> ## 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.
# Knowledge overview
> The website pages, files and notes your agent searches before it answers, and how they get indexed.
Knowledge is the set of website pages, files and notes your agent searches while it talks. When a customer asks about prices, hours, services or a policy, the agent searches this set first and answers from what it finds. If nothing comes back, it says it doesn't have that detail to hand and offers to take a message. It won't make one up.
Open it at **Agents → your agent → Knowledge**. Every channel the agent answers on (website chat, phone and email) searches the same knowledge.
Facts belong in knowledge, not in the agent's instructions. Your agent treats knowledge as more reliable than instructions: if the two disagree on a price, an hour, an address or a policy, it gives the knowledge answer. A fact typed into the instructions has no source, so you can only fix it by editing the instructions again. See [Writing instructions](/docs/persona/instructions).
## Source types
There are three ways to add knowledge, all at the top of the **Knowledge** page.
| Source | How you add it | Best for |
| ---------------- | ------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- |
| **Website** | Type your address and click **Add website**. | Everything your site already says. The crawler reads up to 120 pages. |
| **Notes & text** | Click **Paste text or notes**, paste, then **Add text**. | Facts your website doesn't state clearly: the hours you actually keep, return rules, answers to common questions. |
| **Files** | Click **Upload a file**. PDF, Word, Excel, CSV, HTML or plain text, up to 8 MB. | Price lists, policy documents and spreadsheets that aren't on your site. |
If you entered your website while setting up the agent, it's already in the list as your first source.
Knowledge sources are unlimited on every plan. To add or remove sources you need the Owner or Admin role.
What the crawl reads, its limits, and its error messages.
Filling the gaps your site leaves.
Supported types and size limits.
## How a source is indexed
Whatever you add, WireDesk turns it into plain text and splits it into short **passages** of about 1,200 characters each. Each passage overlaps the next a little, so a sentence at a boundary isn't lost. The passages are then indexed for search by meaning, not by exact keywords.
When the agent needs a fact, it searches with a short phrase and gets back the five closest passages. Passages that are only a weak match are dropped. That's why a short passage that answers a question directly usually beats a long page that mentions it once.
* **Notes** are indexed as soon as you click **Add text**.
* **Websites and files** are indexed in the background. The page checks every few seconds, so you don't need to reload. A website usually takes a minute or two.
## Statuses
Each source in the list is in one of three states.
| State | What the list shows | What it means |
| ----------- | -------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
| In progress | "crawling the site — this can take a minute or two…" for a website, "processing…" for a file | Still being read and indexed. |
| Ready | For example "38 pages · 212 passages — click to see which pages" | Indexed. The agent can answer from it. |
| Failed | The reason, in red | Nothing new was indexed. See [website errors](/docs/knowledge/website#errors-and-what-they-mean) and [file errors](/docs/knowledge/files#errors). |
If a source is still in progress after 30 minutes, WireDesk treats the job as interrupted and shows an error, such as "indexing was interrupted — remove this source and add it again". A website may be retried automatically first, up to three attempts.
The agent's setup checklist shows the same thing on the agent's **Overview**. The **Knowledge indexed** item reads "Indexing now — a minute or two" while sources are in progress, and "Answering from N sources" once they're ready.
## The Sources list
Below the add buttons is the list of everything the agent has read, newest first. Each row shows the source's name and its state:
* A **website** shows the address you entered.
* A **file** shows its file name.
* A **note** shows "Pasted text".
Click a ready source to look inside it:
* **Websites** open as a list of the pages that were indexed, grouped by the first part of their path, with each page's size. You can search the list, sort it by size, and untick pages to leave them out. **Show the passages themselves** switches to the raw passages. See [Keep it current](/docs/knowledge/keep-it-current#exclude-pages-from-a-website).
* **Notes and files** open straight to their passages, numbered in order.
To remove a source, click the bin icon on its row and confirm with **Remove**. Its passages come out of the knowledge base straight away, and the agent stops answering from them.
When at least one source is ready, a **Test a question** box appears above the list. Use it to see exactly which passages a question finds. See [Test a question](/docs/knowledge/test-a-question).
## Turning knowledge off
Whether the agent searches knowledge at all is a setting under **How it talks**: **Knowledge — answer from your website & documents**. It's on for new agents. If you turn it off, the agent answers only from its instructions and actions, even if sources are listed here. See [Remember callers and answer from knowledge](/docs/persona/remember-callers).
## Related
* [Train the agent on your website](/docs/knowledge/website)
* [Test a question](/docs/knowledge/test-a-question)
* [Fix a wrong answer](/docs/knowledge/fix-a-wrong-answer)
* [How answers work](/docs/concepts/how-answers-work)