​

Topics Explorer

​

Why Visual Exploration Matters

• Hidden relationships - Discover which topics are semantically related and should be handled similarly
• Quality validation - Verify that conversations within each topic are actually similar
• Volume drivers - Identify dense topic clusters that represent your highest-impact areas
• Emerging patterns - Spot new conversation clusters that may need their own topics
• Classification accuracy - Find outlier conversations that might be misclassified
• Strategic opportunities - Visualize where improving one topic could benefit related areas

​

Accessing the Topics Explorer

1. Navigate to Analyze > Topics
2. Ensure TopicAI is activated and has processed your conversations
3. Click the Open 3D Visualization button (sparkle icon)
4. The visualization loads in full-screen view

​

Understanding the 3D Visualization

​

What You’re Seeing

• Position is determined by the conversation’s semantic embedding
• Conversations about similar topics cluster together naturally
• Distance between points indicates how similar the conversations are
• Colors indicate topic assignments from TopicAI

• Your conversations exist as high-dimensional vectors (embeddings)
• These are reduced to 3D coordinates while preserving relationships
• Closer points = more semantically similar conversations
• The absolute position doesn’t matter - only relative distances

• Each topic has a unique color
• The legend on the left shows which color represents which topic
• Only topics with assigned conversations appear in the legend
• Hover over any point to see a conversation preview

​

Reading the Patterns

• Conversations are very similar to each other
• The topic is well-defined and coherent
• Your AI likely has consistent knowledge for these

• This is good - your topic model is working well here
• Use these as benchmarks for other topics
• Ensure your knowledge base covers the common patterns

• Two or more topics have blurry boundaries
• Conversations could reasonably belong to multiple topics
• Topics may be too granular or naturally related

• Consider merging related topics if they always overlap
• Check if the topics represent different aspects of the same issue
• Refine topic descriptions to clarify boundaries

• Topic covers diverse conversation types
• May be a catch-all category like “Other” or “General Questions”
• Conversations might not have strong common patterns

• Review scattered topics to see if they should be split
• Look for sub-clusters that deserve their own topics
• Consider if this represents genuinely diverse customer needs

• Unique conversations that don’t fit typical patterns
• Could be miscategorized
• Might represent edge cases or new issues

• Click outliers to read them
• Verify they’re in the correct topic
• Consider if they indicate a new topic emerging
• Check if they’re data quality issues (spam, test conversations)

​

Navigating the 3D Space

​

Basic Controls

• Click and drag: Rotate the view in any direction
• Scroll wheel: Zoom in and out
• Click a point: Open that conversation in a side panel
• Hover over a point: See a text preview of the conversation

• Spacebar: Toggle automatic rotation on/off
• Press spacebar again to stop the rotation

• Click “Toggle Orbiting” in the top-right corner
• Starts automatic rotation to see all angles
• Useful for presentations or discovering hidden patterns
• Click again or press spacebar to stop

​

Exploration Strategies

1. Zoom out to see all conversations at once
2. Identify major cluster locations and sizes
3. Note which topics dominate the landscape
4. Look for unexpected patterns or separations

1. Rotate to find a clear view of a specific cluster
2. Scroll to zoom in close
3. Hover over individual points to preview conversations
4. Click points to read full conversations in the side panel

1. Find two topics you think should be similar
2. Rotate to see both clusters clearly
3. Check if they’re positioned near each other
4. If they’re far apart, they represent very different issues
5. If they overlap, consider merging them

1. Identify a tight cluster for a well-defined topic
2. Look for same-colored points far from the main cluster
3. Click these outliers to read them
4. Verify whether they truly belong to this topic
5. Note patterns in misclassifications to improve topic descriptions

​

Using Topics Explorer for Strategic Insights

​

Identifying Volume Drivers

1. Zoom out to see all conversations
2. Identify the areas with the most dots packed together
3. Check the legend to see which topic that color represents
4. Estimate relative size by visual density

• You’re handling this well at scale
• Document what’s working to apply elsewhere
• Monitor to ensure performance stays high as volume grows

• This is your highest-impact improvement opportunity
• Even small percentage improvements affect many customers
• Prioritize knowledge additions for this topic
• Consider if this indicates a product issue that needs fixing

• Filter to this topic in the main Topics dashboard
• Check resolution rate and CSAT
• If low, this is urgent - many customers are having poor experiences
• Add comprehensive shipping delay knowledge
• Consider product improvements (better tracking, proactive notifications)
• Monitor cluster size in future visualizations to measure impact

​

Finding Emerging Topics

1. Look for groups of 5-10 points that cluster together
2. Check if they’re all the same color (same topic)
3. Look for clusters positioned between existing major topics
4. Identify scattered points that form loose patterns

• Might be a sub-topic worth splitting out
• Click several points to read the conversations
• Determine if they represent a distinct pattern
• Create a new topic if they’re strategically important to track separately

• Emerging pattern not yet well-categorized
• Wait for next Monday’s automatic topic detection
• Or manually trigger “Find New Topics” from the Topics page
• System may identify this as a new topic

• Create a new topic: “Two-Factor Authentication”
• Update topic descriptions to distinguish from general account issues
• This lets you track adoption problems separately
• Monitor if the cluster grows (indicating more issues with the feature)

​

Spotting Declining Topics

1. Take screenshots of the current visualization
2. Note the size and density of major clusters
3. Return weekly or monthly to compare
4. Look for clusters that are visibly smaller or less dense

1. Use the Topics dashboard to identify “Declining” topics
2. Find those topics in the 3D visualization
3. Check if the cluster looks sparse or has fewer points than expected
4. Determine if this is good (problem solved) or bad (customers gave up)

• You added knowledge and resolution improved
• A product bug was fixed
• Seasonal issue passed
• Customers are successfully self-serving

• Low CSAT in this topic (they stopped asking because they’re frustrated)
• Conversations moved to another channel (not captured)
• Customers switching to competitors
• Unresolved issues they gave up on

• Resolution rate increased from 45% to 82%
• CSAT improved from 2.1 to 4.3
• Conversation volume decreased by 67%

​

Discovering Topic Clusters and Relationships

1. Identify a topic cluster you care about
2. Rotate the view to see what other clusters are nearby
3. Note which topics appear in the same “neighborhood”
4. Click conversations at the boundary between clusters

• Related topics can share similar knowledge
• Create a data provider that serves multiple related topics
• Build comprehensive guides that address topic clusters together

• Group related topics for support team training
• Create escalation paths that route related topics to the same specialists
• Build combo solutions that address multiple related issues

• Nearby topics might represent sequential customer experiences
• Example: “Payment Failed” → “Refund Request” → “Subscription Cancellation”
• Improve upstream topics to prevent downstream issues

• “Wrong Item Received” (green)
• “Damaged Package” (yellow)
• “Return Request” (orange)

• Create a comprehensive “Order Issues” knowledge base
• Include information about returns, exchanges, and refunds in all three contexts
• Train your AI to proactively offer return labels when order problems are detected
• Monitor all three topics together as “Fulfillment Quality”
• If one grows, check if the others also grow (systemic issue)

​

Correlating Topics with Performance

1. From Dashboard to Visualization:
   • On the Topics page, note topics with low resolution rates (red on treemap)
   • Find those topics in the 3D visualization
   • Check cluster characteristics (tight vs. scattered, size, position)
2. Pattern Recognition:
   • Do all low-performing topics have scattered clusters?
   • Are high-performing topics always tight clusters?
   • Are certain topic neighborhoods associated with better outcomes?
3. Draw Conclusions:
   • Scattered = diverse conversations = harder to serve with one knowledge base
   • Tight = consistent conversations = easier to optimize with targeted knowledge
   • Topic relationships may reveal systemic issues

• Split broad topics into more specific categories
• Example: “Product Questions” → “Size/Fit Questions,” “Material Questions,” “Compatibility Questions”
• Each new topic will have a tighter cluster and be easier to optimize

• Investigate what changed recently that would affect this area
• Check if a data provider was modified or disabled
• Look for product releases that touched this functionality
• Fix the root cause to improve all three topics at once

​

Practical Use Cases

​

Product Launch Tracking

1. Check if a new topic emerged related to the feature
2. Look for small clusters in unexpected positions (confusion about the feature)
3. Identify which existing topics the feature conversations cluster near
4. Click through conversations to understand common questions

• Customers asking for filters don’t realize the feature exists (poor discoverability)
• Some customers finding the feature struggle to use it (technical issues cluster nearby)
• The feature isn’t in “Usage Questions” cluster, suggesting it’s not part of normal workflow yet

• Improve feature announcement and in-app discovery
• Add knowledge about common advanced filter use cases
• Consider UX improvements to make the feature more intuitive
• Monitor if the cluster moves toward “Usage Questions” as adoption improves

​

Seasonal Trend Identification

1. Take screenshots of the visualization at different times of year
2. Compare cluster sizes and positions across seasons
3. Identify which topics swell during peak season
4. Prepare targeted knowledge before the season hits

• “Gift Wrapping Requests” cluster: Large in November, tiny in January
• “Return Requests” cluster: Moderate in November, massive in January
• “Shipping Questions” cluster: Huge in both months

• Beef up “Gift Wrapping” knowledge in October
• Prepare comprehensive “Holiday Returns” content in December (before January spike)
• Ensure “Shipping” knowledge includes holiday deadlines and delays
• Staff support team according to expected topic volumes

​

Feature Request Prioritization

1. Find the “Feature Requests” topic cluster
2. Look for sub-clusters within it (requests about related features)
3. Measure relative density to estimate demand
4. Check proximity to “Technical Issues” or “Workarounds” clusters

• 45 conversations about “Bulk Actions” (dense sub-cluster)
• 12 conversations about “Dark Mode” (loose grouping)
• 8 conversations about “Export to PDF” (tight small cluster)

1. Build “Bulk Actions” first - highest demand (45 requests), customers actively seeking workarounds
2. Consider “Export to PDF” - smaller volume but tight cluster indicates consistent need
3. Wait on “Dark Mode” - loosely grouped suggests it’s not a burning need

​

Support Strategy Optimization

1. Cross-reference resolution data with cluster characteristics
2. Tight clusters with low resolution = knowledge gaps (automatable)
3. Scattered clusters with low resolution = inherently complex (needs humans)
4. Clusters near sensitive topics (billing, legal) = consider human oversight

• Tight cluster (similar conversations)
• Low resolution rate (35%)
• Positioned between “Billing Questions” and “Subscription Cancellation”

• Scattered cluster (diverse questions)
• Low resolution rate (18%)
• Positioned near “Data Privacy” and “Terms of Service”

• Tight cluster = automatable
• Add comprehensive refund policy knowledge
• Build AI flow: Check order status → Verify refund eligibility → Process automatically
• Goal: 80%+ autonomous resolution

• Scattered cluster = needs expertise
• Set up automatic escalation to legal team
• AI provides initial information but always loops in human
• Goal: Fast routing, not automation

​

Best Practices

​

Regular Exploration Cadence

• Open the visualization
• Zoom out to see the full landscape
• Look for new clusters or dramatic size changes
• Check if anything looks unusual compared to last week
• Screenshot for future comparison

• Explore each major cluster individually
• Click through 5-10 conversations per topic to validate categorization
• Look for opportunities to split or merge topics
• Identify outliers and misclassifications
• Compare density with performance metrics from dashboard

• Compare visualizations across the quarter
• Identify which clusters grew, shrank, or shifted
• Correlate changes with product releases and knowledge updates
• Present findings to product and support leadership
• Plan next quarter’s optimization priorities

​

Combine with Dashboard Analytics

1. Start with the Dashboard to identify topics that need attention:
   • Low resolution rates (red on treemap)
   • Growing trends (green badge in table)
   • High escalation rates (sankey chart)
2. Move to Topics Explorer to understand why:
   • Is the cluster tight (knowledge gap) or scattered (inherently complex)?
   • Are there sub-clusters that should be separate topics?
   • How does this topic relate to others spatially?
3. Click Through Conversations to validate your hypothesis:
   • Read 5-10 conversations from the cluster
   • Check outlier points
   • Verify that your interpretation makes sense
4. Take Action based on combined insights:
   • Add knowledge for tight clusters with low resolution
   • Split scattered clusters into more specific topics
   • Merge overlapping clusters that represent the same need
   • Set up escalations for inherently complex scattered clusters
5. Return to Dashboard to measure impact:
   • Did resolution rates improve?
   • Did the topic trend change to “Declining”?
   • Did related topics also improve?

​

Document Your Findings

• Screenshots with annotations showing what you observed
• Date of the visualization (from the timestamp)
• Cluster descriptions (which topics are where, relative sizes)
• Hypotheses about what patterns mean
• Actions taken based on what you discovered
• Results after implementing changes

• Project wiki or documentation system
• Shared folder with monthly visualization screenshots
• Issue tracker (link actions to visual findings)
• Presentation deck for stakeholder updates

Topics Explorer Review - 2024-11-15

Key Finding: "API Authentication" cluster split into two sub-groups
- Group A (15 conversations): OAuth token issues, tight cluster
- Group B (8 conversations): API key setup questions, looser grouping

Hypothesis: These are two different failure modes requiring different solutions

Actions:
- Created new topic: "OAuth Token Issues"
- Updated "API Authentication" to focus on setup only
- Added detailed OAuth troubleshooting guide

Follow-up: Check visualization 2024-12-01 to see if clusters are clearer

​

Use Color and Space Together

• Confirms or challenges your expectations
• Shows which conversations the system thinks are similar
• Reveals potential misclassifications (same color, far apart)

• Distance = semantic similarity (regardless of topic assignment)
• Nearby points = similar conversations (even if different colors)
• Far points = different conversations (even if same color)

• Same-color points cluster tightly together
• Different-color clusters are far apart
• This indicates good topic model quality

• Same-color points scattered across the space = topic too broad
• Different-color points mixed together = topics too granular
• Same-color outliers far from main cluster = misclassifications

• Click the point to read the conversation
• If it’s indeed about delivery issues, the categorization might update next week
• If this happens frequently, refine topic descriptions to clarify boundaries
• Consider if “Shipping Questions” and “Delivery Issues” should be merged

​

Share Insights with Your Team

• Show feature request clusters to visualize demand
• Demonstrate how product changes affected topic volumes
• Identify pain points that should be fixed in the product itself

• Reveal which topics cluster together (good for training sessions)
• Show the scale of different issue types visually
• Identify complex scattered topics that need specialized expertise

• Demonstrate what customers actually talk about (not what you assume)
• Show improvement over time (comparison screenshots)
• Visualize strategic opportunities (large clusters with low resolution)

• Start zoomed out to show the full landscape
• Zoom into specific clusters to tell stories
• Use automatic rotation to show all angles
• Prepare 3-5 key insights rather than trying to cover everything
• Connect visual findings to business metrics (revenue impact, customer satisfaction)

​

Troubleshooting

​

Visualization Won’t Load

• Blank screen where visualization should appear
• Error message about missing data
• Infinite loading spinner

• Check if TopicAI is enabled for your project
• See Topics activation guide

• You need at least 20-30 conversations with substantial content
• Wait for nightly topic analysis to generate visualization data
• Check the timestamp to see when visualization was last updated

• Use Chrome, Edge, or Safari (recommended)
• Update your browser to the latest version
• Disable browser extensions that might interfere with WebGL
• Check if WebGL is enabled in browser settings

• Initial visualization generation can take several minutes
• Check back in 10-15 minutes after activating topics
• Manually trigger “Categorize Now” and wait for completion

​

Visualization is Laggy or Slow

• Slow rotation or zoom
• Choppy animation
• Browser becomes unresponsive

• Close other browser tabs and applications
• Disable automatic rotation (press spacebar)
• Zoom out less frequently (rendering many points is expensive)

• Topics Explorer requires decent GPU for 3D rendering
• Older computers may struggle with large datasets (1000+ conversations)
• Try on a different device if performance is unusable

• Use date range filters on the Topics page before opening visualization
• This reduces the number of points rendered
• Focus on recent data (last 30-60 days) for better performance

• Old cached data might cause issues
• Clear cache and reload the page
• Close and reopen the browser

​

Can’t Click Points or Hover Doesn’t Work

• Clicking conversations doesn’t open the side panel
• Hovering doesn’t show previews
• Controls don’t respond

• Initial rendering can take a few seconds
• Wait until rotation completes before clicking
• Try clicking again after the view settles

• Some points might be behind others
• Rotate to get a better angle
• Zoom in to spread points apart

• The ScatterGL library can sometimes get into a bad state
• Reload the page to reset
• If it persists, clear cache and reload

• Open browser developer tools (F12)
• Check console for JavaScript errors
• Report persistent errors to support with screenshots

​

Colors Don’t Match Topics

• Legend shows different colors than what you see
• Multiple topics appear to have the same color
• Colors changed since last time

• Colors are generated based on which topics have conversations
• If topics are added or removed, colors may shift
• This is normal behavior

• If you have many topics (30+), some colors may be similar
• Focus on cluster position and density, not just color
• Consider merging very similar topics to reduce color overlap

• If you deleted or merged topics, colors will regenerate
• This won’t affect spatial positioning (positions are based on content, not topics)

• The same point changes color when you refresh
• Legend colors don’t match any points in the visualization
• In these cases, contact support

​

Missing Recent Conversations

• Conversations from today or this week don’t appear
• Visualization seems outdated
• Timestamp is several days old

• Conversations need to be categorized first (happens nightly)
• Visualization is rebuilt as part of that process
• Check the timestamp at the bottom to see when it was last updated

1. Go back to the Topics page
2. Click “Categorize Now” to manually trigger analysis
3. Wait a few minutes for processing to complete
4. Return to Topics Explorer
5. Reload the page to see the updated visualization

• Automatic updates happen every night
• If you need real-time data, the conversation list and dashboard update immediately
• The 3D visualization is for strategic analysis, not real-time monitoring

​

Cluster Doesn’t Make Sense

• Conversations in a cluster don’t seem related
• Topic assignments seem random
• Outliers in unexpected positions

• Some conversations cover multiple topics
• They may appear between topic clusters
• This reflects the reality that customers have complex questions

• Topics are refined weekly
• Recent conversations may shift as the model improves
• Check if conversations started before last Monday (assignments are final)

• The AI sees patterns humans might miss
• Conversations that seem different might share underlying structure
• Click through to understand what the AI detected

• Click outlier points to read them
• If many are clearly wrong, update topic descriptions
• More specific descriptions improve future categorization

• If you provided poor seed topics, the whole model might be off
• Consider deleting the topic model and starting over
• Use industry templates or let the system auto-detect

• Spam, test conversations, or imported garbage can create weird clusters
• Filter these out or delete them from your project
• They’ll disappear from the visualization at the next rebuild

​

Next Steps