Upload the documents you already have. Point it at the pages on the help center. Add the answers that absolutely must be exact. The chatbot answers from those and nothing else — and it shows the visitor which document the answer came from.
Most chatbot launches stall because nobody trusts the answers. This is the screen that gets you past that moment — and then keeps you past it as the underlying content changes.
Point the assistant at your documents, websites, and Q&A — it discovers the pages, indexes them, and only answers from what you approved, with usage and limits visible the whole way.
Upload the documents, add the help-center pages, and write the Q&A pairs that absolutely must be exact. The chatbot answers from those and nothing else. Every answer includes a citation showing which document it came from.
No web scraping outside the pages you approved.
Q&A pairs are higher confidence than documents — use them for the answers that have to be word-perfect.
Every answer shows the source. The visitor knows where it came from. So do you.
The policy changed, the pricing changed, the help center got a new section. A prompt does not know. A source does.
Crawl the whole website and you indexed the team blog, the careers page, and a marketing campaign from last year. The chatbot starts answering from things that should not be answers.
Refund policy. Pricing. Escalation language. You cannot let a retrieval pass guess the wording. The Q&A pair is what stops that.
PDF, DOCX, TXT, Markdown — upload what already exists. The platform parses and indexes it so the chatbot can find the right paragraph at the right moment, and cite it back to the visitor.
Point at the help-center URLs, set the depth, set the include/exclude patterns. Crawl only what should be answers. Schedule re-crawls so the chatbot stays current as the pages do.
For refund policy, pricing, escalation, disclaimers — write the exact answer next to the exact question. The chatbot uses these as the highest-confidence source, ahead of any document or page.
Make a list of the top 20 questions visitors are about to ask.
Map each one to the document, page, or Q&A pair that should answer it.
Upload, crawl, or write — then test all 20 in the playground.
Refresh sources when the underlying content changes, on a cadence you set.
Status per source so you know what is live and what needs attention.
Last crawl, next crawl, what changed — stale answers are visible before the visitor finds them.
Document count, site count, snippet count, Q&A count — the breadth of what the chatbot can answer from.
Support teams replacing the manual FAQ that nobody on staff updates anymore.
Public-sector teams answering citizens from policy documents that are already public.
Sales chatbots that stay aligned with the current pricing and packaging, not last quarter's.
Each client keeps their own knowledge — no sharing between accounts.
Only the sources you marked as approved show up in answers.
Stale or failed sources are visible from the same screen as the live ones.
Yes — files in PDF, DOCX, TXT, and Markdown. Plus pages on websites you control or that are public. The chatbot uses what you uploaded; nothing else.
On plans that allow scheduled re-crawls, yes. Otherwise the team runs the re-crawl manually when the pages change. Either way the chatbot tells you when a source is stale.
No. The screen makes the review fast — you test the top 20 questions in the playground before publishing — but the judgment still belongs to your team.
Tell us what you are trying to ship, and we will show you the pieces of the platform that get you there.