Knowledge bases and sources

What a knowledge base is

A knowledge base holds the documents one assistant answers from. Keep separate audiences in separate knowledge bases: a public help assistant and an internal one should not share content.

How a document becomes an answer

1. The file or page is fetched and converted to text.
2. The text is split into passages that overlap slightly, so a sentence is never cut in half between two of them.
3. Each passage gets an embedding, which is what makes "how do I cancel" find a passage titled "ending your subscription".
4. At question time we search both by meaning and by keyword, merge the two rankings, and give the best passages to the model along with your instructions.

A document that says failed could not be read: a scanned PDF with no text layer, a password protected file, or a page that returned an error. The status tells you which.

Keeping content fresh

• Crawled sites re-crawl on a schedule you choose, and only re-index pages that actually changed.
• Connected sources (Webflow, WordPress) update within seconds of a change through a webhook.
• Uploads are yours to replace; upload the new version and the old passages are removed.

When an answer is wrong

Open the conversation and look at the citations under the answer.

• No citations, and the assistant said it did not know. The content is missing. Add it.
• Citations from the wrong document. Two documents look similar to the model. Make the headings more specific, or split the document.
• Right passage, wrong answer. Adjust the assistant's instructions in the knowledge base settings; that is where tone and rules live.

The Analytics → Questions tab collects the questions that got no answer, so you do not have to find them one conversation at a time.