Knowledge Base

Empower your AI agents with up-to-date, company-specific knowledge by building a Knowledge Group. Each Knowledge Group lets you:

• Scrape and index content from web pages
• Upload and process documents
• Add and manage FAQs
• Check Knowledge usage in AI agents

What is a Knowledge Group?

A Knowledge Group is a collection of information sources—web content, documents, FAQs—organized for your AI agents to reference. An agent that links a group can answer from what’s in it. Once an agent has been published, its live version keeps answering from the group as it was published until you Publish again (see Knowledge Groups and Published Agents).

---

Create and Manage Knowledge Groups

Open Build → Knowledge Base to see your Knowledge Groups as cards. Each card shows the group’s name, its description (or No description provided), and when it was created and last updated. Use Search knowledge groups by name to find a group, and the controls under the cards to show 10, 20 or 50 groups per page. When there are no groups to show, either because you haven’t created one yet or because a search matches nothing, the page reads No Knowledge Groups Yet. How to Create a Knowledge Group:

1. Click Create Knowledge Group.
2. Enter a Name. A Description is optional, but it helps you tell groups apart.
3. Click Create. The new group opens, ready for you to add content.

Click any card to open that group. It has four tabs, described below: Webscrapes (the tab that opens first), Documents, FAQs and Linked Agents. How to Rename a Group or Change Its Description: Click the three dots menu on the group’s card, choose Edit, make your changes, and click Save Changes. If you close the panel with changes you haven’t saved, you’re asked Discard Changes?. How to Delete a Group: Click the three dots menu on the group’s card, choose Delete, and confirm. Deleting a group can’t be undone.

---

1. Webscrapes & Crawls

Scrape content from a single page or crawl multiple pages to keep your knowledge base fresh and comprehensive. How to Add a Webscrape or Crawl:

1. On the Webscrapes tab, enter the website URL. The box already shows https://, and an address you type with http:// is changed to https://.
2. Select Scrape (single page) or Crawl (multi-page). The selector appears once you’ve entered a valid URL, and Scrape is selected by default.
3. Click Add. If you selected Scrape, it starts right away.
4. For a crawl, the Crawl Options panel opens:
   • Set Limit (the maximum number of pages to fetch, from 1 to 100; the default is 10).
   • Set Max Depth (how many link hops from the starting URL the crawler may follow). It can’t be higher than the Limit.
   • Optionally, expand Advanced Options to control which parts of the site the crawler visits (see below).
5. Click Proceed to start the crawl.

The deeper the starting URL’s path, the higher the minimum for these two fields. Limit and Max Depth can’t go below the number of path segments in the URL (for https://example.com/hc/en-us/articles, that’s 3, and it’s never less than 1), and the panel tells you when it has raised the minimum. Max Depth starts at that minimum. The Co-pilot can start a scrape or crawl for you too. Unless you ask for more, a crawl it starts fetches 5 pages, one link level deep, and it can go up to the same maximum of 100 pages as the Limit, with a Max Depth no higher than the page count. Advanced Options: The Advanced Options section of the Crawl Options panel is collapsed until you open it. It has four settings: If the same address is already in the group as the same type (a scrape or a crawl), adding it again doesn’t create a second entry. It runs the existing entry again: a scrape is refreshed, and a crawl runs again with the options you just chose. Each entry shows its web address (or its name, once it has one), a Scrape or Crawl label (crawls also show their Limit and Depth), a status icon, and its created and completed dates. The status icon is a spinner while the item is running, a green check once it has completed, and a red cross if it failed. When a scrape or crawl finishes while the Webscrapes tab is open, the person who started it sees a message such as Web scrape completed or Web crawl failed, followed by the web address. A rescrape or recrawl shows Rescrape completed or Recrawl failed in the same way. If you were on another tab or page at the time, no message appears, so check the entry’s status icon instead. Managing Webscrapes: Click the three dots menu on an item to open its actions. The menu isn’t available while the item is still running.

• View: Review the actual text content that was scraped or crawled. For a crawl, the list is headed Crawled URLs with a count, and you can filter it with Search URLs…. Click the pencil icon to give the entry a name that’s easier to recognise than its address, then press Enter or click the check mark to save it. The name then shows as the entry’s title in the list. A website the Co-pilot adds for you can be given a name when you ask for it, and that name shows the same way. Websites it added before 14 October 2026 were saved without the name, so they show only their address until you rename them here.
• Rescrape / Recrawl: Refresh the content by running it again, available once the item has completed. Rescrape starts immediately. Recrawl opens Recrawl Options, pre-filled with the crawl’s saved settings, including its Advanced Options (the section opens automatically if any of them was changed from its default); adjust them if you like, then click Start.
• Delete: Remove the entry if it’s no longer needed. It’s removed immediately, with no confirmation. If a published agent uses the group, the removal goes into a draft copy of the group and reaches the live agent at the next Publish (see Knowledge Groups and Published Agents).

---

2. Documents

Upload files to enrich your knowledge base with manuals, guides, policies, and more. Supported Formats: PDF, DOC, DOCX, PPTX, TXT, MD, HTML and JSON. Select and more under the upload area to see the full list. How to Upload:

1. Drag and drop files onto the upload area, or click it to select files. You can add up to 10 files at a time. If you drop or select more than 10 at once, none of them are added, so upload larger sets in batches.
2. Wait for upload and processing to complete. A document shows Processing… while it’s being prepared, and Failed if it couldn’t be processed.
3. Review and manage your document list

A file in a format that isn’t supported is rejected with a message naming the file. Once uploaded, you can download a document from the Knowledge Group to review or archive offline. Click the download (arrow) icon at the right of the document’s row. The icon is greyed out while the document is still processing. If a row has no download icon, the original file isn’t available for that document. To delete a document, click the trash icon on its row.

3. FAQs

Manually add frequently asked questions to help your AI agents answer common customer queries quickly and accurately. How to Add FAQs:

1. Click Add FAQ
2. Enter the question and answer
3. Click Save FAQ, then manage your FAQ list

The newest FAQs are listed first. Use the pencil icon to edit one (click Save changes when you’re done) or the trash icon to delete it. Deleting is immediate and asks for no confirmation. Use the Search FAQs… box to quickly find specific entries; it looks at both the question and the answer. On an agent you’ve published, FAQ changes reach callers, chats and emails at the next Publish. FAQs are used on every channel a linked agent works — voice, chat, and email. Every FAQ in a linked group is written into the agent’s instructions, so a long list adds to every call, chat and email the agent handles (see Knowledge Across Channels). An email agent answers a question that matches one of its FAQ entries straight from that entry, rather than composing an answer of its own. See Email Agents.

---

4. Linked Agents

See which AI agents are currently linked to this Knowledge Group. Manage connections to ensure your agents always have access to the right information. How to Manage Agents:

• View all linked agents. If none are linked, the tab says so.
• Click Unlink next to an agent to remove the connection. The group and its content stay as they are. On an agent you’ve published, the live version keeps using the group until you Publish.

Linking happens from the agent’s side, not from this tab. See Linking a Knowledge Group to an Agent.

---

Linking a Knowledge Group to an Agent

An agent only uses the Knowledge Groups linked to it, and it can have at most 2 linked at a time. How to Link a Group:

1. Open the agent (Build → AI Agents) and select its Knowledge Base card. Your Knowledge Groups are listed, with the ones already linked to this agent always first on page 1, however many groups you have, each marked with a Linked badge.
2. Select Link on the group you want. You’ll see a message confirming that the group was linked to the agent. If the group was on a later page, the list moves to page 1, where the group is now listed first.
3. To remove a link later, select Unlink on a linked group, or use Unlink on the group’s Linked Agents tab.

On a wide screen, Link and Unlink appear when you point at a group’s card; on a narrow screen they’re always visible. Linking and unlinking change the agent’s draft. On an agent you’ve published, callers keep the groups it was published with until you Publish (see Knowledge Groups and Published Agents).

How Agents Use Knowledge

Once a Knowledge Group is linked, the platform adds standing rules to the agent’s instructions, on voice, chat and email alike:

• The agent searches its linked Knowledge Groups before it answers a question, and answers concisely from what it finds.
• It is told not to guess, invent, or answer from outside knowledge, and to speak as if the knowledge is its own rather than mention uploaded files or documents.
• If nothing reliable turns up for a complete question, the built-in reply is along the lines of “I don’t have information regarding your query in the available knowledge. Is there anything else I can help you with?” For a short or casual reply, such as “okay”, it asks only “Is there anything else I can help you with?”

So when a customer asks a question your Knowledge Group doesn’t cover, the agent is likely to give that reply instead of a general-knowledge answer. Add the missing answer as an FAQ, a document or a crawled page, then test the agent with Talk or Chat, and Publish if the agent has been published so callers get it too. FAQs work a little differently: they aren’t searched. They are written straight into the agent’s instructions, so the agent has them in front of it on every turn.

---

Knowledge Groups and Published Agents

An agent you haven’t published uses its Knowledge Groups as you leave them, so your changes reach callers, chats and emails as you save them, once any document or web page has finished processing. Once an agent has been published, it answers from the version of each group that was published with it, and your changes wait for the next Publish. Talk and the test Chat always use your latest changes, so you can check an edit before callers meet it. See Publishing Your Agent and What Waits for Publish.

• Web pages and documents make a draft copy. When a published agent uses the group, the first change you make to its web pages or documents (adding, rescraping, replacing or deleting one) makes a copy of the group, under the same name, and your change goes into the copy. Studio switches you to the copy, and the agent shows Unpublished edits. The live agent keeps answering from the version it was published with until you Publish.
• Let the copy finish. While the copy is made, the group’s other documents show Processing… and its web pages show the spinner of a running item. Let that finish before you publish.
• FAQs, linking and unlinking are changes to the agent’s draft too, and they don’t make a copy. The live agent follows them at the next Publish.
• Shared groups: when several agents link the group, the copy takes the group’s place on all of them, and each published agent that uses it shows Unpublished edits.
• Unlinking can hide a group. If you unlink a group from a published agent and no other agent has it linked, the group leaves Build → Knowledge Base, and the agent’s Knowledge Base card, because versions of that agent still use it. Publishing the agent without the group doesn’t bring it back: the agent’s earlier versions still use it. You can’t link it again or delete it from the list; see the next point.
• Deleting needs more than a Publish. A group that any version of a published agent uses can’t be deleted, even after you unlink it from the agent. An attempt shows “This knowledge base is in use by a published agent and cannot be deleted. Remove it from the agent and publish again first.” Publishing is only the first step. Unlink the group, Publish so the live version no longer uses it, then select Versions in the agent’s header, open the menu on each older version that used the group and choose Delete Version (see Version History). A deleted version is gone for good, so you can no longer make it live or export it. Deleting the last version that uses an unlinked group deletes the group itself, with its content, and that can’t be undone. Deleting a published agent releases its groups, so you can then delete them.

---

Knowledge Across Channels

Everything in a Knowledge Group — scraped web content, uploaded documents, and FAQs — is available to a linked agent on voice, on chat, and on email. An email agent answers a matching question authoritatively from the FAQ entry, and searches its linked Knowledge Groups before it writes a reply. This is new behaviour for email. Previously FAQs were honoured on voice and chat but ignored on email replies, so an email agent answered from general model knowledge, hedged, or escalated a thread even when the exact answer was sitting in its FAQ list. Nothing needs configuring. Your existing FAQ entries and Knowledge Groups start being used on the agent’s next email turn. See Email Agents.

Behaviour worth knowing

• Email replies may be worded differently than before. Anything covered by an FAQ is now answered from that entry, so a reply that used to be composed freely will follow your FAQ wording instead — and an FAQ entry that is stale or wrong will start showing up in email replies as well as on calls and chats. Review your FAQ list before relying on it for email.
• A long FAQ list makes each email reply cost more. The agent’s full FAQ list is included in every email reply it composes, so the longer the list, the higher the cost of each email turn. Keep FAQs to genuinely frequent questions, and move detail into documents — those are searched only when they are relevant.
• A long FAQ list also makes every call and chat heavier. The same full list is written into the agent’s instructions on voice and chat too, so the longer it is, the more the agent has to work through on every turn of every conversation.

---

Import and Export

When you export an agent, the bundle now carries its linked Knowledge Groups’ full content — name, description, every FAQ, and every URL together with its crawl settings (crawl format, crawl mode, page and depth limits, crawl-entire-domain, ignore-sitemap, and the include/exclude path lists). Crawl-entire-domain, ignore-sitemap and the path lists correspond to the Include Subdomains, Skip Sitemap, Exclude Paths and Only Include Paths settings described under Webscrapes & Crawls. Previously a bundle recorded knowledge groups by name only, so importing into another workspace produced empty groups you had to rebuild by hand.

• Within the same region, already-crawled page content and uploaded files are copied straight across rather than re-crawled, so the imported group is usable immediately.
• Across regions, URLs are re-crawled from scratch — content can differ if the site has changed — and uploaded documents are omitted, because a file cannot be regenerated from a URL. Re-upload those manually.

Other outcomes you may see:

• If the destination already holds a group whose content is identical, that group is reused and nothing is re-crawled.
• If the destination’s copy has moved on independently, the bundle’s content is ignored and your newer content is kept.
• If the two have genuinely diverged — or if either side’s content fingerprint is unknown — the importer creates a second, separate Knowledge Group rather than clobbering your edits, leaving two similarly named entries to reconcile by hand.

A file from Export Version (see Version History) follows different rules from the ones above, because it holds the knowledge groups as they were in that version. A group is linked as it is only if it’s still in the workspace with the same FAQs as that version. Otherwise the importer creates a new Knowledge Group from the file, and it never matches a group by name, so a group with the same name doesn’t take its place. A group it creates follows the same-region and across-region limits above. Studio shows no warning when a group couldn’t be brought over, so check Build → Knowledge Base after importing. Duplicating an agent inside a single workspace is unchanged: it shares the existing Knowledge Group and none of this applies. Because the copy shares the group, a change to the group reaches the copy as soon as you save it (a duplicate starts as Not published), and reaches a published original at its next Publish. Expect billable crawl activity when an import does re-crawl URLs.

---

Best Practices

• Consolidate Knowledge Groups: An agent can have at most two Knowledge Groups linked at a time (see Linking a Knowledge Group to an Agent), so it works best to add related documents, web content and FAQs to a single Knowledge Group. It also makes it easier to see at a glance what the agent knows.
• Use FAQs for Common Questions: Add frequent questions directly to the FAQ section. These go straight into the agent’s instructions instead of being searched for, resulting in faster responses for common queries. Keep the list to genuinely frequent questions, because every FAQ in a linked group is part of the agent’s instructions on every turn (see Knowledge Across Channels).
• Regular Updates: Keep your webscrapes and documents up to date to ensure the AI has the latest information. On an agent you’ve published, Publish after you update them so callers get the change.
• Review Connections: Periodically check which agents are linked to ensure they have the correct access.
• Avoid Conflicting Information: Ensure your documents do not contain contradictory facts, which can confuse the AI Agent.
• Keep it Relevant: Only add information relevant to the specific use case. Avoid overloading the Knowledge Group with excessive or irrelevant data to maintain optimal performance.
• Use Structured Formatting: Format documents with clear headings and bullet points. This helps the AI understand the structure and relationship of information.
• Describe Visuals: If your documents contain charts or images, ensure the key takeaways are also written in text so the AI can read them.
• Test After Updates: After modifying the Knowledge Group, ask your AI Agent specific questions to verify it correctly understands and uses the new material. Talk and Chat use your latest changes, even on a published agent, so test there before you Publish.

---

Troubleshooting