Overview

The AI Answer Assessment feature includes functionalities like Hallucination Checks. This tool ensures that users receive accurate and reliable information by systematically evaluating the quality and adherence of messages to retrieved sources.

Who is it for

This feature is ideal for:

Benefits

• Accuracy: Ensures that messages are evaluated for factual correctness.

• Reliability: Reduces the risk of misinformation by categorizing hallucination levels.

• User-Friendly: Automatically generates assessments without requiring additional user actions.

• Transparency: Provides clear insights into the quality of the information provided.

Use Cases

• Content Verification: Assessing the accuracy of information in chat messages.

• Regulatory Compliance: Ensuring that messages meet predefined compliance standards (future functionality).

• Educational Purposes: Teaching users about the importance of source adherence and factual accuracy.

Step-by-Step Guide

How to Activate the Answer Assessment

The Hallucination Check feature is automatically enabled if the Hallucination Check option is activated in Internal Knowledge Search.

How It Works

The message assessment is generated instantly after the assistant's message is completed and appears in the chat interface without requiring a page reload.

Interface

Once the message assessment is available, it appears as an accordion at the bottom of the message.

Assessment Types & Sentiments

Assessments are classified by:

• Type:  String e.g HALLUCINATION

• Sentiment: GREEN, YELLOW, or RED

Positive/Verified assessment	Unverified assessment	Negative assessment

Positive/Verified assessment	Unverified assessment	Negative assessment
Open Assessment with low hallucination level	Open Assessment with medium hallucination level	Open Assessment with high hallucination level

For more information on the meaning of the hallucination level, please visit ourhttps://unique-ch.atlassian.net/wiki/spaces/PUBDOC/pages/1410531424 documentation.

Tips & Tricks

• Understand Sentiments:

  • Green: Positive/Verified assessment.

  • Yellow: Unverified assessment.

  • Red: Negative assessment.

• Learn from Examples: Review the explanations provided for each sentiment to understand the reasoning behind the evaluation.

Limitations

• Compliance Checks: This functionality is not yet supported.

• Scope: The feature currently focuses on hallucination levels and does not cover all aspects of message quality.

• Dependence on Settings: The Hallucination Check must be activated in the Internal Knowledge Search settings to function.