. It grounds what the AI replies to a ticket. You write it once and it works in all three places.

Adding a file:

1. Open the Knowledge Base tab. The Knowledge Sources panel sits at the top.
2. Choose a file and click Add To Knowledge. Markdown, plain text, HTML, CSV and JSON are read, up to 20 MB each.
3. It is read and indexed on the spot, and the table tells you how many searchable pieces it added.

Uploading a file with the same name again replaces it. That matters more than it sounds: an updated FAQ should never leave its old wording in the index arguing with the new one.

How each kind of file is read:

• Markdown is split at its headings, because a heading is you already saying where one answer stops and the next begins. Code fences and link clutter are stripped so matching sees words rather than punctuation.
• Plain text is taken whole and split into pieces of roughly four hundred words.
• HTML has its scripts, styles, navigation and footers dropped, and the readable text kept.
• CSV and TSV become one piece per row, written as "Column: value", so a search for either the column name or the value finds it.
• JSON becomes one piece per record, with nested fields flattened into readable paths. It finds the list inside whatever wrapper the file happens to use.

Members only:

Every uploaded file, web address, blog and pasted text has a Members Only button. Press it and only the Members Helper in your members area reads that source; support answers, the sales chats and your help pages never see it. That is how you give the helper your training library without putting your lessons on a public page. See The Members Helper As A Coach, And Members Only Knowledge.

Pausing and removing:

Pause takes a source out of the search without deleting it, which is the quick way to find out whether one file is dragging your answers off course. Remove deletes the source and everything it added to the index.

The search that answers:

On your knowledge base page a visitor types a question in their own words. The desk finds the passages across your whole index that best match it, and the AI writes the answer from those passages only, listing the sources underneath so nothing has to be taken on trust. Under that sits a way through to a human, and it carries the question with it, so nobody has to type the same thing twice.

If your documentation does not cover the question, it says so and shows the closest articles instead of inventing an answer. That is deliberate. A help page that makes something up is worse than one that says ask us.

It works with no API key:

With no Claude key saved, the search box still searches your whole index and lists what matches, which is all a search box was ever meant to do. The AI only upgrades the answer. It is never the thing that makes search work.

Keeping the bill sane:

That page is public, so every answer spends your own Claude credit on somebody you have never met. Under Public Answers you can switch the AI answers off, and the page then still searches and lists the matching articles. Two caps hold it down, both on the same screen. Answers per visitor per hour, ten by default. Answers per day for the whole site, three hundred by default. Set either to zero to remove that limit. Your monthly budget in Settings, AI still applies on top of both. A visitor who hits a cap still gets the matching articles rather than a dead end, and the desk declines to spend a call at all when nothing in your documentation is a decent match.

Getting better answers:

• Put the real question in the heading, in the words a customer would actually use.
• Keep one topic per heading or article. Long catch-all pages match everything weakly and answer nothing well.
• Read the sources listed under an answer. If the wrong document keeps being cited, that document is the one to fix.
• Watch what visitors ask and fail to get answered. Each one is an article worth writing.