Train Atlas on your knowledge base
Atlas drafts answers using whatever you teach it. Here’s how to get the most out of that - what to upload, what to skip, and how to keep it sharp over time.
What Atlas can ingest
You add sources from the dashboard’s knowledge-base config. There are three:
- Single page. Point Atlas at a URL and it will crawl the page (and a few configurable levels deeper). Best for help docs, FAQs, and product pages you already maintain on your site.
- Discord channel. Import the message history from a Discord channel - useful when your support team has already answered the same questions in public, and you want Atlas to learn from those answers.
- GitHub repository. Link a repo and an optional path; Atlas indexes every Markdown file (
.md,.mdx) it finds there. Good for engineering teams who keep their docs in the same place as their code.
Structuring entries
Smaller, focused sources work better than one giant page. A help center with a separate article per topic (“Refund policy”, “Cancellations”, “Region availability”) is much easier for Atlas to pull clean answers from than a single “Everything you need to know” mega-doc. Atlas cites the source it used in its reply, so granular sources also make the citations more useful.
Letting staff teach inline
When staff correct an Atlas draft before sending, the correction is captured as a Q&A entry and added to the knowledge base. It then competes with your other sources on the next similar question - ranked by relevance and by community votes - so good corrections rise on their own over time. No retraining job to run; new entries are searchable immediately.
What not to put in
- Secrets, passwords, API keys, customer data.
- Outdated policies. Atlas treats every entry as current.
- Long-running incident notes. Use canned responses for those instead.
Quality checks
Open the knowledge base list in the dashboard every so often and skim it for outdated entries. If an answer changed (pricing, policy, supported region), edit or remove the entry so Atlas stops citing the stale version. Re-crawl any URL-based entries after big docs changes so the index reflects what your customers will actually find on your site.