Train on websites and documents

What the agent actually reads

Every channel - web chat, voice, WhatsApp, email - pulls from the same knowledge base. Train once, reuse everywhere. That sounds convenient until one stale fee sheet poisons every channel at once. Treat knowledge like production data, not a one-time onboarding checkbox.

You have two main feeds: website crawl and file/cloud uploads. Use the site for public facts (services, hours, bios, FAQ). Use uploads for stuff you won’t put on the marketing site (intake scripts, internal fee tables, “never book X without a deposit” rules).

Website crawl

Go to Knowledge → Websites and paste the canonical https URL customers type - not a preview link, not a localhost tunnel, not a WordPress staging clone with “TEST” in the title.

We follow links from that seed up to your plan’s crawl cap (commonly 30, 100, or 500 pages). Tag archives, calendar plugins, and PDF mountains burn that budget fast. Prefer a tidy set of service pages and a real FAQ. If your blog has 400 posts about industry news, don’t start the crawl from the blog index.

When the job finishes, open a couple of stored pages in the UI. If the copy is wrong on the live site, don’t blame the model. Fix the page, re-crawl, then re-test the same question in chat.

Files and cloud links

Upload PDFs or DOCX for material that shouldn’t be public. Name files like a human (“intake-script-2026.pdf”), not “final_final_v7.” You’ll thank yourself when something goes stale.

Drive, Dropbox, and OneDrive imports need a link that downloads without a login. “Anyone in the org” links fail silently from our servers because we are not in your org. Use a public share or a downloadable link your IT team already vetted.

Watch extraction status. A scanned PDF with no OCR text trains nothing useful - the agent can’t invent words that aren’t there. If status fails, re-export with selectable text or paste the critical rules into a short DOCX.

When marketing changes Friday afternoon

Retrain after price or service changes. Prompt tweaks will not override a document that still says $99. The model is obedient to the knowledge you gave it; that’s a feature until the knowledge is wrong.

Use Conversations → Text chats as your punch list. When three people ask the same thing and get a shrug, that is a missing page or a missing PDF - not a “smarter model” problem. Update the source, retrain, ask the question again yourself before you close the ticket.

Keep a short “truth list” for the front desk: parking, cancellation window, insurance accepted or not, languages spoken. If those four change, update knowledge the same day you change the website footer.

Related guides

• Go-live checklist for first bookings - The go-live order we recommend when a new org wants chat live today and phone soon after - without skipping calendar truth or hours.
• Embed the chat widget - Snippet install, mobile checks, CSP traps, and how to tell booking failures apart from chat widget failures before you blame the model.
• Set up inbound voice - Get a business phone number, short greeting, handoff rules, and the after-hours test call you should not skip.
• WhatsApp setup without carrier risk - Enable WhatsApp the way your market already messages, stay inside platform rules, and keep quiet hours honest.