Skip to main content
The knowledge base is the material an agent quotes from. Without it, an agent has only your business description and its general training — it will not know your prices, your policies or your hours, and it is instructed not to guess them. The library belongs to the workspace, not to an agent. One upload, one indexing cost, one reindex, and every agent can read it. Each agent then chooses how much of it to use. Only owners and admins can add, edit or delete documents. Everyone else sees the library read-only.

​
Three ways to add something

Knowledge → Crawl a page, paste a URL, press Add page.One page, not the whole site — add your pricing page and your delivery page as two documents. The page must be public: the crawler is not signed in, and a page that renders entirely in JavaScript comes back empty. Private, internal and loopback addresses are refused, and a redirect chain is followed at most three hops with every hop re-checked.A crawled document keeps its URL, and the refresh button on its row re-reads the page. Editing a crawled document turns it into your own text, and a later re-read will not overwrite it.
One topic per document retrieves more accurately than one long page. Long documents are split into chunks automatically — at sentence ends, with a little overlap so a fact is never cut in half — and a reply is built from the handful of chunks that match the question, up to five per reply. A document’s title is part of what its chunks are matched on, so a clear title (“Shipping and delivery”) helps.

​
How it is searched

Two kinds of search, and the card at the top of the tab says which one is running:
  • Meaning-based search on — documents are embedded, so a paraphrased question still finds the right passage. On built-in AI this works with no setup of your own; on your own provider key it needs an embeddings key saved under Provider & credits.
  • Keyword search only — passages are found by matching words. Everything still works; it is just less forgiving of the words a customer chooses.
Both searches run on every question and their results are combined: a passage found by both ranks highest, and a passage that only one of them found still counts — so an exact word (a product code, a place, a policy name) is found even when the meaning search ranks something vaguer. Passages that are not related to the question at all are left out rather than sent to the agent. When a customer’s message is a short follow-up such as “and for kids?”, their previous message is searched with it.

​
Document status

Every document carries a status, because a document that saved but is invisible to the agent is the worst outcome here. Reindex all re-processes every document with the current settings. Use it after changing the embeddings provider, or after any document turns amber.

​
Which documents this agent reads

The control at the top of the tab has two positions:
  • Everything — the agent reads every document in the workspace, including ones added later.
  • Only these — tick the documents on the rows below and press save. New documents are not added automatically.
Selecting nothing means exactly that: the agent answers with no documents at all. That is why an empty selection is stored as a real choice rather than read as “all of them” — unticking your last document must not silently re-grant the whole library. Switching from Everything to Only these starts from the full list, so you narrow down from what you were looking at rather than from nothing.

​
Editing and removing

Click a document to open it and edit its title and body; saving re-indexes it. The bin icon deletes it — the agent stops using it to answer immediately, on every agent that had it selected.

​
What it costs

Indexing is metered only on built-in AI credits, at roughly a hundredth of what generating a reply costs — one credit per 25,000 tokens, with a minimum of one per document. On your own provider key nothing is metered. See Credits and your own key.