Core Concepts

Knowledge sources

Where the agent looks for answers. Add the right sources and the bot stops guessing.

What the agent draws on

On every message, the agent pulls from, roughly in this order of authority:

Your operating manual — the plain-English instructions you give the agent (tone, policies, what to escalate). Set it under Knowledge.
Your docs & uploaded content — everything you've indexed (crawled pages, files, pasted markdown).
Past resolved conversations — successful answers get folded back in for similar future questions.
Live tool data — for account-specific questions, the agent can read your connected tools.
The model's general knowledge — used only when nothing above applies, and the agent will say when it's unsure.

Adding a docs site

From dashboard → Knowledge → Docs, paste your docs URL.
Pick how to ingest it: a single URL, a sitemap, or a deep crawl that follows links.
Click crawl. A source indexes up to 200 pages; the initial crawl typically completes in a few minutes.

Re-indexing

When your docs change, hit Resyncon the source row to re-crawl and re-embed it. There's no automatic schedule today — resync is on demand, so trigger it after a docs update.

Uploading files & markdown

Besides crawling, you can upload files or paste markdowndirectly. Use this for internal content that isn't (and shouldn't be) public: troubleshooting runbooks, escalation notes, "known broken" lists, or anything behind auth that the crawler can't reach.

Don't paste secrets

Uploaded content is searchable by the agent and can be quoted in replies. Don't put API keys, customer data, or anything you wouldn't want shown back to a user.

How retrieval actually works

Each source is chunked, embedded, and stored in a per-org vector index. On every message, the agent retrieves the most relevant chunks across all your sources and includes the top-ranked ones in the prompt. The model is instructed to ground its answer in those sources and to say "I don't know" (and escalate) when no chunk is sufficient.

What if the agent ignores a source

Usually the chunk wasn't retrieved — the source isn't indexed yet, the question doesn't lexically match, or the page was behind auth. Open the conversation in History: each reply lists the sources the agent retrieved, with a match score, so you can see what it actually considered.