Stage 1: Chit-Chat Detection

Overview

The first stage of the Atlas Reasoning Engine determines whether the input is a valid query within scope or casual conversation (chit-chat).

Processing Time: 50-100ms
Type: Fast validation
Purpose: Filter out off-topic requests and casual greetings

---

What Happens in This Stage

Input Analysis

Atlas analyzes the incoming text to classify the intent:

Classification Categories

Category	Description	Atlas Response
✅ Valid Query	Business-related question or request	Continue to Stage 2
⚠️ Chit-Chat	Greeting, pleasantry, or social interaction	Simple conversational response
❌ Out of Scope	Request outside agent's capabilities	Polite rejection with guidance

---

Examples

✅ Valid Queries (Pass Through)

These queries proceed to Stage 2:

"What's the status of order 12345?"
"Show me all open cases for Acme Corp"
"Create a new opportunity for John Smith"
"What are my tasks for today?"
"Find contacts in the San Francisco region"

⚠️ Chit-Chat (Simple Response)

These get handled with simple responses:

"Hello" → "Hi! How can I help you today?"
"How are you?" → "I'm here to help! What can I do for you?"
"Good morning" → "Good morning! What would you like to know?"
"Thank you" → "You're welcome! Is there anything else?"

❌ Out of Scope (Rejected)

These get politely rejected:

"Tell me a joke" → "I'm here to help with Salesforce tasks. How can I assist you?"
"What's the weather?" → "I focus on Salesforce data. How can I help with your business?"
"Sing a song" → "I'm a business assistant. What would you like to know about your data?"

---

How It Works

1. Text Preprocessing

Original Input: "Hey there! What's the status of my order?"
↓
Normalized: "what status order"
↓
Intent Tokens: ["status", "order"]

2. Intent Classification

Atlas uses a classifier to determine the primary intent:

Signal	Weight	Indicates
Business terms	High	Valid query
Question words	Medium	Information request
Greetings only	High	Chit-chat
Action verbs	High	Valid command
Social phrases	Medium	Chit-chat

3. Confidence Scoring

# Pseudocode
intent_score = calculate_intent_confidence(tokens)

if intent_score > 0.8:
    return VALID_QUERY
elif intent_score > 0.3:
    return CHIT_CHAT
else:
    return OUT_OF_SCOPE

---

Benefits

⚡ Performance Optimization

• Fast Rejection: Out-of-scope queries don't consume processing resources
• Token Savings: Avoids expensive LLM calls for simple greetings
• Quick Response: Users get immediate feedback for casual interactions

💰 Cost Reduction

Without Chit-Chat Detection	With Chit-Chat Detection
All queries → full pipeline	Only valid queries → full pipeline
"Hello" costs 1000+ tokens	"Hello" costs < 50 tokens
Wasted processing time	Optimized resource usage

---

Configuration

Adjusting Sensitivity

Admins can tune the chit-chat detector:

// Example configuration
{
  "chitchat_detection": {
    "enabled": true,
    "confidence_threshold": 0.8,
    "allow_greetings": true,
    "handle_pleasantries": true,
    "out_of_scope_message": "I'm here to help with Salesforce. What would you like to know?"
  }
}

Custom Responses

Customize responses for chit-chat:

{
  "responses": {
    "greeting": "Hello! How can I assist you today?",
    "farewell": "Goodbye! Feel free to reach out anytime.",
    "thanks": "You're welcome! Anything else I can help with?",
    "out_of_scope": "I focus on Salesforce. How can I help with your business?"
  }
}

---

Monitoring

Performance Considerations

Typical Chit-Chat Detection Performance:

• Processing Time: 50-100ms per request
• Chit-Chat Detection Rate: 10-20% of incoming queries
• Out-of-Scope Detection Rate: 3-8% of queries
• False Positive Rate: Less than 5% with well-tuned prompts

note

Monitor your agent's performance through Setup → Einstein → Einstein for Service. Salesforce provides built-in analytics for query classification and detection accuracy.

---

Troubleshooting

Issue: Valid Queries Marked as Chit-Chat

Symptoms:

• Business queries get simple responses
• Users report "agent doesn't understand"

Solutions:

1. Lower confidence threshold (0.8 → 0.7)
2. Add business terms to allowlist
3. Review false positive logs
4. Retrain classifier with examples

Issue: Too Many Queries Pass Through

Symptoms:

• High processing costs
• Many irrelevant queries in pipeline

Solutions:

1. Raise confidence threshold (0.8 → 0.9)
2. Expand out-of-scope patterns
3. Add domain-specific filters
4. Review query patterns

---

Best Practices

✅ Do's

• ✅ Allow greetings at conversation start
• ✅ Combine greeting + query in single input
• ✅ Monitor false positive/negative rates
• ✅ Customize responses for your brand
• ✅ Track chit-chat patterns over time

❌ Don'ts

• ❌ Set threshold too high (blocks valid queries)
• ❌ Set threshold too low (wastes resources)
• ❌ Ignore user feedback on misclassification
• ❌ Use generic responses (poor UX)
• ❌ Disable monitoring

---

Integration with Stage 2

Queries that pass Stage 1 move to Stage 2: Query Evaluation:

---

Related Documentation

• Stage 2: Query Evaluation - Next stage
• Complete Pipeline - All stages overview

---

Stage 1 ensures that only valid business queries consume expensive processing resources, optimizing both performance and cost.