Answer Corrections

Answer corrections let you fix bad AI answers by creating Q&A override pairs. When a visitor asks a question similar to one you've corrected, the AI uses your correction to inform its response.

How It Works

1. You create a correction with a trigger question and corrected answers (per language)
2. UserBoost embeds the question using the same AI that powers the chatbot's knowledge base
3. When a visitor asks something semantically similar, the correction is injected into the AI's context
4. The AI generates a natural response informed by your correction

The AI doesn't just repeat your correction word-for-word — it uses it as authoritative source material while maintaining a natural conversational tone.

Creating Corrections

From the Corrections Page

1. Go to your domain's Corrections tab
2. Click Add correction
3. Enter the trigger question (what the visitor would ask)
4. Add one or more languages with the corrected answer
5. Click Save

From a Conversation

1. Go to Conversations and find a conversation with a bad AI answer
2. Click Correct answer next to the assistant message
3. You'll be taken to the Corrections page with the visitor's question pre-filled
4. Add your corrected answer and save

Per-Language Answers

Corrections support multiple languages. Add a language tab for each market you serve:

• EN: "Returns must be initiated within 30 days via our returns portal."
• DA: "Returneringer skal påbegyndes inden 30 dage via vores returportal."

The system matches the correction to the visitor's language. If no answer exists for the visitor's language, the correction is skipped.

Semantic Matching

Corrections use semantic similarity — not keyword matching. This means:

• "How do I return an item?" matches "I want to send something back"
• "What's your return policy?" matches "Can I get a refund?"
• Matching works across languages (a Danish question can match an English correction)

A correction triggers when similarity is 75% or higher.

Duplicate Detection

When creating a correction, the system checks if a similar one already exists. If it finds a match above 85% similarity, you'll see a warning — but can still save if the questions are genuinely different.

Limits

• Maximum 50 corrections per domain
• Question: up to 500 characters
• Answer: up to 2,000 characters per language