# Adding knowledge

What the agent can answer from: crawled pages, uploaded documents, and the authored answers that do the real work.

Source: https://docs.omazy.ai/agent/knowledge/

Knowledge is what the agent is allowed to know. Everything it says should trace
back to something here, and an answer that cannot be traced is an answer waiting
to be wrong.

## Three ways in, in ascending order of value

**Crawl your site.** Fastest. Gets a working agent today. It captures what you
have already published, which is a lot, and it captures it in marketing voice,
which is not always what a question needs.

**Upload documents.** Good for the things that exist as files: policies, price
lists, manuals.

**Author answers.** Slowest, and the one that actually makes an agent good. Every
business has questions its website has never answered plainly. Delivery on
public holidays. What happens if the size is wrong. Whether you take that one
payment method everyone asks about.

A dozen authored answers usually outperform a thousand crawled pages, because
they are the dozen people ask.

## Keeping it honest

Knowledge goes stale quietly. Nobody files a ticket saying "your agent is citing
last year's returns policy", they just get a wrong answer and leave.

Two habits prevent most of it:

- Re-crawl when the site changes, not on a calendar.
- When a human corrects the agent in the inbox, write that correction into
  knowledge. Otherwise you will correct it again next week.

The second one is the whole game. An inbox is a list of things your agent did
not know, delivered free, sorted by how often people ask.

## Sources

An answer can carry its sources, and the widget will show them. Citations cost
a line and buy trust, and they let a customer check the thing themselves rather
than escalating to ask whether the agent was right.

## When an answer must be exact

Some answers cannot be paraphrased: a legal line, a refund policy, a regulated
disclosure. Those belong in saved responses, which return the text you wrote,
word for word, rather than the model's summary of it.