Your knowledge base: sources, sync and search
Connect help-center and docs sites, Confluence, Notion, SharePoint/OneDrive, Google Drive, Guru and Document360; keep them in sync; write articles TR-AI can use well.
TR-AI crawls your sources, splits pages into sections (keeping headings, numbered steps, tables and FAQ question-and-answer pairs together) and indexes them for search by meaning and by keyword. The best matching sections — plus the next section of the same page — are what the AI reads before it answers.
- 1Add sourcesDashboard → Knowledge base: add your help-center or docs URL, and connect Confluence or Notion if you use them. Login-protected sites can be crawled with a login or header.
- 2Choose auto-syncWeekly, daily or every 6 hours depending on plan. Only changed pages are re-indexed.
- 3Re-crawl after big changesUse “Re-crawl knowledge base” after publishing many new articles or restructuring your site.
- 4Exclude noiseExclude URL patterns (release notes archives, legal pages, old versions) so they never become answers.
Document and knowledge-tool sources
| Source | What you enter | Notes |
|---|---|---|
| SharePoint / OneDrive | Tenant ID, client ID and secret of a Microsoft Entra app with Sites.Read.All (application) permission; site URL; optional folder | Word, PDF, text, Markdown and HTML files |
| Google Drive | Service-account JSON key and a folder ID (share the folder with the service account email) | Google Docs are exported as text; Word, PDF and text files are read too |
| Guru | Guru user email and user token | All cards the user can see, or one collection |
| Document360 | API token (Settings → API tokens) and data region | Articles of the main project version in the language you choose (default en) |
These sources sync on the same schedule as your other sources; TR-AI checks each document’s modified time first and only re-indexes when something changed. Secrets are stored encrypted and never shown again after saving.
Cleaner crawling
- Text repeated across many pages — “Still need help?” footers, cookie notices — is removed, so every section is about its own topic.
- Sidebars, “Related articles” lists and tables of contents are dropped from page text.
- Short sections keep their own heading instead of being merged into a neighbour.
- A page served under several URLs that declare the same canonical URL is crawled once.
- Search combines meaning with exact keywords (error codes, product names), and the best sections reach the AI in ranked order. On our internal 109-question benchmark the right article ranked first rose from 91.7% to 94.5%.
- Existing sources are re-processed this way on their next crawl or sync — nothing to do on your side.
Writing articles TR-AI answers well from
- One topic per article, with the question in the title (“How do I reset my password?”).
- Use real headings and numbered steps — TR-AI keeps steps together and quotes them in order.
- Name things exactly as customers do, including product names, error messages and codes; keyword search finds exact terms.
- Put the answer in the first paragraph of each section; context and edge cases after.
- Keep FAQs as question-and-answer pairs (accordions are fine — the question and answer are kept together).
- Avoid “see the screenshot” instructions — the text has to stand on its own.
Screenshots customers send
When a customer attaches screenshots (all 8 help desks — Zendesk, Freshdesk, Intercom, HubSpot, Salesforce, Zoho Desk, ServiceNow and Jira Service Management — plus e-mail, SMS/WhatsApp, Messenger and the chat widget), TR-AI reads the visible error messages, codes and screen names — up to 3 images, one AI call per ticket — and searches your knowledge base with them. Turn it off under Settings → Resolution engine (“Read screenshots”).
- Put exact error messages and codes in your articles: a screenshot of “ERR_403” then finds the article that names it.
- Images are read on the fly and never stored by TR-AI. On a follow-up, only images added in the last 15 minutes are read, so old screenshots are not sent again. With “Read screenshots” off, attachments are not even downloaded.
- Salesforce: files on the case and on the customer’s e-mails to it. HubSpot: images in the ticket’s conversation thread. ServiceNow and Jira SM: image attachments on the record.
- Text read from images is treated as the customer’s words, never as instructions, and goes through the same personal-data masking as the ticket.