Skip to main content
Shir is an AI scheduling voice agent for Maccabi Healthcare Services. It speaks fluent, natural Israeli Hebrew and is designed to schedule new appointments for Family Doctors or Pediatricians .

View Agent

Open in Wonderful Platform

Overview

FieldValue
MarketIsrael - Healthcare Services
VerticalHealthcare
CompanyMaccabi
ModelGPT Realtime 25-08-2025
PersonaFemale, Professional, Empathetic
LanguageHebrew Only
ChannelVoice

Skills

maccabiAppointments - Handles the end-to-end flow of scheduling new medical appointments including patient verification, clinic/location search, appointment slot presentation, and booking confirmation

Key Tools

ToolPurpose
verify()Verifies customer identity using ID number and full name
select_relative()Selects a family member for scheduling appointments on their behalf
predict_location_shared()Predicts locations from Hebrew queries using Google Maps API and fuzzy search
identify_location_maccabi()Identifies and sets specific location/city code for appointment search
get_available_clinics()Gets available clinics by city name or coordinates
get_personal_clinics()Gets personal clinics based on verified member’s branch
pick_clinic()Picks a clinic and retrieves available appointment dates
pick_date()Picks a date and retrieves available appointment times
schedule_appointment()Books the appointment at the selected time
forward()Transfers the call to a human agent with appropriate transfer reason

Prompting Techniques

1. Conversation State Machine with Linear Phases
  • The agent follows a strict 7-phase conversation flow (Triage → Context Setup → Authentication → Search Parameters → Presentation → Booking → Closing) ensuring no steps are skipped.
  • You must move linearly through these phases. Do not skip steps. If there are instructions in the tools output, you must follow them, it describes your next steps.
2. Hebrew TTS Optimization Special punctuation rules to improve voice synthesis quality:
  • NO COMMAS: Do not use commas (,). Use periods (.) to force pauses between thoughts.
  • NO HYPHENS: Do not use hyphens (-). Use colons (:) or separate sentences instead.
  • IDs & Codes: Speak digits individually, separated by periods: “1. 2. 3. 4. 5.”
3. Tool Output as Instructions
  • Tool responses contain notes that the agent must treat as mandatory instructions: Always follow any notes from the tool output (treat notes as mandatory instructions.
4. Confidentiality Guards Strict rules to avoid revealing sensitive information:
  • Existing Appointment (Confidentiality): Only block when get_future_appointments() returns hasFamilyPedsAppointments=true. If so, do NOT say they have one. Say: “According to my check, I will transfer you to a human representative”
5. Scope Control with Silent Metrics
  • Out-of-scope requests are tracked silently before transferring:
  • Out of Scope: If the user asks to change/cancel an appointment, or asks for Specialists… → Apologize you are unable to assist, call track_metric(metricType: “RestrictedCount”, shouldReset: false), then call transfer_call(phone: “11”, transferReason: 0).

Lessons Learned

What worked
  • Using clear conversation phases/stages enabled better mapping and analysis of conversations using custom metrics (tags)
  • Requesting ID number input via DTMF (typing followed by #) improved identification accuracy
  • Treating tool notes as mandatory instructions allowed dynamic flow control from backend
Challenges
  • Transcription of numbers and names resulted in difficulty identifying some customers. Solution: Developed a feature that allows switching between transcription models during a conversation; learned through working with different transcription models
  • Hebrew TTS pronunciation issues affected user experience Solution: Implemented strict punctuation rules (no commas, no hyphens) and specific pronunciation guidelines for dates, addresses, and phone numbers