Skip to main content

Topic Collision Analysis

Use Einstein AI to detect and resolve overlapping topics that could cause message routing issues.

The Problem​

As your agent implementation grows with more topics, ensuring they don't overlap or conflict becomes increasingly complex.

Teams need to:

  • 🎯 Prevent Routing Confusion: Ensure each user message is routed to the correct, most specific topic
  • 🔍 Identify Overlaps: Detect when multiple topics could handle the same query
  • ⚖️ Maintain Consistency: Provide predictable, reliable responses to users
  • 📊 Scale Confidently: Add new topics without breaking existing routing
  • 🎨 Optimize Scope: Define clear boundaries for each topic's responsibility
  • ✅ Validate Changes: Test topic updates before deploying to production

In short: You need a way to ensure your topics are orthogonal (non-overlapping) so users get routed to the right place every time.

How GenAI Explorer Solves This​

GenAI Explorer uses Einstein AI analysis to:

✅ Automatic Detection: Einstein analyzes all topic pairs and identifies overlaps with precision scores

✅ Overlap Scoring: See exactly how much topics overlap (0-100%) and why

✅ Root Cause Analysis: Einstein explains why topics collide with clear reasoning

✅ Specific Resolutions: Get 2-3 actionable suggestions per collision (merge, refine scope, add exclusions)

✅ Visual Comparison: See overlapping vs orthogonal topics side-by-side with diagrams

✅ Resolution Strategies: Learn 5 proven approaches to fix collisions (scope refinement, merging, exclusions, splitting, hierarchy)

Impact: Achieve 95%+ topic classification accuracy, eliminate routing inconsistencies, and build reliable agents users can trust.

Data Cloud Integration

Overview​

The Topic Collision Analysis feature uses Einstein AI to detect when multiple topics in a GenAI Planner have overlapping scopes. Since topics are used to classify incoming messages, they must be orthogonal (non-overlapping) to ensure proper message routing and classification.

Why This Matters​

When topics have overlapping scopes:

  • Classification Ambiguity: A single user message could match multiple topics
  • Routing Failures: The system cannot determine which topic should handle the message
  • Inconsistent Behavior: Similar messages may be routed to different topics unpredictably
  • Poor User Experience: Users receive incorrect or incomplete responses

How It Works​

1. Analysis Trigger​

When viewing a GenAI Planner, you'll see an alert prompting you to analyze topics:

Topic Scope Analysis
Use Einstein to analyze topic scopes for potential collisions. 
Topics must be orthogonal to properly classify messages.
[Analyze Button]

2. Einstein Analysis Process​

When clicked, the system:

  1. Gathers Topic Information: Collects all topic metadata including:

    • Master Label and Developer Name
    • Description
    • Scope definition
    • Plugin Type

  2. Calls Einstein API: Sends a structured prompt asking Einstein to:

    • Compare each pair of topics
    • Calculate overlap scores (0.0 to 1.0)
    • Identify why topics overlap
    • Suggest specific resolutions

  3. Displays Results: Shows a comprehensive analysis with actionable recommendations

3. Analysis Results​

No Collisions Detected ✓​

✓ No Topic Collisions Detected
All topics have orthogonal scopes and should classify messages correctly.

Collisions Detected ⚠️​

When collisions are found, the system displays:

For each collision:

  • Topic Pair: Shows which two topics overlap
  • Overlap Score: Percentage indicating severity (e.g., "75% overlap")
  • Overlap Reason: Detailed explanation from Einstein
  • Suggested Resolutions: 2-3 specific actions to resolve the collision

Example:

⚠️ Topic Scope Collisions Detected
Found 2 potential collision(s). Topics with overlapping scopes may cause classification failures.

┌─────────────────────────────────────────────────────────────┐
│ [Customer Support] overlaps with [Product Inquiries] 75%    │
│                                                              │
│ Reason: Both topics handle customer questions about         │
│ product features and troubleshooting.                       │
│                                                              │
│ Suggested Resolutions:                                      │
│ • Narrow Customer Support to only handle technical issues   │
│ • Limit Product Inquiries to pre-sales questions only       │
│ • Merge both into a comprehensive "Customer Service" topic  │
└─────────────────────────────────────────────────────────────┘

[Re-analyze Button]

Resolution Strategies​

Einstein typically suggests these types of resolutions:

1. Scope Refinement​

Narrow the focus of one or both topics to eliminate overlap.

Example:

  • Original: "Customer Support" (handles all customer questions)
  • Refined: "Technical Support" (handles only technical issues)

2. Topic Merging​

Combine overlapping topics into a single, comprehensive topic.

When to use:

  • Topics have >70% overlap
  • Separating them adds no value to users
  • They share the same underlying knowledge base

3. Exclusion Criteria​

Add explicit boundaries to topic scopes.

Example:

  • Topic A: "Product Questions" (excludes pricing)
  • Topic B: "Pricing Inquiries" (only handles pricing)

4. Topic Splitting​

Break down overly broad topics into more specific ones.

Example:

  • Original: "General Inquiries" (too broad)
  • Split into:
    • "Account Questions"
    • "Billing Questions"
    • "Technical Questions"

5. Hierarchy Creation​

Establish parent-child relationships between topics.

Example:

  • Parent: "Customer Service"
    • Child: "Returns & Refunds"
    • Child: "Order Status"
    • Child: "Product Support"

Best Practices​

1. Regular Analysis​

  • Run analysis after adding new topics
  • Re-analyze when modifying topic scopes
  • Check periodically as part of planner maintenance

2. Topic Design Guidelines​

DO:

  • ✓ Create specific, focused topics
  • ✓ Use clear, descriptive scope definitions
  • ✓ Document exclusions and boundaries
  • ✓ Test with sample messages

DON'T:

  • ✗ Create overly broad "catch-all" topics
  • ✗ Use vague scope descriptions
  • ✗ Ignore overlap warnings
  • ✗ Duplicate functionality across topics

3. Testing Recommendations​

After resolving collisions:

  1. Test with real user messages
  2. Monitor classification accuracy
  3. Review routing patterns
  4. Adjust scopes based on actual usage

User Workflow​

  1. Navigate to a GenAI Planner in SF Explorer
  2. Click the "Analyze" button in the Topic Scope Analysis section
  3. Review any detected collisions
  4. Implement suggested resolutions:
    • Edit topic scopes in Salesforce
    • Merge or split topics as needed
    • Update descriptions and instructions
  5. Re-analyze to verify fixes
  6. Test with sample messages to confirm proper routing

Technical Details​

API Integration​

The feature uses Einstein's LLM Completions API:

const einsteinEndpoint = `/services/data/v60.0/einstein/llm/completions`

const requestBody = {
   prompt: analysisPrompt,
   model: "sfdc_ai__DefaultGPT4Omni",
   maxTokens: 2000,
   temperature: 0.3, // Low temperature for consistent analysis
   responseFormat: { type: "json_object" }
}

Fallback Analysis​

If Einstein API is unavailable, the system performs basic text similarity analysis:

  • Compares topic names, descriptions, and scopes
  • Calculates Jaccard similarity coefficient
  • Suggests generic resolutions

This ensures the feature remains functional even without Einstein access.

Troubleshooting​

Issue: False Positives​

Symptom: Topics flagged as overlapping but shouldn't be

Solution:

  • Review Einstein's reasoning
  • Check if scope descriptions are unclear
  • Add more specific scope details
  • Re-run analysis

Issue: Einstein API Errors​

Symptom: Analysis fails with API error

Solution:

  • Fallback analysis will run automatically
  • Check Einstein API permissions
  • Verify org has Einstein features enabled
  • Contact Salesforce support if persistent

Issue: No Collisions Detected (But Problems Exist)​

Symptom: Analysis shows no issues but messages route incorrectly

Solution:

  • Check topic instructions (not just scopes)
  • Review actual message classification logs
  • Test with specific problematic messages
  • Consider semantic overlaps Einstein might miss

Next Steps​