Reecall
Data Extraction

Best practices & use cases

Choose the fields you need, describe how to extract them, and test the results on representative conversations. Use the schemas below as starting points for your Assistant's dataExtractionSchema.

Design focused schemas

  • Extract only the data points you actually need. A smaller, well-scoped schema is easier to test and refine than a broad one.
  • Prefer flat structures over deeply nested objects when the data allows it.
  • Each additional field gives you another extraction result to validate. If a field is "nice to have" rather than essential, leave it out.

Prefer this:

JSONCode
{ "type": "object", "properties": { "motive": { "type": "string" }, "product": { "type": "string" }, "issueResolved": { "type": "boolean" } }, "required": ["motive", "product", "issueResolved"], "additionalProperties": false }

Over this:

JSONCode
{ "type": "object", "properties": { "motive": { "type": "string" }, "product": { "type": "string" }, "issueResolved": { "type": "boolean" }, "sentiment": { "type": "string" }, "language": { "type": "string" }, "summary": { "type": "string" }, "keywords": { "type": "array", "items": { "type": "string" } }, "followUpNeeded": { "type": "boolean" }, "department": { "type": "string" }, "priority": { "type": "string" } }, "required": ["motive", "product", "issueResolved", "sentiment", "language", "summary", "keywords", "followUpNeeded", "department", "priority"], "additionalProperties": false }

Guide the LLM with descriptions

  • Always include a description field on each property. The LLM uses these descriptions to understand what to extract and from which parts of the conversation.
  • Describe the information to extract. For example, use "The main reason why the user is calling" for motive.
  • Be specific about edge cases in the description. For instance, specify what value to use when the caller does not mention a product.

Example:

JSONCode
{ "type": "object", "properties": { "motive": { "type": "string", "description": "The main reason why the user is calling." }, "product": { "type": "string", "description": "The product or service the conversation relates to, as mentioned by the caller." } }, "required": ["motive", "product"], "additionalProperties": false }

Use enums for categorisation

The example below is a properties fragment to place inside an object schema.

  • When a field has a known set of possible values, define them as an enum. This constrains the LLM output to your predefined categories, making downstream processing predictable.
  • Keep enum values concise and mutually exclusive.
  • If the list of possible values is open-ended, use a free-text string instead of trying to enumerate every possibility.

Example:

JSONCode
{ "motive": { "type": "string", "description": "The main reason why the user is calling.", "enum": ["Technical Support", "Billing", "Appointment request", "General enquiry"] }, "urgency": { "type": "string", "description": "How urgent the caller's request is, based on the tone and content of the conversation.", "enum": ["low", "medium", "high"] } }

Handle optional data with nullable fields

The example below is a properties fragment to place inside an object schema.

  • Use nullable types such as ["string", "null"] for fields that may not be mentioned in the conversation.
  • Include guidance in the description about when null is the appropriate value. This helps the LLM distinguish between "the caller did not mention it" and "the caller said something ambiguous".

Example:

JSONCode
{ "referralSource": { "type": ["string", "null"], "description": "How the caller heard about the service, or null if not mentioned during the conversation." } }

See Nullable Fields in the schema reference for the full syntax.


Iterate and test

  • Test your schema against real conversations before deploying to production. Review the extracted data on several Exchange records to verify the LLM is interpreting your descriptions correctly.
  • If a field is consistently misclassified, refine its description or narrow its enum values rather than adding more fields.
  • Start with a minimal schema and expand incrementally. Adding one or two fields at a time makes it easier to identify which change caused a regression in accuracy.

Use cases

Copy a schema, adjust its descriptions and enum values for your use case, and set it as the Assistant's dataExtractionSchema.

Support line

Extract the reason for the call, the affected product, how the issue was resolved, and any follow-up commitment.

JSONCode
{ "type": "object", "properties": { "motive": { "type": "string", "description": "The main reason why the caller is contacting support.", "enum": ["Technical issue", "Billing enquiry", "Account management", "Complaint", "Other"] }, "product": { "type": ["string", "null"], "description": "The product or service the issue relates to, or null if not identified during the call." }, "issueResolved": { "type": "boolean", "description": "Whether the caller's issue was fully resolved during this call." }, "resolutionMethod": { "type": ["string", "null"], "description": "A brief description of how the issue was resolved, or null if the issue remains open." }, "followUpRequired": { "type": "boolean", "description": "Whether a follow-up action (callback, escalation, ticket) was promised to the caller." } }, "required": ["motive", "product", "issueResolved", "resolutionMethod", "followUpRequired"], "additionalProperties": false }

Example output:

JSONCode
{ "motive": "Technical issue", "product": "Fibre broadband", "issueResolved": true, "resolutionMethod": "Router restarted remotely by the agent.", "followUpRequired": false }

User need qualification

Capture the caller's need, their profile, and a qualification outcome to route or prioritise the lead.

JSONCode
{ "type": "object", "properties": { "need": { "type": "string", "description": "A concise summary of what the caller is looking for." }, "category": { "type": "string", "description": "The broad category that best fits the caller's need.", "enum": ["Product enquiry", "Service request", "Partnership", "Pricing", "Other"] }, "companyName": { "type": ["string", "null"], "description": "The name of the caller's company, or null if they are an individual or did not mention it." }, "budget": { "type": ["string", "null"], "description": "The caller's stated budget or budget range, or null if not discussed.", "enum": ["Under 1k", "1k-10k", "10k-50k", "Over 50k", null] }, "urgency": { "type": "string", "description": "How urgently the caller needs a response, inferred from the tone and content of the conversation.", "enum": ["low", "medium", "high"] }, "qualified": { "type": "boolean", "description": "Whether the caller's need matches the services offered and warrants a follow-up." } }, "required": ["need", "category", "companyName", "budget", "urgency", "qualified"], "additionalProperties": false }

Example output:

JSONCode
{ "need": "Looking for an AI voice assistant to handle after-hours calls for a dental practice.", "category": "Product enquiry", "companyName": "Bright Smile Dental", "budget": "1k-10k", "urgency": "medium", "qualified": true }

Appointment booking

Extract the details needed to confirm or follow up on a scheduling request.

JSONCode
{ "type": "object", "properties": { "appointmentType": { "type": "string", "description": "The type of appointment the caller wants to book.", "enum": ["Initial consultation", "Follow-up", "Annual review", "Urgent visit", "Other"] }, "preferredDate": { "type": ["string", "null"], "description": "The caller's preferred date for the appointment (as mentioned in the conversation), or null if no preference was stated." }, "preferredTimeSlot": { "type": ["string", "null"], "description": "The caller's preferred time of day, or null if no preference was stated.", "enum": ["morning", "afternoon", "evening", null] }, "callerName": { "type": ["string", "null"], "description": "The full name of the person the appointment is for, or null if not provided." }, "contactPhone": { "type": ["string", "null"], "description": "A phone number to confirm the appointment, or null if not provided." }, "bookingConfirmed": { "type": "boolean", "description": "Whether a specific date and time were agreed upon during the call." } }, "required": ["appointmentType", "preferredDate", "preferredTimeSlot", "callerName", "contactPhone", "bookingConfirmed"], "additionalProperties": false }

Example output:

JSONCode
{ "appointmentType": "Initial consultation", "preferredDate": "next Tuesday", "preferredTimeSlot": "morning", "callerName": "Sophie Martin", "contactPhone": "+33 6 12 34 56 78", "bookingConfirmed": false }

FAQ

Identify which questions were asked, whether the Assistant was able to answer them, and any remaining gaps.

JSONCode
{ "type": "object", "properties": { "questions": { "type": "array", "description": "The distinct questions the caller asked during the conversation.", "items": { "type": "object", "properties": { "question": { "type": "string", "description": "A concise restatement of the caller's question." }, "answered": { "type": "boolean", "description": "Whether the assistant provided a satisfactory answer to this question." } }, "required": ["question", "answered"], "additionalProperties": false } }, "allQuestionsAnswered": { "type": "boolean", "description": "Whether every question raised by the caller was answered during the conversation." }, "topicCategory": { "type": "string", "description": "The primary topic area of the caller's questions.", "enum": ["Product features", "Pricing", "Account & billing", "Technical setup", "Policies", "Other"] } }, "required": ["questions", "topicCategory", "allQuestionsAnswered"], "additionalProperties": false }

Example output:

JSONCode
{ "questions": [ { "question": "What are the supported languages?", "answered": true }, { "question": "Is there a free trial?", "answered": true }, { "question": "Can I export call recordings?", "answered": false } ], "allQuestionsAnswered": false, "topicCategory": "Product features" }
Last modified on