> ## Documentation Index > Fetch the complete documentation index at: https://docs.voicy.co/llms.txt > Use this file to discover all available pages before exploring further. # Knowledge base > A knowledge base gives your agent information about your business, so it can answer callers' questions. There are two types: a Markdown file and a website. Jump to: [Markdown file](/guides/knowledge-base#markdown-file) · [Website](/guides/knowledge-base#website) ## Which type to choose | | Markdown file | Website | | - | - | - | | Where the information comes from | A file you upload | The pages of your website | | Size | Up to 15,000 tokens | Your whole site, hundreds of pages | | Keeping it current | Upload a new file (**Replace**) | Automatic every night, or **Re-crawl** | | How the agent uses it | The whole file is in front of the agent all the time | The agent looks up the relevant pages when a caller asks | | Best for | Short, fixed information that must always be available | A large site, or information that changes often | ## Markdown file Upload a Markdown file about your business, and your agent has its whole content in front of it for the entire call. ### What it is A Markdown knowledge base is a text file you write and upload yourself. Use it for opening hours, branches, prices, policies and frequently asked questions. The whole file is part of the agent's context on every turn, so the agent can answer from it straight away, with no search step. It works best for short, stable information that must always be available. For a large site or information that changes often, use a [website knowledge base](/guides/knowledge-base#website) instead. You can attach both to the same agent. ### Create a Markdown knowledge base Go to **Knowledge Bases** and click **+ Create KB**. Set **Type** to **Markdown**. Give it a name (for example, "Hours and locations") and a short description of what it contains. Pick a `.md` file (`.markdown` and `.txt` are accepted too). The file can be up to **15,000 tokens**, which is about a medium-length document. ### Update the content On the knowledge base's page, click **Replace** and upload the new version of the file. Each call reads whatever the knowledge base holds when that call starts, so your change reaches the next call. You don't need to redeploy the agents that use it. ### The knowledge base page | Control | What it does | | - | - | | **Rendered** / **Raw** | **Rendered** shows the formatted Markdown. **Raw** shows, character for character, the exact block the agent receives. | | Direction toggle | Switches the preview between right-to-left and left-to-right. | | Token count | Shows how much of the 15,000-token limit the file uses. | | **Copy** / **Replace** / **Delete** | Copy the content, upload a new file, or delete the knowledge base. Deleting also removes it from every agent it's attached to. | | **Properties** | The description, plus a list of the agents that use this knowledge base. | | **Test Call** | Ask the knowledge base a question the way a caller would. See Testing a knowledge base. | ### Tips * Use clear headings and short sections. The agent reads the file as written. * The agent answers factual questions only from its sources. It won't guess from general knowledge. If callers ask for something the file doesn't cover, add it to the file. * Check **Raw** view after you upload, to see exactly what the agent sees. ## Website Connect your business's website, and your agent answers callers' questions by looking up the relevant pages during the call. ### How it works Voicy reads your site's pages and keeps a stored copy up to date. When a caller asks about the business, the agent searches that copy, reads the most relevant pages, and answers from them. Nothing is fetched from your site during the call. ### Create a website knowledge base In **Knowledge Bases**, click **+ Create KB**, set **Type** to **Website**, and enter your domain, for example `example.co.il`. Each domain can be added once per account. The first crawl starts on its own. The status shows **Crawling · A of L** with an estimate of the time left, and the **Pages** table fills in as pages are read. The agent can answer from a page as soon as it's stored. **Only pages listed in your sitemap are read.** Voicy doesn't follow links around your site. A page missing from it (a hidden page, a new page not yet added) won't be read, and the agent can't answer from it. Add such pages under **Extra URLs** in the knowledge base's properties, one full URL or path per line, up to **500**. They're read on every crawl from then on. Pages that aren't web pages, such as PDF price lists and images, aren't read. ### Keeping it up to date * **Schedule:** your site is re-crawled automatically at night, between 01:00 and 06:00 Israel time. * **Scan notes:** if your sitemap dates its pages reliably, Voicy checks only changed pages each night and does a full scan weekly ("Delta scan daily. Full scan weekly."). A red **(!)** means the dates are missing or unreliable, so the site gets a full scan every night. This isn't an error. * **Re-crawl:** click **Re-crawl** after you change your site, for example after a price update. It reads every page. The button then shows **Queued** and **Crawling…**. After a run it's disabled for about 5 minutes, with a countdown. During a re-crawl, agents keep answering from the current pages. If a crawl fails, the page gives the reason, for example "the site could not be reached" or "no pages were found in the sitemap". ### The Pages table This table shows every page read, as a tree by URL path, with title, tokens, status and last update. A page that couldn't be read is marked in amber (**not read**), and folders holding one open automatically. Click a title to open the live page. ### Lookup results in the transcript In a Test Call, each lookup appears as a system row. Expand it to see the question searched, the answer, the pages read (and which were cited), and the time taken. | Row | Meaning | | - | - | | **נמצא** (found) | The answer came from a page on your site. | | **לא מופיע באתר** (not on site) | The pages read don't contain it. Either the information is missing from the site, or it's on a page the search didn't reach. | | **נדרש פירוט** (needs detail) | The site gives several answers (for example, a different price per branch), so the agent asks the caller which one they mean. | | **נכשל** (failed) | The lookup couldn't run. The most common cause is that no pages are stored yet because the first crawl hasn't finished or failed. | When a website is attached, the agent looks up every question about the business. If the answer isn't on the site, the agent says it doesn't have that information instead of guessing. For anything the agent must always know, use the prompt or a [Markdown knowledge base](/guides/knowledge-base#markdown-file). ## Attach a knowledge base to an agent Open the agent. In its knowledge base section, click **+ Add Knowledge Base** and choose the knowledge base. The agent card shows the knowledge base's name, description and token count. To detach it, click **✕** on its card. An agent can have any number of Markdown files and **one website**. All Markdown knowledge bases attached to one agent share a single limit: together they can hold up to **15,000 tokens**. If an attachment would go over that, the save is refused. Attaching a knowledge base changes the agent, so deploy the agent before you expect live calls to use it. See Draft and deploy. ## Test it Ask the knowledge base a question the way a caller would, and see where the answer came from. See Test a knowledge base. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.