Intelligence Features

​

Overview

​

Gender Detection

​

What is Gender Detection?

• Analyzes pitch, resonance, and vocal patterns
• Determines probability: male, female, or uncertain
• Makes result available to agent via variable
• Detection occurs early in conversation (first 5-10 seconds)

​

Enabling Gender Detection

1. Open Call Settings section
2. Locate User Gender Detection toggle
3. Enable the toggle
4. Save changes

• No additional settings required
• Detection is automatic once enabled
• Results available immediately after detection

​

Using Gender Detection in Prompts

# Greeting with Gender

When greeting users, be respectful and use appropriate pronouns.

If {{user_gender}} is detected:

- Male: Use "sir", "he/him" as appropriate
- Female: Use "ma'am", "she/her" as appropriate
- Unknown: Use gender-neutral language

Examples:

- "Yes sir, I can help you with that"
- "Certainly ma'am, let me look that up"
- "Of course, I'd be happy to help you" (neutral)

# Gender-Aware Responses

When recommending products:
If {{user_gender}} == "male":

- Suggest men's product lines first
- Use masculine style references

If {{user_gender}} == "female":

- Suggest women's product lines first
- Use feminine style references

If {{user_gender}} == "unknown":

- Suggest unisex or ask preference
- Use neutral language

​

Use Cases for Gender Detection

Detected: Male → "Good morning sir, how can I help you?"
Detected: Female → "Good morning ma'am, how can I help you?"

Healthcare Products:
- Male: Focus on men's health products
- Female: Focus on women's health products

Arabic-speaking agents:
- Adjust verb endings and pronouns based on gender
- Use appropriate formal address (سيد / سيدة)

Fashion Retail:
- Male: Men's section, masculine styles
- Female: Women's section, feminine styles
- Ask: "Are you shopping for yourself or someone else?"

Fitness/Wellness:
- Tailor advice based on common gender-specific needs
- Adjust communication style if appropriate

​

Best Practices

• Use gender for grammatical agreement (languages with gendered grammar)
• Personalize formal address (sir/ma’am) when culturally appropriate
• Provide gender-specific product recommendations when relevant
• Always have gender-neutral fallback

• Make stereotypical assumptions about interests or preferences
• Use gender to limit options or suggestions
• Over-emphasize gender when not relevant
• Assume gender defines needs or wants

​

Privacy and Compliance Considerations

• Consider mentioning in privacy policy
• Disclose if asked by caller
• Explain it’s for personalization, not discrimination

• Gender is not binary in all cultures
• Some users may not identify with detected gender
• Always allow for correction or preference

​

Accuracy and Limitations

• Voice quality and clarity
• Background noise levels
• Age of speaker (children harder to detect)
• Vocal training or characteristics
• Trans individuals may not align with detected gender

If {{user_gender}} == "unknown" or confidence is low:

- Default to gender-neutral language
- Don't mention or call attention to uncertainty
- Proceed naturally with conversation

​

Smart Call End

​

What is Smart Call End?

• LLM monitors conversation for completion signals
• Detects farewell indicators (“goodbye”, “that’s all”, “thanks, bye”)
• Identifies when objectives are met
• Triggers graceful call conclusion
• Prevents unnecessary conversation extension

​

Enabling Smart Call End

1. Open Call Settings section
2. Locate Smart Call End toggle
3. Enable the toggle
4. Save changes

​

How Smart Call End Detects Completion

• “Goodbye”
• “Bye”
• “That’s all I needed”
• “Thank you, goodbye”
• “I’m all set”
• “Have a good day” (from user)

• Task explicitly completed
• User confirms satisfaction
• No additional questions when asked
• Natural conversation conclusion

User: "Okay perfect, my appointment is confirmed for Tuesday at 2 PM?"
Agent: "Yes, that's correct. Confirmation number is 12345. Anything else?"
User: "Nope, that's everything. Thanks!"
Smart Call End: [Detects objective met + farewell → end call]
Agent: "You're welcome! Have a great day."
[Call ends]

​

Configuring Smart Call End Behavior

# Smart Call End Configuration

When you detect the conversation is complete:

1. Provide brief summary if appropriate
2. Offer final assistance: "Is there anything else I can help with?"
3. If user says no or gives farewell: End gracefully

Final Message Template:
"Thank you for calling [Company]. Have a great [day/evening]!"

Then the call will automatically end.

# Smart Call End with Confirmation

Before ending calls:

1. Summarize actions taken
2. Confirm user is satisfied
3. Provide reference number (if applicable)
4. Offer final help

Example:
"Just to confirm, I've scheduled your appointment for [date/time],
confirmation number [number]. Is there anything else I can help with today?"

If user indicates they're done: Execute graceful farewell and end.

​

Use Cases for Smart Call End

Flow:
1. Book appointment ✓
2. Provide confirmation ✓
3. Ask if anything else
4. User says no
→ Smart Call End triggers graceful goodbye

Flow:
1. Look up order ✓
2. Provide status ✓
3. Ask if they need anything else
4. User satisfied
→ Smart Call End closes conversation

Flow:
1. Answer question ✓
2. Ask if they need clarification
3. User says "nope, that's perfect, thanks"
→ Smart Call End detects completion + farewell

Flow:
1. Gather information ✓
2. Schedule demo ✓
3. Confirm details
4. User says "great, see you then"
→ Smart Call End handles natural conclusion

​

Benefits of Smart Call End

• No awkward “how do I end this call?” moments
• Natural conversation flow
• Users feel in control

• Prevents unnecessarily long calls
• Reduces average call duration
• Eliminates “dead air” at end

• Smooth, confident conclusions
• No stammering or uncertain endings
• Clean, polished experience

• No need to manually script every ending
• Adapts to various conclusion patterns
• Handles unexpected farewell styles

​

Smart Call End vs. Max Call Duration

• Smart Call End: Handles 90%+ of calls naturally
• Max Call Duration: Safety net for edge cases

Max Call Duration: 300 seconds (5 minutes)
Smart Call End: Enabled

Expected: Most calls end naturally at 90-180 seconds via Smart Call End
Backup: If call goes long, max duration prevents runaway at 300 seconds

​

Troubleshooting Smart Call End

# Smart Call End Criteria

Only end the call when ALL of these are true:

1. User's primary objective is complete
2. You've asked "anything else?" and user says no
3. User uses farewell language OR explicitly indicates they're done

If any doubt remains, ask: "Is there anything else I can help you with today?"

• Verify Smart Call End is enabled
• Check prompt doesn’t override with “never end calls”
• Simplify farewell detection in prompt

• Review call transcripts to find patterns
• Adjust prompt to consistently ask if anything else needed
• Ensure clear task completion markers

​

Speaker Identification

​

What is Speaker Identification?

• Analyzes voice biometrics (pitch, timber, resonance)
• Assigns unique identifiers to each speaker
• Tracks conversation by speaker
• Maintains context per speaker

​

Enabling Speaker Identification

1. Open Call Settings section
2. Locate Speaker Identification toggle (marked Beta)
3. Enable the toggle
4. Save changes

​

Use Cases for Speaker Identification

Scenario: Couple calling about joint account

Agent: "Are you both on the line?"
Speaker 1: "Yes, this is John"
Speaker 2: "And I'm Sarah"

Agent tracks:
- Speaker 1 = John
- Speaker 2 = Sarah

Throughout call, agent knows who is speaking and can address appropriately.

Scenario: Parent and child calling about appointment

Parent: "I'm calling to book an appointment for my daughter"
Agent: "Great, and what's your daughter's name?"
Child: "I'm Emily"

Agent distinguishes:
- Parent's voice
- Child's voice
Can direct questions appropriately to each speaker.

Scenario: Multiple stakeholders on conference call

Agent identifies:
- Decision maker
- Technical person
- Budget authority

Can tailor responses to who is speaking.

Scenario: Patient with caregiver

Agent can:
- Get medical info from patient
- Get logistics info from caregiver
- Maintain HIPAA compliance by knowing who is who

​

Technical Implementation

[Speaker 1]: "Hello, I'd like to schedule an appointment"
[Agent]: "Of course! What's your name?"
[Speaker 1]: "This is David"
[Speaker 2]: "And I'm his wife, Sarah. We need a joint appointment"
[Agent]: "Perfect David and Sarah. Let me help you both."

When multiple speakers are detected:

1. Ask each person to identify themselves
2. Address them by name throughout conversation
3. Direct questions to appropriate speaker
4. Confirm with correct person before finalizing

Example:
"David, you mentioned [X]. Sarah, did you have anything to add?"

​

Accuracy and Limitations

• Distinct male/female voices
• Clear audio quality
• Minimal background noise
• Significant voice characteristic differences
• Speakerphone or conference calls

• Similar voices may be confused
• Background noise reduces accuracy
• Very brief utterances harder to identify
• More than 3-4 speakers becomes challenging
• Requires distinct speech patterns

​

Best Practices

• Ask speakers to identify themselves by name
• Use speaker ID to enhance conversation flow
• Address speakers by name when identified
• Gracefully handle when identification is uncertain

• Use for authentication or security
• Assume perfect accuracy
• Ignore verbal identification in favor of voice only
• Make critical decisions based solely on speaker ID

​

Agentic RAG

​

What is Agentic RAG?

1. User asks question
2. System retrieves potentially relevant documents
3. LLM uses retrieved docs to answer
4. Single-pass retrieval

1. User asks question
2. Agent reasons about what information is needed
3. Agent searches knowledge base strategically
4. Agent evaluates results
5. Agent may search again with refined query
6. Agent synthesizes final answer
7. Multi-step, iterative retrieval

​

Enabling Agentic RAG

1. Open Knowledge Base section in Configuration
2. Locate Agentic RAG toggle
3. Enable the toggle
4. Save changes

​

How Agentic RAG Works

1. Retrieve docs matching "return policy electronics online holiday"
2. Pass all docs to LLM
3. Generate answer

1. Agent reasons: "Complex question with three components"
2. First search: "return policy electronics"
3. Evaluates: "Found general electronics policy, but missing holiday info"
4. Second search: "holiday sale terms"
5. Evaluates: "Now have both pieces"
6. Synthesizes: Combines both results into comprehensive answer
7. Response addresses all three components accurately

​

Benefits of Agentic RAG

• More precise answers to complex questions
• Better handling of multi-faceted queries
• Fewer incorrect or partial answers

• Understands nuanced questions
• Identifies what information is actually needed
• Handles ambiguity better

• Can answer questions requiring multiple pieces of information
• Synthesizes information from different sources
• Handles “compare X and Y” questions

• Only searches when actually needed
• Refines search based on what’s found
• Avoids information overload

​

Trade-offs

• Additional LLM calls for reasoning
• Multiple search operations
• Adds 500-2000ms per response

• More LLM API calls
• More processing overhead
• Higher token usage

• More moving parts
• Harder to debug
• Requires well-structured knowledge base

​

When to Use Agentic RAG

Questions requiring multiple documentation sections:
"How do I reset my password if I don't have access to my recovery email?"

Requires:
1. Password reset procedure
2. Alternative authentication methods
3. Account recovery process

Comparative questions:
"What's the difference between the Pro and Enterprise plans?"

Requires:
1. Pro plan features
2. Enterprise plan features
3. Pricing comparison
4. Use case recommendations

Conditional questions:
"Can I return an item if I opened it but didn't use it?"

Requires:
1. General return policy
2. Opened item policy
3. Condition requirements

Questions spanning topics:
"I want to schedule a consultation to discuss upgrading my plan and migrating my data"

Requires:
1. Scheduling process
2. Plan upgrade information
3. Data migration procedures

"What are your business hours?"
- Single fact, standard RAG is sufficient

Emergency support, urgent issues
- Latency is critical, accuracy threshold is lower

<10 knowledge items
- Standard RAG works fine
- Overhead not justified

Single-topic, common questions
- No reasoning needed
- Fast retrieval more important

​

Optimizing Agentic RAG Performance

- Break complex topics into logical chunks
- Use descriptive titles
- Include clear headers and sections
- Moderate document size (1-5 pages)

- Huge monolithic documents
- Vague titles ("Information")
- No clear organization
- Mixed unrelated topics in one doc

• 20 well-written docs > 100 poorly organized docs
• Clear, concise content
• Up-to-date information
• Consistent formatting

​

Monitoring Agentic RAG

• Compare answer accuracy vs standard RAG
• Track questions correctly answered
• Monitor user satisfaction

• Average response time increase
• P95/P99 latency
• User feedback on speed

• Additional LLM costs
• Token usage increase
• ROI vs improved accuracy

• How often does it search multiple times?
• Average number of searches per query
• Which questions trigger multi-step reasoning?

​

Agentic RAG Configuration in Prompts

# Agentic RAG Instructions

When using the knowledge base:

1. First determine if the question is simple or complex
2. For simple questions: Single search is sufficient
3. For complex questions: Use multi-step reasoning

Complex Question Indicators:

- Multiple sub-questions
- Comparisons ("X vs Y")
- Conditional logic ("if... then...")
- Requiring multiple sources

Search Strategy:

1. Identify key information needed
2. Search for each component
3. Evaluate if you have enough information
4. If gaps remain, search with refined query
5. Synthesize comprehensive answer

If information not found after 2-3 searches:
"I don't have that specific information in my knowledge base.
Let me connect you with someone who can help."

​

Agentic RAG + Other Features

• Complex questions answered thoroughly
• Natural conclusion once complete
• Better user satisfaction

• Multiple people with different questions
• Agent handles each person’s needs
• Retrieves relevant info for each speaker

• Gender-specific information retrieval
• Personalized recommendations from knowledge base

​

Language Dialect Switcher

​

What is Language Dialect Switcher?

• Analyzes pronunciation patterns
• Identifies regional dialect
• Switches speech recognition model
• Improves transcription accuracy

​

Enabling Language Dialect Switcher

1. Open Call Settings section
2. Locate Language/Dialect Switcher toggle
3. Enable the toggle
4. Save changes

​

How Dialect Switching Works

• Agent configured with Egyptian Arabic voice
• Caller speaks in Saudi dialect

• Speech recognition optimized for Egyptian
• Saudi pronunciation may be misunderstood
• Some words transcribed incorrectly

• Detects Saudi dialect from first few sentences
• Switches to Saudi-optimized recognition
• Better transcription accuracy
• Agent still speaks in Egyptian voice

• Agent configured with US English
• Caller has strong Scottish accent

• US English speech recognition
• Accent causes transcription errors
• May misunderstand key words

• Detects UK/Scottish accent
• Adapts recognition
• Better understands caller
• Agent still speaks in US English voice

​

Supported Dialects

• Egyptian (EGY)
• Saudi Arabian (KSA)
• Emirati (UAE)
• Jordanian (JOR)
• Lebanese (LEB)
• Syrian (SYR)
• Palestinian (PLS)
• Iraqi (IRQ)
• Bahraini (BAH)

• US English
• UK English
• Australian English
• Canadian English
• Indian English
• Various regional accents

​

Benefits of Dialect Switching

• Better transcription accuracy
• Fewer misunderstood words
• Reduced need for repetition

• Single agent handles multiple dialects
• No need for separate agents per region
• Consistent brand voice across markets

• Callers feel understood
• Natural conversation flow
• Less frustration

• One agent instead of many region-specific agents
• Reduced development complexity

​

Latency and Performance

• Detection: 2-5 seconds (during first few sentences)
• Switching: ~50-100ms
• Total impact: Minimal (one-time cost at call start)

• Dialect detection: 85-95% accuracy
• Recognition improvement: 10-30% better transcription for non-standard dialects

​

Use Cases

Company serves entire Middle East:
- Configure agent with widely-understood Egyptian Arabic
- Enable Dialect Switcher
- Handles callers from Saudi, UAE, Jordan, etc.
- Each caller understood in their dialect

Global SaaS company:
- Configure with US English
- Enable Dialect Switcher
- Handles UK, Australian, Indian callers
- Better understands various accents

National retail chain:
- Customers from all regions
- Single agent handles all calls
- Adapts to each caller's way of speaking

​

Limitations

• Agent continues speaking in configured dialect
• Only affects how agent understands caller
• For voice change, need separate agents

• Cannot switch English ↔ Arabic
• Only within same language
• For multi-language, need separate agents

• Works best with standard dialects
• Heavy accents may still challenge system
• Non-standard speech patterns may not be recognized

​

Best Practices

• Enable for agents serving multi-region markets
• Test with speakers from target dialects
• Monitor transcription accuracy improvements
• Combine with standard RAG or knowledge base

• Expect it to change agent’s speaking dialect
• Rely on it for completely different languages
• Assume 100% accuracy across all accents
• Skip testing with actual dialect speakers

​

Combining Intelligence Features

​

Recommended Combinations

✓ Gender Detection (personalization)
✓ Smart Call End (natural conclusions)
✓ Language Dialect Switcher (understand everyone)

✓ Agentic RAG (complex questions)
✓ Smart Call End (after thorough answers)
✓ Language Dialect Switcher (diverse callers)

✓ Speaker Identification (who's speaking)
✓ Gender Detection (personalization per speaker)
✓ Agentic RAG (comprehensive information)

✓ Smart Call End (efficiency)
✓ Language Dialect Switcher (broad reach)
Skip: Agentic RAG (latency sensitive)
Skip: Speaker Identification (simple 1:1 calls)

​

Feature Selection Matrix

​

Performance Considerations

​

Latency Impact by Feature

• Detection overhead: ~400ms (one time)
• Agentic RAG: ~2000ms (when used)
• Total: ~2400ms for complex knowledge query

• Detection overhead: ~200ms (one time)
• Smart call end: 0ms
• Total: Negligible impact

​

Cost Impact

• Gender Detection: +1-2 API calls per call
• Smart Call End: +0-1 API calls per call (minimal)
• Speaker Identification: +1-3 API calls per speaker
• Agentic RAG: +2-5 API calls per knowledge query
• Dialect Switcher: +1-2 API calls at start

1000 calls/month baseline: $200
+ Gender Detection: +$10-20
+ Smart Call End: +$5
+ Speaker ID (if used): +$15-30
+ Agentic RAG (avg 2 queries/call): +$50-100
+ Dialect Switcher: +$10-15

Total: $290-370 (45-85% increase)

​

Troubleshooting Intelligence Features

​

Gender Detection Issues

• Review prompt to handle uncertainty gracefully
• Use gender-neutral fallbacks
• Don’t over-rely on detection for critical decisions

• Check audio quality
• Verify feature is enabled
• Check for background noise interference
• Some voices are naturally ambiguous

​

Smart Call End Issues

# Stricter Smart Call End

Only end when:

1. Objective explicitly complete
2. Asked "anything else?" and user said no
3. User used farewell ("goodbye", "that's all")

• Review transcripts for patterns
• Simplify farewell detection
• Ensure agent asks confirming question before end

​

Speaker Identification Issues

• Ask speakers to identify themselves verbally
• Use names instead of relying solely on voice
• Check audio quality (speakerphone distance)

• Always ask “who am I speaking with now?”
• Confirm name when speaker changes
• Use verbal identification as primary method

​

Agentic RAG Issues

• Review knowledge base size (too large?)
• Reduce complexity of documents
• Consider disabling for simple questions
• Check for prompt encouraging excessive searching

• Review knowledge base content quality
• Ensure information is up to date
• Check for contradictory information
• Improve document structure and titles

​

Dialect Switcher Issues

• Verify caller’s dialect is supported
• Check for heavy accent vs standard dialect
• Review audio quality
• May need custom training for very specific accents

• First few seconds may be ambiguous
• System typically corrects after more speech
• Check for code-switching (mixing dialects)

​

Best Practices Summary

​

Do’s

• Test each feature with real calls
• Measure latency impact
• Verify value added

• Begin with Smart Call End + Dialect Switcher
• Add features one at a time
• Measure impact of each

• Track latency metrics
• Monitor cost increases
• Measure user satisfaction

• Handle feature failures elegantly
• Always have fallbacks
• Don’t assume 100% accuracy

​

Don’ts

• Only enable features you need
• More features = more latency
• Each feature has trade-offs

• Test with real target audience
• Verify benefits outweigh costs
• Check edge cases

• Review legal requirements
• Update privacy policy if needed
• Consider data retention

• All ML features have error rates
• Build robust fallbacks
• Handle uncertainty gracefully

​

Feature-Specific Recommendations

​

Always Recommend

• Minimal cost
• No latency
• Better UX
• Enable by default

• Minimal cost (~50ms one-time)
• Improves understanding
• Broader reach
• Enable for multi-region

​

Situationally Recommend

• If culturally appropriate
• If personalization adds value
• If prompt can use it effectively
• Check legal compliance

• If knowledge base is large/complex
• If accuracy > latency priority
• If questions are multi-faceted
• If budget allows

• If multi-person calls expected
• If call context benefits from it
• If willing to handle beta limitations
• If accuracy is acceptable

​

Next Steps