Software & SaaS · website AI assistant
AI chatbot for Docs & knowledge bases
Updated
The short version
Replace keyword search on your docs or help center with an assistant that answers the question and links the passage it used.
Documentation search is usually the weakest part of good documentation. A reader types a question and gets ten blue links ranked by keyword overlap, none of which is obviously the right one. The information was there; the retrieval wasn't.
An assistant over the same content changes the interaction from *find the page* to *answer the question, then show me the page*. That second half is what makes it safe for technical content — the reader gets a direct answer and a citation to verify it against, which is exactly how a good engineer reads documentation anyway.
MySiteGPT chunks your docs, retrieves the passages that actually match the question, and answers from those passages with visible source cards. When nothing matches well enough, it says so — which is far more useful to a reader than a confidently wrong config example.
What documentation sites get asked
Real questions, in visitors' own words. Every one of these is answerable from content you probably already publish.
- “How do I do X — and which page explains it properly?”
- “What does this error message mean?”
- “Is there an example of this in a language I use?”
- “What's the difference between these two settings?”
- “Has this behavior changed recently?”
- “Where do I configure this — in the file or the dashboard?”
How MySiteGPT helps documentation sites
Answers, then the passage
Readers get a direct answer plus an openable source card, so verifying takes one click instead of a search session.
Finds what your search box misses
Retrieval works on meaning as well as keywords, so a question phrased in the user's vocabulary still lands on the page written in yours.
Says 'not documented' out loud
An honest fallback beats a plausible invented flag. Every fallback is logged as an unanswered question — a documentation gap list ranked by real demand.
Stays current with your docs
Sources re-sync on a daily schedule you choose, so a merged docs PR is reflected without anyone re-training anything by hand.
What to train it on, in order
The answers are only as good as the index behind them. Start at the top — these are ordered by how much question volume they absorb per page trained.
- 1
The full docs site, crawled from its sitemap
Breadth matters more than curation here — the long tail is the whole point. Check the sitemap first so you know the page count going in.
- 2
Troubleshooting and error-message pages
The highest-intent content you have. Someone searching an error string is stuck right now.
- 3
Tutorials and quickstarts
Give the assistant a worked example to cite rather than making it assemble one from reference pages.
- 4
Anything still trapped in PDFs
Spec sheets and legacy manuals upload directly and become answerable alongside the web docs.
Not sure your pages are ready to be retrieved from? Run one through the RAG readiness checker first, or audit the whole site with the sitemap analyzer.
What to measure once it's live
We don't publish other people's numbers as if they were yours. Track these instead — they are all visible in your own dashboard within a fortnight.
- Searches that end without a click, before vs after
- Support tickets whose answer already existed in the docs
- Unanswered questions grouped by section — your writing queue
- Repeat questions after a docs fix ships (they should stop)
AI chatbot for Docs & knowledge bases: frequently asked questions
How is this different from the search box we already have?
Keyword search returns a ranked list of pages and leaves the synthesis to the reader. This returns the answer, in their phrasing, with the source passage attached — and admits when the docs don't cover it, which a search box never does.
Will it invent API parameters?
It answers only from retrieved passages of your content, and shows the source for each claim. When retrieval confidence is low it returns your fallback instead of composing something plausible. That's the single most important property for technical docs.
How much documentation can it hold?
The Pro plan indexes 1,000 trained pages, which covers most product documentation sites comfortably. The token and chunk estimator will tell you what your docs work out to before you start.
Can it cover private internal docs?
Only content you train it on is used, and only the widget you embed can reach it — but the widget itself is public wherever you place it. For internal-only knowledge, embed it behind your own authentication.
Try it on your own content first.
The chat bubble on our homepage is the product, trained on our own site — ask it anything and watch every answer cite its source. Then train one on yours: 14 days, full features, no card.