Data sources

In Data sources, you manage the content your AI agent uses in its answers. Each website, file, question and answer, or text is one data source. Together, these sources form the agent’s knowledge base.

Adding a data source does not train the underlying model. The agent uses the content you add as the basis for its answers.

Your daily work happens on one page:

1. See the latest knowledge-base update, automatic refresh setting, and usage at the top of the page.
2. Add a source from one of four source cards.
3. Search, filter, and open existing sources in one list.
4. Inspect, edit, and test a selected source in the panel on the right.

Choose the format that is maintained most reliably. If content changes on a website, add it as a website. If it lives in an internal guide, upload a file or add a text source.

Choose Website. The panel offers three ways to add content:

• Fetch links discovers pages for you to review.
• Use sitemap uses a sitemap URL when one is available.
• Single page adds one exact URL.

1. Enter an address

   Enter the full website or page URL, for example https://company.com.

2. Refine if needed

   Check advanced settings before fetching if you need more control. Otherwise, continue to selecting content.

   Path filters, query parameters, and fetch speed

   Open advanced settings to exclude paths, handle query parameters, or slow down fetching.

3. Choose content

   After discovery, search and review the found URLs, select relevant pages, and exclude anything unnecessary.

4. Save the selection

   Save the pages you selected.

   Expected result: the selected sources are saved as drafts. Publish the changes under Update the knowledge base .

1. Choose File.
2. Drop files into the upload area or choose them from your computer.
3. Follow the upload queue in the panel. Review file-specific errors; you do not need to restart the entire upload.

Supported formats are PDF, DOCX, TXT, Markdown, CSV, and HTML. The old .doc format is not supported; convert it to DOCX or PDF first.

If processing fails, check the format, size, and protection. A scanned PDF may look readable to a person while containing an image instead of machine-readable text.

Choose Question and answer when you need one controlled answer for one or more customer phrasings. Give it a clear title, add questions in the customer’s own words, and format the answer in the editor.

A good question and answer:

• uses natural customer language;
• answers one subject at a time;
• includes relevant conditions and limits;
• does not conflict with other data sources.

Try the following structure with your own verified information:

Title: Delivery area
Questions: Where do you deliver? Can I order from abroad?
Answer: [State the actual delivery area and any restrictions.]

Text in brackets is a placeholder. Replace it with your company’s information before saving; do not publish this example unchanged.

Choose Text, give the content a title, and write it in the editor. The editor shows formatted content as it will be saved and the text size in relation to the limit.

Text is useful for special opening hours, campaign terms, or short information that is not yet on your website. Keep one text source focused on one topic; create separate sources for separate topics.

The source list brings every data source together. It shows the source name, type, status, content amount, and last update.

• Use search to find a title or URL.
• Filter by data-source type or status.
• Open a data source from its name to see its details.
• If you have management permission, you can select multiple sources for shared actions.

On mobile, the same information is shown in compact source cards so that important status and actions remain easy to read.

When you open a data source, the panel on the right includes these views as appropriate for the source type:

For a website source, you can restore an excluded URL, remove a URL from the knowledge base, or retry a failed URL when needed.

Test search checks retrieval from the selected source. It is not a full-agent conversation test. Try a naturally phrased customer question, review the returned content and compare it with the original source. If information is missing or outdated, correct the source and update the knowledge base. Finally, test the question in the agent preview, where instructions and conversation context also affect the answer.

Update knowledge base publishes pending data-source changes for the AI agent on every plan.

1. Choose Update knowledge base.
2. Follow the current stage, processed source count, and any failed or skipped sources in the update panel.
3. Review the completed update’s result on the same page. Then try an important customer question in Test search.

Expected result: successfully published content is available to the agent.

Automatic refresh updates website sources weekly when included in the workspace subscription. Check availability in the table below and in the workspace. Otherwise, use manual updates. File, question-and-answer, and text changes are always published with the manual Update knowledge base action.

• Essential service, pricing, delivery, and contact content is ready.
• Outdated or conflicting data sources have been archived or corrected.
• Necessary repeated questions and answers have been added.
• You have tested natural customer questions in Test search.
• Someone owns data-source maintenance when company information changes.

Also check availability and any applicable add-on in your workspace. This table gives minimum tiers for these features; it is not a workspace usage limit or a list of every integration.