AI Triage
AI Triage
Section titled “AI Triage”DeskFork’s AI virtual agent automatically answers common questions using your knowledge base. When the agent is uncertain, it routes the ticket to a human immediately — it never guesses.
How the AI works
Section titled “How the AI works”- Customer message arrives → DeskFork embeds the message with Workers AI
- Vectorize semantic cache searched (cosine similarity threshold: 0.95)
- Cache hit (similarity ≥ 0.95): cached answer returned instantly, no LLM call
- Cache miss: LLM called via AI Gateway with your knowledge base as context
- Response stored in Vectorize cache for future similar questions
- AI response sent to customer via WebSocket (async, non-blocking)
Safe escalation principle
Section titled “Safe escalation principle”The AI will never generate an answer it is not confident about. If:
- Similarity is below 0.95 and no relevant KB article is found
- The question matches a payment/refund/legal keyword
- Negative sentiment exceeds the configured threshold
- The same customer asked 2 follow-up questions in a row without resolution
…the ticket is immediately routed to a human agent with the reason logged in the ticket timeline.
Once a human agent sends a reply, the bot is permanently disabled for that ticket until an agent explicitly re-enables it via “Route to bot” in the ticket sidebar. This prevents the bot from interfering with ongoing human conversations.
Configuring AI triage
Section titled “Configuring AI triage”Go to Settings → AI → Virtual Agent.
| Setting | Description |
|---|---|
| Enabled | Toggle AI on/off per inbox |
| Similarity threshold | Default 0.95; lower = more cache hits but higher error risk |
| Sentiment threshold | Negative sentiment score that triggers escalation (0–1) |
| Escalation keywords | Payment, refund, legal, etc. — triggers immediate handoff |
| Follow-up limit | Number of unanswered follow-ups before escalation (default: 2) |
Knowledge base
Section titled “Knowledge base”The AI uses articles from your knowledge base as its source of truth. Go to Knowledge Base → Articles to:
- Create articles (Markdown supported)
- Organise into collections
- Publish/unpublish articles
- Review AI citation accuracy
Monitoring AI performance
Section titled “Monitoring AI performance”Reports → AI shows:
- Cache hit rate
- Escalation rate by reason
- Average similarity score on cache misses
- Top escalated topics (for KB gap analysis)