> ## Documentation Index > Fetch the complete documentation index at: https://wiredesk.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # The agent says it does not know > Tell a real knowledge gap from a retrieval miss, and fix each: check the source, test the question, add a note, or hand over to a person. When the agent says it does not know, it is often right. "Our hours are not listed in what I have here" is an honest answer when your website never lists its hours, and no setting fixes a fact nobody wrote down. The agent is told to answer from what it finds in your knowledge and to say when it has nothing, rather than make something up. So start by finding out which case you are in. Either the fact is missing, or the fact is there and the agent did not find it. The fixes are different. Open **Agents → your agent → Knowledge** and type the customer's question into **Ask a question it should know**, then choose **Search**. You see the passages the agent would get, each with its source and a match score. If you get "No passages matched — the agent would say it's not sure and offer to take a message", the agent never saw the fact. See [Test a question](/docs/knowledge/test-a-question). Only ready sources are searched. A website source reads "crawling the site — this can take a minute or two…" and a file or note reads "processing…" until it is done. After that, the source shows how many passages it holds. If a source shows an error instead, it was not read. A freshly built agent whose sources are still processing will seem to know nothing about a site it just crawled. Open the page in a private window. The crawler sees what an anonymous visitor sees, so it cannot read anything behind a login, a password or a cookie wall. The same goes for text that is only inside an image. On a website source, open the list of pages and search for the page. If it was excluded, or the crawl never reached it, add its address with **Add website** or put the fact in a note. Under **How it talks**, **Knowledge — answer from your website & documents** must be ticked. With it off, the agent does not search your knowledge at all. ## The fact is missing If the fact is nowhere in your knowledge, you have a content gap, and the agent behaved correctly. Choose **Paste text or notes** under **Knowledge** and write the fact as the answer to the question. Keep it to a sentence or two, in the words a customer would use, then choose **Add text**. See [Notes](/docs/knowledge/notes). The same gap probably exists on your website too, because every customer who did not ask the agent had the same problem. ## The fact is there, but the agent misses it This is a real retrieval miss, and it has a typical cause. The sentence you want is buried in a page that is mostly navigation, header and footer, so the passage that matches best is the boilerplate, not the answer. A long policy page with one relevant sentence loses to a short page that is entirely about the topic. The reliable fix is still a note. Short, specific notes are found far more reliably than long documents. Then ask the question again, first in **Ask a question it should know** and then in **Test chat**. If a note that plainly states the fact still loses, [tell us](https://wiredesk.ai/contact). That is a retrieval problem for us to fix in the product, not something to fix in your settings. ## It knew this yesterday Knowledge is a snapshot, not a live copy of your site. If a page changed or a source was removed, the agent keeps working from what it last read. Re-read the site after a meaningful change. See [Keep it current](/docs/knowledge/keep-it-current). ## Do not tell it to always answer You could write "never say you don't know" into the instructions, but don't. An agent told to always have an answer will invent one, and an invented price or policy reaches a real customer with complete confidence. Put the facts in knowledge instead. Instructions are for how the agent behaves, not for facts. ## When it should not answer at all Some questions should go to a person, not get a better answer: an account-specific dispute, a medical or legal judgement, anything where a confident wrong answer is expensive. For those, switch on **Request a human** under **Actions** for chat and email, or **Hand calls to you** for the phone, with a real person behind it. An agent that says it does not know and offers nothing has failed. An agent that says it does not know and hands the customer to someone who does is working as intended. See [Handover](/docs/concepts/handover). ## Find these before customers do Read your conversations. Every "I don't have that here" in a week of transcripts is a note waiting to be written. ## Related * [Fix a wrong answer](/docs/knowledge/fix-a-wrong-answer) * [Test a question](/docs/knowledge/test-a-question) * [Notes](/docs/knowledge/notes) * [How answers work](/docs/concepts/how-answers-work)