LLM-Assisted Attendant Features

General Bots provides AI-powered assistance to human attendants during customer conversations. These features help attendants respond faster, more professionally, and with better context awareness.

Overview

When the bot transfers a conversation to a human attendant (via TRANSFER TO HUMAN), the LLM orchestrator continues working in the background to assist the human. This creates a hybrid experience where AI augments human capability rather than replacing it.

The system uses the same PROMPT.md and bot personality configured for the bot, ensuring consistency between bot responses and attendant assistance.

Features

Configuration

Add these settings to your bot’s config.csv:

name,value

# Enable all LLM assist features
attendant-llm-tips,true
attendant-polish-message,true
attendant-smart-replies,true
attendant-auto-summary,true
attendant-sentiment-analysis,true

# Optional: Set bot personality for context
bot-system-prompt,You are a friendly customer service assistant for Acme Corp
bot-description,Premium support for enterprise customers

Selective Enablement

Enable only the features you need:

name,value
attendant-smart-replies,true
attendant-sentiment-analysis,true

Feature Details

1. Real-time Tips (attendant-llm-tips)

When a customer sends a message, the LLM analyzes it and provides actionable tips to the attendant.

Tip Types

Example Tips

Customer says: “This is ridiculous! I’ve been waiting 3 days for a response!”

Tips generated:

• ⚠️ Customer is frustrated - use empathetic language and apologize
• 🎯 Customer has been waiting for support response
• ✅ Acknowledge the delay and provide immediate assistance

API Usage

' Internal API - automatically called by UI
POST /api/attendance/llm/tips
{
    "session_id": "uuid",
    "customer_message": "message text",
    "history": [{"role": "customer", "content": "..."}]
}

2. Message Polish (attendant-polish-message)

Before sending, attendants can polish their message with one click. The LLM improves grammar, clarity, and tone while preserving the original meaning.

Supported Tones

• professional (default)
• friendly
• empathetic
• formal

Example

Original: “ya we can do that but u need to wait til tmrw”

Polished: “Yes, we can certainly help with that! Please allow until tomorrow for us to process your request.”

Changes: Fixed grammar, improved clarity, added professional tone

API Usage

POST /api/attendance/llm/polish
{
    "session_id": "uuid",
    "message": "original message",
    "tone": "professional"
}

Response:

{
    "success": true,
    "original": "ya we can do that...",
    "polished": "Yes, we can certainly...",
    "changes": ["Fixed grammar", "Improved tone"]
}

3. Smart Replies (attendant-smart-replies)

Generate three contextually appropriate reply suggestions based on the conversation history and bot personality.

Reply Categories

• greeting - Opening responses
• answer - Direct answers to questions
• acknowledgment - Empathetic acknowledgments
• solution - Problem-solving responses
• follow_up - Continuation questions
• closing - Conversation wrap-up

Example

Customer: “How do I reset my password?”

Suggested Replies:

1. Empathetic: “I understand how frustrating it can be when you can’t access your account. I’ll help you reset your password right away.”

2. Solution-focused: “To reset your password, please go to the login page and click ‘Forgot Password’. You’ll receive an email with reset instructions.”

3. Follow-up: “I can help you with that! Are you trying to reset the password for your main account or a sub-account?”

API Usage

POST /api/attendance/llm/smart-replies
{
    "session_id": "uuid",
    "history": [
        {"role": "customer", "content": "How do I reset my password?"},
        {"role": "attendant", "content": "Hi! Let me help you with that."}
    ]
}

4. Auto Summary (attendant-auto-summary)

When an attendant takes a conversation, they receive an instant summary of what’s happened so far. This is especially useful for:

• Long conversations
• Transfers between attendants
• Complex multi-issue discussions

Summary Contents

Example Summary

{
    "brief": "Customer requesting refund for damaged product received yesterday",
    "key_points": [
        "Order #12345 arrived damaged",
        "Customer sent photos as proof",
        "Previous agent offered replacement"
    ],
    "customer_needs": [
        "Full refund instead of replacement",
        "Confirmation email"
    ],
    "unresolved_issues": [
        "Refund approval pending"
    ],
    "sentiment_trend": "stable",
    "recommended_action": "Escalate to supervisor for refund approval"
}

API Usage

GET /api/attendance/llm/summary/{session_id}

5. Sentiment Analysis (attendant-sentiment-analysis)

Real-time analysis of customer emotional state to help attendants respond appropriately.

Analysis Components

Example Analysis

Customer message: “I’ve been trying to get help for TWO WEEKS! This is absolutely unacceptable!”

{
    "overall": "negative",
    "score": -0.8,
    "emotions": [
        {"name": "frustration", "intensity": 0.9},
        {"name": "anger", "intensity": 0.7}
    ],
    "escalation_risk": "high",
    "urgency": "high",
    "emoji": "😟"
}

The UI shows a warning: ⚠️ High escalation risk detected

API Usage

POST /api/attendance/llm/sentiment
{
    "session_id": "uuid",
    "message": "customer message",
    "history": []
}

WhatsApp Attendant Commands

Attendants using WhatsApp can access LLM assist features via commands:

Queue Management

AI Assistance

Example Usage

Attendant: /queue
Bot: 📋 Queue (3 waiting)
     1. John Smith (whatsapp)
        Status: waiting | ID: abc123
     2. Maria Santos (web)
        Status: waiting | ID: def456
     ...

Attendant: /take
Bot: ✅ Conversation assigned
     Customer: John Smith
     Session: abc123
     
Attendant: /tips
Bot: 💡 Tips for this conversation

     🎯 Customer is asking about billing
     ✅ Check their account status before responding
     ⚠️ Previous conversation ended unresolved
     
Attendant: /polish thank u for waiting ill check ur account now
Bot: ✨ Polished message

     "Thank you for your patience! I'll check your account right now."
     
     Changes:
     • Fixed grammar
     • Improved professionalism
     
Attendant: /resolve
Bot: ✅ Conversation resolved
     Session abc123 marked as resolved.

Web Attendant Console

The web attendant console (/suite/attendant/) integrates LLM assist features directly:

Insights Panel

The right sidebar shows:

• Customer Sentiment - Real-time with emoji indicator
• Detected Intent - What the customer wants
• Conversation Summary - Auto-generated when you select a conversation

Suggested Replies

Below the insights, three AI-generated reply suggestions appear. Click any suggestion to insert it into the message input.

Polish Button

The ✨ button next to the message input polishes your text before sending.

Tips Notifications

When a customer sends a message:

1. Tips appear in the insights panel
2. High-priority tips show as toast notifications
3. Warning tips (escalation risk) are highlighted

Bot Personality Integration

LLM assist uses your bot’s personality when generating suggestions. Set this in config.csv:

name,value
bot-system-prompt,You are a friendly tech support agent for CloudSoft Inc. Be helpful and patient.
bot-description,Enterprise software support

Or in your start.bas header:

REM CloudSoft Support Bot
REM Friendly, patient, and technically knowledgeable
REM Always offer to escalate complex issues

TALK "Welcome to CloudSoft Support!"

The LLM reads these comments to understand the bot’s personality and applies the same tone to:

• Smart reply suggestions
• Message polishing
• Tips generation

API Reference

Endpoints

Check Configuration

GET /api/attendance/llm/config/{bot_id}

Response:
{
    "tips_enabled": true,
    "polish_enabled": true,
    "smart_replies_enabled": true,
    "auto_summary_enabled": true,
    "sentiment_enabled": true,
    "any_enabled": true
}

Fallback Behavior

When LLM is unavailable, the system provides fallback functionality:

Best Practices

1. Start with Smart Replies

If you’re unsure which features to enable, start with attendant-smart-replies. It provides immediate value with low overhead.

2. Enable Sentiment for High-Volume Support

For teams handling many conversations, attendant-sentiment-analysis helps prioritize frustrated customers.

3. Use Polish for Quality Consistency

Enable attendant-polish-message to ensure consistent, professional communication regardless of individual writing skills.

4. Tips for Complex Products

For products with many features or complex workflows, attendant-llm-tips helps attendants quickly understand context.

5. Summary for Shift Changes

Enable attendant-auto-summary if your team has shift changes or frequent transfers between attendants.

Troubleshooting

“Feature is disabled” Message

Add the feature to your config.csv:

attendant-smart-replies,true

Slow Response Times

LLM calls add latency. If responses are slow:

• Use a faster LLM model
• Enable only essential features
• Check your llm-url configuration

Generic Suggestions

If suggestions seem generic:

• Set bot-system-prompt in config.csv
• Add personality comments to start.bas
• Ensure conversation history is being passed

WhatsApp Commands Not Working

1. Verify the attendant is registered in attendant.csv
2. Check that the phone number matches exactly
3. Ensure crm-enabled,true is set

See Also

• Transfer to Human - Bot-to-human handoff
• Attendance Queue - Queue configuration
• LLM Configuration - LLM setup
• config.csv Format - Configuration reference