> For the complete documentation index, see [llms.txt](https://help.usejimo.com/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://help.usejimo.com/docs/experiences/agent/knowledge.md).
# Knowledge
Manage everything the Agent knows. The quality and coverage of your knowledge base directly determines how helpful the Agent is to your users.
### Overview
The Knowledge section is where you train and configure your Agent's knowledge base. Everything the Agent can answer comes from what you provide here.
**Dashboard path:** [Agent > Train > Knowledge](https://i.usejimo.com/agent/knowledge)
Knowledge is organized into four tabs:
* [**Interface**](/docs/experiences/agent/knowledge/interface.md): Let the Agent read the pages of your own product. Scan a page once, or let the Agent read it live, so it can answer questions about what users actually see on screen.
* [**Sources**](/docs/experiences/agent/knowledge/sources.md): Static training data that feeds the Agent. Add URLs (site crawls or individual pages), plain text, or upload files and folders.
* [**Custom Answers**](/docs/experiences/agent/knowledge/custom-answers.md): Curated question-answer pairs that the Agent prioritizes.
* [**User Context**](/docs/experiences/agent/knowledge/user-context.md): The user properties the Agent is allowed to read, so its answers fit the person asking.
{% hint style="warning" %}
#### **Tours and Actions are no longer knowledge tabs (replaced by** [**Skills**](/docs/experiences/agent/skills.md)**).**
Since September 2026 the Agent guides and acts through Skills only: a Guidance Skill replaces a Tour connected to the Agent, an Execute Skill replaces a No-Code Action.
Your Tours and Actions still exist as experiences, they are just not triggered from the chat anymore.
See [Migrating to Skills](/docs/experiences/agent/skills/migrating-to-skills.md).
{% endhint %}
***
### What the Agent reads when it answers
A plain question gets one fast answer built from a small set of inputs. Knowing what reaches the Agent, and when, explains most "why did it answer that" questions.
| Source | Read when | How much | Good to know |
| ------------------------------------------------------ | ----------------------------------------------------------------- | ---------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Custom Answers** | Every question, matched question to question | At most one, when the user's question and yours ask the same thing | Returned faithfully, no extra facts. Can carry a button. |
| **Sources** (URLs, files, text, help center) | Every question, searched with the user's own words | The most relevant passages, a few at a time | Something is always retrieved; whether it covers the question is the Agent's call. A follow-up like "and on mobile?" is searched as typed, so it retrieves poorly: encourage full questions. |
| **Skills** | Every question | The best matches above a relevance floor | See [Skills](/docs/experiences/agent/skills.md). |
| **User Context** | Every answer in the live widget | Only the allowlisted properties, 20 at most | Not available in the dashboard test chat. See [User Context](/docs/experiences/agent/knowledge/user-context.md). |
| **Conversation history** | Every turn | The last few turns of the current thread | One thread per browser tab. New chat clears it. No memory across conversations. |
| **Scanned pages** (Interface) | Eye on, when a Skill runs or when the Agent plans a longer answer | Every scanned page of the domain, the current page's landmarks and the elements you kept | Flagged as possibly outdated after 30 days without a rescan. |
| **The live page** | Eye on | The visible text and interactive elements, with privacy masking | No screenshots, no raw HTML. Eye off: the URL only. |
| **General Instructions, language, project name, time** | Always | Whole | See [Chat](/docs/experiences/agent/chat.md). |
| **Escalation rules** | Only if escalation is configured | Whole | See [Escalation](/docs/experiences/agent/escalation.md). |
**What wins when sources overlap**
1. A matching Custom Answer beats Sources.
2. A matching Guide or Execute Skill beats written steps: the Agent offers the Skill instead of describing it.
3. The live page beats documentation for values currently shown on screen.
4. When nothing covers the question, the Agent says so in one sentence, does not guess, and offers to contact support if [Escalation ](/docs/experiences/agent/escalation.md)is configured.
{% hint style="info" %}
**The Sources list under an answer shows what was retrieved, not only what was used.** Every retrieved source with a URL is listed, including passages the answer did not rely on.
{% endhint %}
**Privacy on the live page.** When the eye is on, emails, IBANs, card numbers and phone numbers are masked; password, email, phone and one-time-code fields are hidden; anything inside an element marked `data-jimo-no-capture` is never read.
***
### Test Knowledge
Click the **Test Knowledge** button (top-right) at any time to open a private Agent chat and test responses against your current knowledge base.
This is the fastest way to verify that:
* New sources have been trained correctly.
* Custom Answers trigger on the right questions.
* Skills are suggested in the right context.
{% hint style="warning" %}
#### **SDK and Actions are disabled in preview.**
CTAs show as expected in the chat but are **disabled**: nothing runs on the page. To test a Skill end to end, use **Try this skill** from the Skill panel, or test live.
To test the full behavior you'll have to **test it live**.
{% endhint %}
{% hint style="info" %}
**The test chat always behaves as if the user's eye is on, and does not read User Context.** It shows the best case for Skills and ignores user properties. For the real thing, test live in your product, in a new chat.
{% endhint %}
***
### Best Practices
{% stepper %}
{% step %}
**Building your knowledge base**
* **Scan your key pages first.** The [Interface](/docs/experiences/agent/knowledge/interface.md) tab teaches the Agent what your users actually see. Start with your 3-5 most visited screens.
* **Then layer Sources.** A few well-targeted source crawls (help center, product docs) perform better than a massive crawl of your entire website.
* **Use Custom Answers for wording you must control** (pricing, legal, security disclaimers). For anything that needs steps, build [Skills](/docs/experiences/agent/skills.md).
* **Add two or three properties in User Context** (plan, role, language) so the Agent stops suggesting what the user cannot do.
{% endstep %}
{% step %}
**Teaching procedures**
* **Procedures are Skills.** Guidance for a walkthrough, Execute for a task the Agent does itself, Explain for an answer built from the screen.
* **Write clear descriptions.** Whether for a scanned element or a Skill, describe intent in the words your users use ("How do I invite someone?" not "team member invitation procedure").
{% endstep %}
{% step %}
**Monitoring and iterating**
* **Retrain regularly.** If your product changes often, set auto-retraining to Daily or Weekly in Sources. Rescan Interface pages when your UI changes.
* **Review conversations.** Check [Agent > Analyze](/docs/experiences/agent/analyze.md) to spot questions the Agent struggles with.
* **Test after every change** with the Test Knowledge button, then once more live in your product, eye off and eye on, in a new chat each time.
{% endstep %}
{% endstepper %}
***
{% hint style="info" icon="calendar-clock" %}
### Coming Soon
The Agent sidebar also shows upcoming capabilities under **Coming Soon**:
* **Contextual Triggers**: Context-aware triggers that let the Agent proactively engage users based on what they are doing in the app.
* **Observe**: Monitor and analyze how the Agent interacts with your users in real-time.
{% endhint %}
