Skip to main content
Message Search is your most powerful tool for understanding exactly what users and your AI are saying. Unlike conversation-level analysis, message search lets you drill down into individual messages across all conversations, helping you discover knowledge gaps, validate AI responses, and identify patterns in how users phrase their questions.

​
Why Message Search Matters

Every message in your system contains valuable signals about your AI’s performance. Message search helps you:
  • Find knowledge gaps - Discover when your AI says “I don’t know” or provides vague answers
  • Validate accuracy - Search for specific product information to verify your AI is giving correct answers
  • Understand user language - See how real users phrase questions to improve your knowledge base
  • Track tool usage - Find when specific tools or actions are used (or should have been)
  • Identify patterns - Spot repeated phrases that indicate systematic issues
  • Export insights - Extract data for training, compliance, or quality assurance
Message search is especially valuable during the first few weeks after launch. Search for phrases like “I don’t have information” or “I’m not sure” to quickly identify and fill knowledge gaps.
Message search scans the content of individual messages, not entire conversations. This distinction is important: Conversation Search (in Analyze → Conversations)
  • Filters by conversation metadata (status, rating, topic, channel)
  • Shows complete conversation threads
  • Best for understanding context and user journeys
Message Search (in Analyze → Message Search)
  • Searches actual message text content
  • Filters by message-level attributes (sentiment, completeness, sender)
  • Shows individual messages with conversation context
  • Best for finding specific phrases, patterns, or information
Use message search when you know what you’re looking for but don’t know which conversations contain it.

​
Full-Text Search Capabilities

The search bar supports powerful full-text search syntax to help you find exactly what you need. Type any word or phrase to find messages containing that text: 
refund
Finds all messages mentioning “refund”

​
AND Operator (Implicit)

Multiple words without operators are treated as AND - all words must appear: 
refund policy
Finds messages containing both “refund” AND “policy” (in any order)

​
OR Operator

Search for messages containing any of several terms: 
refund or cancellation
Finds messages mentioning either “refund” OR “cancellation” (or both)

​
NOT Operator (Exclusion)

Exclude messages containing specific terms using the minus sign: 
refund -approved
Finds messages about “refund” but NOT containing “approved” Use quotes to search for exact phrases: 
"annual subscription"
Finds only messages with the exact phrase “annual subscription”

​
Complex Queries

Combine operators for sophisticated searches: 
"password reset" -email
Finds exact phrase “password reset” excluding mentions of “email” 
(billing or invoice) payment
Finds messages about payment that also mention billing or invoice
The search interface includes a help icon (ℹ️) next to the search bar with quick examples of these search operators.

​
Advanced Filters

Beyond text search, you can filter messages by numerous attributes to narrow your results.

​
Date Range

Filter messages by when they were sent:
  • Last 7 days - Recent conversations and emerging issues
  • Last 30 days - Monthly patterns and trends
  • Last 90 days - Quarterly analysis
  • Custom range - Any specific date range for targeted analysis
When investigating a specific deployment or knowledge base change, set the date range to start just before the change. This helps you compare “before” and “after” message patterns.

​
Channel Filter

Filter by the source channel of messages:
  • Website - Messages from your web widget
  • Zendesk - Messages from Zendesk integration
  • Salesforce - Messages from Salesforce integration
  • Slack - Messages from Slack integration
  • WhatsApp - Messages from WhatsApp integration
Use channel filters to:
  • Compare AI performance across channels
  • Find channel-specific issues
  • Analyze if certain channels have different types of questions

​
Rating Filter

Filter by the CSAT rating of the conversation containing the message:
  • Abandoned (0) - User didn’t respond to survey
  • Terrible (1) - Very dissatisfied
  • Bad (2) - Dissatisfied
  • OK (3) - Neutral
  • Good (4) - Satisfied
  • Amazing (5) - Very satisfied
  • Unoffered - No survey presented
Use Cases:
  • Search AI messages in Terrible/Bad rated conversations to see what went wrong
  • Find user questions in Amazing rated conversations to identify what works well
  • Look at Unoffered conversations to understand when surveys aren’t triggered

​
Completeness Filter

Filter by how thoroughly the AI answered the user’s question:
  • Complete - Full, comprehensive answer provided
  • Partial - Some information provided but incomplete
  • No Answer - Unable to provide relevant information
Use Cases: 
Search: "I don't have"
Filter: Completeness = No Answer
Result: All instances where AI explicitly couldn't answer
Combine completeness with text search to understand gaps:
  • Complete answers about “pricing” → Validate information accuracy
  • No Answer about “integrations” → Missing knowledge gap
  • Partial answers about “API” → Need more comprehensive documentation

​
Sentiment Filter

Filter by the emotional tone detected in messages:
  • Positive - User expresses satisfaction, happiness, appreciation
  • Neutral - Factual or emotionally neutral
  • Negative - User expresses frustration, anger, dissatisfaction
Use Cases:
  • Negative sentiment in user messages → Find frustration points
  • Negative sentiment after AI responses → AI may be making things worse
  • Positive sentiment → Identify what’s working well
Track sentiment progression within conversations: 
Search: "thank you"
Filter: Sentiment = Positive
Result: Successful resolution patterns

​
Language Filter

Filter by the detected language of the message:
  • English
  • Spanish
  • German
  • French
  • And more…
Use Cases:
  • Verify your AI handles non-English queries properly
  • Find translation issues or language-specific problems
  • Analyze volume by language to prioritize localization efforts
  • Export messages in specific languages for native speaker review
Language detection is automatic but may occasionally misidentify short messages or mixed-language content. Use this filter for general segmentation rather than precision filtering.

​
File Types Filter

Filter messages that include specific file types:
  • Image - Messages with image attachments
  • PDF - Messages with PDF documents
  • Document - Word, Excel, PowerPoint files
  • Video - Video file attachments
  • Audio - Audio file attachments
Use Cases:
  • Find messages where users shared screenshots of errors
  • Identify documentation requests (PDFs)
  • Track support cases with visual evidence
  • Analyze how AI handles messages with attachments

​
Visited Pages Filter

Filter messages by the webpage URL where the conversation occurred (website channel only): Enter full URLs or partial paths: 
https://example.com/pricing
https://example.com/docs/api
/checkout
Use Cases:
  • Find all questions asked on your pricing page
  • Analyze support needs for specific product pages
  • Identify pages where users struggle most
  • Compare question types across different sections of your site
Example Workflow:
  1. Notice high bounce rate on /features/integrations page
  2. Search messages from that page: pages = /features/integrations
  3. Discover users frequently ask “Does this work with Salesforce?”
  4. Add clear Salesforce integration information to that page

​
Handoff Filter

Filter by whether the conversation was escalated to human support:
  • Offered - AI offered handoff to user
  • Accepted - User accepted the handoff offer
  • Requested - User explicitly asked for human help
  • None - No handoff occurred
Use Cases: 
Filter: Handoff = Requested
Result: See what triggers users to ask for humans

Filter: Handoff = Offered
Search: (empty to see all)
Result: When AI proactively offered escalation
Combine with text search: 
Search: "speak to a person"
Filter: Handoff = None
Result: Cases where users asked for humans but weren't escalated

​
Labels Filter

Filter messages from conversations with specific labels: Labels are custom tags you create for categorization. The label filter has three states: Has Label (+ icon, green)
  • Show messages from conversations WITH this label
  • Example: Has “billing-dispute” → Only billing dispute messages
Does Not Have Label (− icon, red)
  • Show messages from conversations WITHOUT this label
  • Example: Not “resolved” → Exclude already-resolved cases
Unselected (gray)
  • No filtering on this label
Clicking Behavior: Click the label name to cycle through states:
  1. Unselected → Has
  2. Has → Does Not Have
  3. Does Not Have → Unselected
Or click the + or − buttons directly to toggle each state independently. Use Cases: 
Labels: Has "technical-support", Does Not Have "resolved"
Result: Active technical support cases

Labels: Has "bug-report"
Search: "workaround"
Result: Bug reports where users found workarounds

​
Combining Filters

The real power comes from combining multiple filters: Find Knowledge Gaps 
Text: "I don't have information"
Completeness: No Answer
Date Range: Last 30 days
Result: Recent knowledge gaps to prioritize
Validate Pricing Information 
Text: "pricing" or "cost" or "price"
Rating: Good, Amazing
Channel: Website
Result: Successful pricing discussions to use as templates
Identify Escalation Triggers 
Sentiment: Negative
Handoff: Requested
Language: English
Result: What makes English-speaking users escalate
Analyze Product Launch 
Text: "new feature"
Date Range: Last 7 days
Completeness: No Answer
Result: Questions about new features AI can't answer yet

​
Active Filters Display

All active filters appear as removable chips below the search bar. Click any chip to remove that specific filter, or click “Clear All” to reset everything. This makes it easy to:
  • See exactly what filters are applied
  • Quickly adjust your search
  • Remove filters one at a time to broaden results
  • Save time by not re-opening filter menus

​
Working with Search Results

​
Message Cards

Search results display as cards showing:
  • Message content - The actual text of the message
  • Sender - User or AI (with AI model indication)
  • Conversation context - Brief preview of surrounding messages
  • Metadata - Timestamp, channel, sentiment, rating
  • Conversation link - Click to view full conversation

​
Infinite Scroll

Results load progressively as you scroll down. This allows you to:
  • Browse hundreds or thousands of results efficiently
  • Find patterns across large datasets
  • Load only what you need (faster performance)

​
Selection Mode

Select individual messages for batch operations:
  1. Click the checkbox on any message card to enter selection mode
  2. Select multiple messages by clicking additional checkboxes
  3. Batch toolbar appears showing selected count
  4. Available actions:
    • Add label to all selected conversations
    • Export selected messages
    • Mark conversations as resolved
Keyboard Shortcut: Press ESC to clear all selections and exit selection mode.
Use selection mode to quickly label similar messages. For example, select 10 messages about “API authentication” and apply the “api-docs-needed” label to track knowledge gaps.

​
Exporting Search Results

Click the Export button to download your search results with all active filters applied.

​
Export Options

CSV Format
  • One row per message
  • Columns: message ID, content, sender, timestamp, conversation ID, sentiment, rating, etc.
  • Best for spreadsheet analysis, reporting, sharing with stakeholders
JSON Format
  • Complete message objects with all metadata
  • Best for programmatic analysis, importing to other systems

​
What Gets Exported

Exports include:
  • All messages matching your current search and filters
  • Message metadata (sender, timestamp, sentiment, etc.)
  • Conversation context (ID, rating, status, channel)
  • User information (anonymized if configured)
  • Message tags and labels
Exports respect your current filters. If you have 10,000 total messages but filter to 100 about “refunds”, only those 100 will export. This makes exports precise and manageable.

​
Export Use Cases

Quality Assurance 
Filter: Rating = Terrible, Bad
Export: CSV
Use: Share with team for review and improvement planning
Compliance and Auditing 
Filter: Date Range = Last Quarter
Export: JSON
Use: Archive for compliance requirements
Training Data 
Filter: Rating = Amazing, Completeness = Complete
Export: CSV
Use: Extract high-quality conversations to train new team members
Knowledge Base Creation 
Search: (topic-specific keywords)
Filter: Completeness = Complete
Export: CSV
Use: Extract AI answers to validate and convert to documentation
Bug Tracking 
Labels: Has "bug-report"
Export: CSV
Use: Import to project management tool

​
Common Search Patterns

Learn these proven patterns for maximum productivity.

​
Finding Knowledge Gaps

Pattern 1: Direct Admissions 
Search: "I don't have information" or "I'm not sure" or "I don't know"
Filter: Sender = AI
Result: Explicit knowledge gaps
Pattern 2: Vague Responses 
Search: "general" or "typically" or "usually"
Completeness: Partial
Result: Answers that lack specificity
Pattern 3: External Search Usage 
Search: "search_web" or "According to my search"
Result: When AI used web search instead of internal knowledge

​
Validating AI Accuracy

Pattern 4: Fact-Checking Pricing 
Search: "$" or "price" or "cost"
Rating: Terrible, Bad
Result: Pricing information that led to dissatisfaction
Pattern 5: Product Claims 
Search: "supports" or "compatible with" or "integrates"
Completeness: Complete
Result: Verify compatibility claims are accurate
Pattern 6: Policy Verification 
Search: "policy" or "terms" or "conditions"
Channel: Zendesk
Result: Check if AI cites correct policies in support tickets

​
Understanding User Language

Pattern 7: Question Phrasing 
Search: "how do I" or "how can I" or "how to"
Sentiment: Neutral
Result: Common question structures users employ
Pattern 8: Problem Description 
Search: "not working" or "broken" or "error"
Sentiment: Negative
Result: How users describe technical issues
Pattern 9: Feature Discovery 
Search: "can you" or "is it possible" or "do you have"
Result: What features users are looking for

​
Tracking Tool Usage

Pattern 10: Handoff Effectiveness 
Search: "connect you" or "transfer"
Handoff: Offered
Result: How AI phrases handoff offers
Pattern 11: Search Tool Overuse 
Search: "search" (in AI messages)
Completeness: No Answer
Result: Cases where search didn't find internal knowledge
Pattern 12: Action Completion 
Search: "I've" or "I have" (in AI messages)
Sentiment: Positive
Result: Successful tool executions

​
Identifying Systematic Issues

Pattern 13: Repeated Escalations 
Search: "representative" or "human" or "agent"
Handoff: None
Result: Escalation requests that weren't handled
Pattern 14: Confusion Markers 
Search: "confused" or "unclear" or "I don't understand"
Rating: Bad, Terrible
Result: When AI causes confusion
Pattern 15: Positive Exceptions 
Search: "helpful" or "perfect" or "exactly"
Rating: Amazing
Result: What makes users very happy

​
Best Practices

​
Search Strategy

  1. Start broad, then narrow
    • Begin with simple keywords
    • Add filters progressively
    • Don’t over-filter initially or you might miss patterns
  2. Use OR for variations
    • Search: "refund" or "reimbursement" or "money back"
    • Captures different ways users express the same need
  3. Exclude noise
    • Search: billing -"thank you"
    • Find billing issues, exclude successful resolutions
  4. Validate with completeness
    • After finding AI answers, filter by Completeness = Complete
    • Ensures you’re looking at full responses, not partial attempts

​
Analysis Workflow

  1. Weekly Knowledge Gap Review 
    Schedule: Every Monday
Search: "I don't" or "I'm not sure"
Filter: Date Range = Last 7 days
Action: Create snippets for top 5 gaps
  2. Monthly Accuracy Check 
    Schedule: First day of month
Search: (product/pricing keywords)
Filter: Rating = Terrible, Bad
Action: Verify all facts are current
  3. Post-Deployment Validation 
    Timing: 3 days after each deployment
Search: (topics you changed)
Filter: Date Range = Since deployment
Action: Confirm improvements are working
  4. Quarterly Export and Archive 
    Schedule: End of each quarter
Search: (none - all messages)
Filter: Date Range = Last quarter
Export: JSON for archives
Action: Compliance and trend analysis

​
Performance Tips

Be Specific with Text Search
  • "password reset" (exact phrase) is faster than password alone
  • Reduces irrelevant results
  • Speeds up search processing
Use Date Ranges
  • Smaller date ranges search faster
  • Default to Last 30 days for most analyses
  • Expand only when looking for historical patterns
Combine Filters Before Searching
  • Set all filters first, then enter search text
  • Each search refinement triggers a new query
  • Setting filters together = one query instead of many
Export Strategically
  • Export only what you need
  • Use filters to reduce export size
  • Large exports (10,000+ messages) may take time

​
Organization Tips

Create Repeatable Search Links Bookmark frequently-used searches:
  1. Configure your search and filters
  2. Copy the URL (contains all filter parameters)
  3. Bookmark or save to documentation
  4. Share with team members
Example Bookmarks:
  • “Weekly Knowledge Gaps” → Saved URL with “I don’t” search + last 7 days
  • “Pricing Questions” → Saved URL with pricing keywords + website channel
  • “Escalation Triggers” → Saved URL with negative sentiment + handoff requested
Use Labels to Track Follow-Up When you find issues in message search:
  1. Click through to the conversation
  2. Apply a label like “needs-snippet” or “guidance-issue”
  3. Later, filter by that label to track progress
  4. Remove label when fixed
Document Common Patterns Keep a team document with:
  • Frequent searches your team runs
  • What each search reveals
  • Actions taken based on findings
  • Improvement metrics over time

​
Troubleshooting

​
No Results Found

Possible Causes:
  1. Too many filters active
    • Solution: Remove filters one by one to broaden search
    • Check Active Filters chips below search bar
  2. Overly specific search phrase
    • Solution: Try simpler keywords
    • Example: Change "I don't have that information" to "don't have"
  3. Date range too narrow
    • Solution: Expand date range
    • Try Last 90 days instead of Last 7 days
  4. Typos in search
    • Solution: Check spelling
    • Try alternative phrasings (refund vs reimbursement)

​
Too Many Results

Solutions:
  1. Add date filter - Focus on recent messages first
  2. Use exact phrases - Add quotes around multi-word searches
  3. Add completeness filter - Focus on Complete or No Answer only
  4. Filter by channel - Narrow to specific integration
  5. Use NOT operator - Exclude common but irrelevant terms

​
Search is Slow

Performance Tips:
  1. Narrow date range - Search last 30 days instead of all time
  2. Add more filters - Reduce dataset before searching
  3. Use specific search terms - Avoid single common words like “the” or “is”
  4. Check your connection - Slow network affects load times

​
Wrong Messages in Results

Check These:
  1. Verify search syntax
    • Are operators (OR, AND, -) typed correctly?
    • Are quotes balanced for exact phrases?
  2. Check Active Filters
    • Remove filters you didn’t intend to apply
    • Look at Active Filters chips
  3. Understand search scope
    • Search looks at message content only
    • Not conversation titles or metadata
  4. Consider word stemming
    • Search may match variations: “run”, “running”, “runs”
    • Use exact phrase search if this is a problem

​
Next Steps

Now that you understand message search:

Review Conversations

View complete conversation threads and filter by status, rating, and topic

Track Topics

Understand automatic topic detection and analyze conversation trends

Monitor Metrics

Dashboard overview with CSAT, resolution rate, and performance metrics

Improve Answers

Use message search findings to refine knowledge and guidance

​
Key Takeaways

  1. Message search finds needles in haystacks - Locate specific phrases, patterns, or issues across thousands of messages
  2. Combine filters for precision - Text search + filters + date ranges = powerful targeted analysis
  3. Export for deeper analysis - CSV exports let you analyze data in spreadsheets or share with stakeholders
  4. Regular searches catch problems early - Weekly “I don’t know” searches identify knowledge gaps before they multiply
  5. Use it to validate improvements - After adding knowledge or updating guidance, search to confirm changes work
  6. Label what you find - When message search reveals issues, label those conversations to track follow-up
Message search is your detective tool for understanding exactly what’s happening in individual interactions. Combined with conversation-level analysis and metrics, it gives you complete visibility into your AI’s performance.