About

To help your agent answer visitors’ questions, you give it content to draw on, such as videos, documents, web pages, or written information. This content is called the agent’s knowledge.

This article explains how your agent uses that knowledge to build answers and what happens if it doesn’t contain the information a visitor needs.

How knowledge retrieval works

When you add videos, audio files, or documents, the agent extracts the text from captions or document content, divides it into short sections, and prepares those sections for searching.

When a visitor asks a question, the agent searches by meaning, rather than requiring the same words. For example, “How much time off do I get?” can match a section about “annual vacation allowance.”

The agent then uses the relevant sections to compose an answer. This combination of retrieving information and generating an answer is called Retrieval-Augmented Generation (RAG).

Web sources and free text work differently, as explained below.

How the agent uses each type of knowledge

Type	How the agent uses it	Limits and requirements
Videos, audio files, and documents (added through the Media Manager)	Reads document text and the captions of videos and audio files. It also reads on-screen text if you order OCR (on-screen text recognition) through REACH.	New files can take up to 3 hours to process. Up to 1 GB per file. Videos and audio files need captions. Image files aren't supported. Uploaded files use your knowledge storage allowance. How much storage you have depends on your plan.
Web sources	Reads the page and up to five levels of subpages live, each time a visitor asks a question. Uses the latest content from those pages. Reads only text, not images or videos.	Up to 10 URLs per agent. Public pages only, not pages behind a login or firewall. Doesn't use your knowledge storage allowance.
Free text	Available to the agent for every question, without a search. Useful for short, important facts, such as opening hours or prices.	Up to 10,000 characters. Doesn't use your knowledge storage allowance.

Key behaviors

• The agent answers in the visitor's language, even if your knowledge is in a different language.
• Each agent has its own knowledge. Media, web sources, and free text aren't shared between agents.

How to guide the agent to answer accurately

By default, if the agent can't find an answer in its knowledge, it can answer from general AI knowledge. This means it might give answers that aren't based on your content.

To keep answers based on your knowledge:

1. On the Playbook tab, go to Agent role.
2. Add an instruction such as: "If the information isn't in your knowledge, say that you don't know."
3. Click Publish.

To stop the agent from discussing specific subjects, add them as restricted topics on the Playbooktab. You can add up to 10.

How to build your agent's knowledge

Use these guides to add content and prepare knowledge that helps your agent give better answers:

• Add knowledge to your agent — Learn how to add or delete media, web sources, and free text.
• Best practices for adding knowledge to your Agentic Avatar — Learn how to choose and prepare content, test your agent, and fix answer problems.