Extractors

Welcome to the Extractors guide! Extractors are a powerful feature that allows your AI assistant to automatically extract specific information from call transcripts without explicitly asking questions during the conversation.

What are Extractors?

Extractors are specialized actions that analyze call transcripts to find and extract specific pieces of information. Unlike regular questions that are asked directly to the caller, extractors work behind the scenes by:

• Analyzing conversation transcripts after or during the call
• Finding relevant information based on your instructions
• Extracting data even when it wasn't explicitly asked for
• Working silently without interrupting the conversation flow
• Using the assistant's language to accurately understand and extract information

Language Support: Extractors automatically work in your assistant's configured language. If your assistant is set to English or Multilingual, extractors will analyze transcripts in English. For German-language assistants, extractors will use German for analysis. This ensures accurate extraction based on how the conversation actually took place.

How Extractors Differ from Questions

The main difference between extractors and questions is how they gather information:

Feature	Questions	Extractors
How they work	Ask callers directly during the conversation	Extract information from the transcript
Visibility	Caller hears the question	Works silently in the background
When to use	Need specific information from the caller	Want to capture information mentioned naturally
Question field	Required: "Question to the user"	Not used: Only needs extraction instructions

When to Use Extractors

Extractors are ideal when:

• Information is mentioned naturally in conversation without prompting
• You want a smoother conversation flow without too many direct questions
• Multiple data points can be captured from a single topic discussion
• The same question might be answered in different ways or contexts
• You want to avoid repetition if information was already provided

Examples of Good Extractor Use Cases

1. Company Name: Instead of asking "What's your company name?", extract it when the caller naturally mentions it
2. Contact Details: Capture phone numbers, emails, or addresses mentioned during the conversation
3. Problem Description: Extract the reason for calling from their natural explanation
4. Availability: Note when they mention their preferred times or dates
5. Preferences: Capture preferences they mention without direct prompting

Creating an Extractor

Step 1: Access Extractors

1. Navigate to Actions in the main menu
2. Click on the Extractors tab
3. Click "Add Extractors" or "Create new extractor"

Step 2: Choose the Type

Select the extractor type based on what information you want to extract:

• Yes/No Action: Extract boolean information (e.g., "Are they interested?" → true/false)

  • Best for simple yes/no decisions or true/false determinations
  • Results in a boolean value that can be easily filtered

• Single Choice: Extract one option from predefined choices (e.g., department mentioned)

  • Requires defining a list of possible choices
  • The AI will select the best matching option from your list
  • Useful for categorization and structured data

• Open Action: Extract free-form text information (e.g., company name, problem description)

  • No predefined options - extracts any text
  • Best for names, descriptions, or unique information
  • Most flexible but requires clear instructions

Tip: Choose the most restrictive type that fits your needs. Single Choice is better than Open Action if you know all possible values, as it ensures consistency.

Step 3: Configure the Extractor

Fill in the required fields:

1. Identifier: A unique name in lowercase (e.g., company_name, contact_reason)

   • Use only lowercase letters and underscores
   • Make it descriptive but concise
   • This identifier will be used to reference the extracted data

2. Instruction on Answer: Clear instructions for what to extract

   • Explain what information to look for in the transcript
   • Describe how to identify the information
   • Specify what to do if the information isn't found

   Example for Company Name:

   Find the company name mentioned by the caller in the transcript. Look for phrases like "I'm calling from...", "I work at...", or "This is regarding [company name]". If multiple company names are mentioned, extract the one that belongs to the caller, not the assistant or other parties. If no company name is found, leave blank.

3. Examples (for Open Actions): Provide examples of expected answers

   • Show what good extractions look like
   • Include variations of how the information might appear
   • Help the AI understand the format you expect
   • Examples: If extracting company names, provide: "Acme Corp", "Smith & Associates", "Global Tech Inc"

4. Choices (for Single Choice): Define the possible options

   • List all valid choices the AI can select from
   • Keep options clear and mutually exclusive
   • Consider adding an "Other" or "Unknown" option
   • Examples: For department: "Sales", "Support", "Billing", "Technical", "Other"

Step 4: Save and Test

1. Click Save to create your extractor
2. Attach the extractor to an assistant in the Extractors tab
3. Test with sample calls to verify it extracts correctly
4. Adjust instructions if needed based on results

Note: The extractor will automatically use the language of the assistant it's attached to. Make sure your extraction instructions match the expected language of the conversations.

Using Extractors with Assistants

Attaching Extractors to an Assistant

1. Go to Assistants and select your assistant
2. Click on the Extractors tab
3. Click "Add Extractors"
4. Select the extractors you want to use from the list
5. Click Save

The extractors will now automatically analyze all calls for that assistant.

Managing Extractors

From the Extractors tab in your assistant:

• View all attached extractors and their configurations
• Edit instructions by clicking on an extractor
• Add or remove extractors as needed - changes are saved together when you click Save
• Reorder extractors if processing order matters

Note: When you add, edit, or remove extractors, remember to click the Save button to apply all your changes at once.

Best Practices

Language Considerations

Extractors work in the language configured for your assistant:

• English/Multilingual Assistants: Extractors will analyze transcripts in English
• German Assistants: Extractors will analyze transcripts in German

Tips for multi-language use:

• Write your extraction instructions in the same language as your assistant
• If you work with multiple languages, create separate assistants for each language
• Test extractors with calls in the correct language to ensure accuracy

Writing Good Instructions

✅ Do:

• Be specific about what to extract
• Explain how to identify the information
• Describe what to do if information is missing
• Mention what NOT to extract (e.g., assistant's information)
• Use clear, simple language

❌ Don't:

• Make instructions too vague or general
• Assume the AI knows context without explaining
• Forget to handle cases where information isn't present
• Write overly complex instructions

Examples of Good Instructions

For Contact Email:

Extract the email address provided by the caller during the conversation.
Look for patterns like "my email is...", "you can reach me at...", or
"send it to...". Only extract email addresses from the caller, not from
the assistant. If no email is provided, leave blank.

For Preferred Contact Time:

Find when the caller mentioned they prefer to be contacted. Look for
phrases about availability, preferred times, or when they're usually
free. Extract as specific time ranges (e.g., "mornings", "2-4 PM",
"weekdays"). If they mention they're available anytime, extract "anytime".
If no preference is mentioned, leave blank.

For Interest Level (Yes/No):

Determine if the caller expressed interest in the product/service discussed.
Look for positive indicators like agreement, asking for more information,
requesting follow-up, or expressing enthusiasm. Consider hesitation or
requests to "think about it" as not interested. Return true if interested,
false if not interested.

Combining Questions and Extractors

You can use both questions and extractors in the same assistant for optimal results:

1. Ask direct questions for critical information you must have
2. Use extractors for supplementary information mentioned naturally
3. Avoid redundancy by not asking questions for information you're extracting
4. Review extracted data regularly to ensure accuracy

Example Strategy

For a sales assistant:

• Question: "What product are you interested in?" (Direct question for critical info)
• Extractor: Company name (Extract if mentioned naturally)
• Extractor: Budget range (Extract if they mention it during discussion)
• Question: "What's the best email to send information?" (Direct question for contact)
• Extractor: Urgency level (Extract based on their language and tone)

Searching and Organizing Extractors

Search Functionality

Use the search box to quickly find extractors:

1. Type any part of the extractor's identifier
2. Results filter in real-time as you type
3. Search is case-insensitive

Organization Tips

• Use consistent naming: Follow a pattern like [category]_[field] (e.g., contact_email, contact_phone)
• Group related extractors: Create extractors for related information together
• Document purpose: Keep notes on what each extractor does and why it's needed
• Review regularly: Check if extractors are still relevant and accurate

Troubleshooting

Extractor Not Finding Information

If your extractor isn't working as expected:

1. Review the instruction: Make it more specific and detailed
2. Check examples: Ensure examples match the actual format in calls
3. Test with transcripts: Review call transcripts to see how information appears
4. Adjust keywords: Include more variations of how people might phrase things
5. Simplify: Sometimes simpler instructions work better

Extracting Wrong Information

If the extractor captures incorrect data:

1. Add negative examples: Specify what NOT to extract
2. Be more specific: Narrow down exactly what you want
3. Check for conflicts: Ensure no other extractors overlap
4. Review edge cases: Test with various scenarios

Missing Data

If extractors frequently leave fields blank:

1. Verify the information is mentioned: Check if callers actually provide it
2. Broaden criteria: Make instructions less restrictive
3. Consider a question instead: For critical info, ask directly
4. Add prompts: Guide callers to mention the information naturally

Next Steps

• Review the Actions Overview for general action concepts
• Learn about Searching Actions
• Explore Call Reasons and Questions

---

If you have questions about extractors or need help setting them up, please contact your administrator or support team!