> ## Documentation Index
> Fetch the complete documentation index at: https://watermelon.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# How to handle hallucinations

> Learn how to reduce hallucinations and help your AI Agent give more reliable answers.

<Warning>
  Hallucinations are a natural part of how all large language models work.

  They can be **reduced**, but never completely eliminated — because models are rewarded for making confident predictions, even when uncertain. This is normal behaviour for every LLM.
</Warning>

## **What are hallucinations?**

Your AI Agent is hallucinating when it produces an answer that **sounds correct** but is **not based on real data**.

This happens when the model predicts a likely-sounding response instead of acknowledging uncertainty.

A hallucination happens when the Agent:

* **Fills in missing information** with its best guess
* **Invents details** instead of saying “I don’t know”
* **Gives confident answers** without reliable knowledge to support them

**Example:**

Question: *“What is the delivery time for product X?”*

Agent (hallucinating): *“Delivery time is 24 hours.”* (even though no such data was provided)

### Why does this happen?

Large language models work by predicting the **most likely next words**.

They do **not automatically verify facts** or check databases—unless you provide that structure.

Hallucinations often occur when:

* Knowledge is **unclear**
* Information is **incomplete**
* Data is **outdated**
* Instructions are **vague**
* The Agent is **uncertain but still rewarded for guessing**

## What can you do when you Agent hallucinates?

While you can’t remove hallucinations entirely, you *can* significantly reduce them by giving your AI Agent strong, structured, and up-to-date knowledge.

### **Tips to reduce hallucinations**

* **Keep your content up-to-date**: outdated information leads to incorrect answers. Regularly update your pricing, terms & conditions, product specifications, processes and workflows, manuals and guides.
* **Document key topics thoroughly**: if important topics are incomplete, the Agent will fill in gaps on its own.
* **Use structured knowledge**: headings, bullets and clear definitions help the model.

  Bad example:

  > We have three packages. Prices change sometimes. The main package is Pro.

  Good example:

  > Our packages and pricing for 2026 are:
  >
  > * Basic – €19/month
  > * Pro – €49/month
  > * Enterprise – custom pricing
* **Write clearly and specifically**: avoid vague language and assumptions.

  Bad example:

  > Product X usually costs around €20.

  Good example:

  > Product X costs €19.95 (fixed price).

  Bad example:

  > You can order on our website.

  Good example:

  > You can place an order via [www.company.com/order](http://www.company.com/order) via the button in the top right.

### **How adding knowledge sources helps reduce hallucinations**

Using **structured, high-quality knowledge Sources** helps your AI Agent rely on real information instead of predictions.

By adding sources such as:

* [**Feeds**](/help-center/features/xml-connection) (product data, pricing, availability)
* [**Web content**](/help-center/features/website-source) (help center, FAQ, policy pages)
* [**Documents**](/help-center/features/files) (PDFs, manuals, price lists)
* [**Real-time search**](/help-center/features/search) (Search)

…your Agent can:

* reference **exact**, **machine-readable** information
* reduce the need to guess
* base answers on **current and verified** content
* give more consistent and reliable responses

The more structured and complete your knowledge, the fewer hallucinations your Agent will produce.