Skip to main content
A female Hebrew voice AI agent for Maccabi Healthcare Services that helps callers find medical services, clinics, doctors, pharmacies, and urgent care centers across Israel, providing location-based search results with operating hours and contact details.

View Agent

Open in Wonderful Platform

Overview

FieldValue
MarketIsrael - Healthcare Services
VerticalHealthcare
CompanyMaccabi (מכבי שירותי בריאות)
ModelGPT Realtime 28-08-2025
PersonaFemale, Professional, Empathetic
LanguageHebrew Only
ChannelVoice

Skills

maccabiServiceLocator - Finds and locates Maccabi healthcare services including clinics, doctors, pharmacies, urgent care centers, labs, and dental services

Key Tools

ToolPurpose
identify_user_request_service_locator()Classifies user requests as FAQ/service/out-of-scope; identifies medical terms using LLM
pick_term_service_locator()Resolves ambiguity when multiple service terms are identified
note_taker_service_locator()Saves search filters (doctor name, gender, language, time preferences)
predict_location_service_locator()Predicts cities/streets/POIs from Hebrew text using Google Maps + fuzzy matching
identify_location_service_locator()Confirms and stores resolved location coordinates for search
search_service_locator()Executes search with all gathered context; returns formatted results
forward_service_locator()Forwards call to human agent with auto-determined extension
external_message_service_locator()Sends SMS with service details (name, address, hours, phone)
trolling_counter_service_locator()Silent counter for inappropriate behavior; forwards after threshold
on_start_service_locator()Session initialization; handles blocklist and authentication

Prompting Techniques

1. No Echoing Rule

Critical instruction to never repeat user’s words. No Echoing Rule (CRITICAL): NEVER repeat, summarize, or paraphrase what the user just said. Do not restate any details from the user’s request (names, locations, services, dates). Jump directly to action or result.

2. Filler Restrictions

Prevents information leakage in acknowledgments. Filler Restrictions: Fillers must be generic and short. They must NOT include: doctor/entity names, locations, service types, dates/times, or any other detail the user mentioned. The purpose of a filler is only to signal you’re working - not to confirm understanding.

3. Voice-First Result Presentation

Present one result at a time. Present one result. STEP 1: Give name + address briefly. If unsure it’s what they want, ask: ‘זה מה שחיפשת?’ / ‘זה מתאים לך?’ STEP 2 (if confirmed): First, offer to send full details via SMS. STEP 3 (if they decline SMS or ask verbally): If they asked for specific info (hours/phone/referral) → give only that.

4. Strict Grounding

No hallucination allowed. Strict Grounding: You are prohibited from inventing or guessing information. Your response must be based solely on the data returned by the tools.

5. FAQ-First Classification

Comprehensive FAQ database with LLM routing The agent uses a large FAQ database with keywords and semantic matching. LLM classifies requests into: FAQ, Service (search), or Out-of-Scope. FAQ responses take priority to handle common requests without search.

Lessons Learned

What Worked
  • State machine architecture - Breaking the flow into discrete steps (identify request → identify location → search) with clear transitions improved reliability and debugging
  • Short base prompt - Keeping the base prompt concise while moving detailed instructions into tool descriptions reduced latency and improved response consistency
  • KV store for context - Using key-value storage to persist conversation state (identified terms, coordinates, notes, search history) enabled stateless tool execution while maintaining context across the conversation
  • Dynamic instructions in tools - Returning agent_instructions from tools based on results allowed context-aware guidance without bloating the base prompt
Challenges
  • Tools latency - Multiple sequential tool calls (identify → predict location → confirm location → search → format) added latency. Mitigated through parallel API calls, LLM racing (Gemini vs GPT), and code-based formatting instead of LLM formatting
  • Agent humanity - Making the voice agent sound natural rather than robotic required careful prompt engineering: no echoing rules, filler restrictions, voice-first formatting (Hebrew words for times/distances), and conversational flow that lets the caller lead