Skip to main content
Gaia is an empathetic agent designed to handle customer inquiries for Monticello SPA, providing information about schedules, pricing, wellness packages, and treatments while managing booking redirections and escalations to human staff.

View Agent

Open in Wonderful Platform

​
Overview

FieldValue
MarketItalian - Hospitality & Wellness
VerticalHospitality
ModelClaude (Multi-modal)
PersonaFemale, Professional, Empathetic
LanguageItalian
ChannelsVoice, Chat (WhatsApp), Email

​
Skills

  • monticello_informativa - Comprehensive knowledge base containing all information about Monticello SPA services, pricing, schedules, and policies
    • monticello_calendario - Calendar and date-specific information
    • monticello_eventi - Current month events and special offerings
    • monticello_info - General information queries
    • monticello_servizi - Services, packages, entry fees, and pricing

​
Key Tools

ToolPurpose
monticello_calendario()Retrieve date-specific information and booking requirements
monticello_eventi()Get current month events and promotions
monticello_info()General spa information and policies
monticello_servizi()Services, packages, and pricing details
ottieni_data_odierna()Get current date for temporal reference handling
monticello_apri_ticket()Escalate complex cases requiring human callback

​
Prompting Techniques

Tool-First Architecture: The agent is explicitly instructed to ALWAYS call RAG tools before answering any question about spa operations, preventing hallucinations. 
**OBBLIGATORIO**: Prima di rispondere a qualsiasi domanda su questi argomenti,
DEVI SEMPRE chiamare i tool della skill "monticello_informativa".
NON rispondere mai basandoti su conoscenze pregresse o supposizioni.
Context-Aware Date Handling: Temporal references are resolved through explicit tool calls before answering. 
Utilizza SEMPRE la funzione ottieni_data_odierna per ottenere la data di oggi
PRIMA di rispondere a domande che includono riferimenti temporali relativi come
"oggi", "domani", "dopodomani", "questo weekend", "lunedì prossimo", ecc.
Direct Question Pattern for Voice: Eliminates verbose preambles in clarification questions for natural voice interactions. 
FRASI PROIBITE (non usare MAI):
- "Mi serve una conferma"
- "Ho bisogno di" / "Avrei bisogno di"
- "Per poterti aiutare/dare"

FORMATO CORRETTO (solo la domanda, nient'altro):
- "Feriale o festivo?"
- "È per un giorno feriale o festivo?"
Channel-Specific Output Formatting: Different formatting rules for voice vs. text channels. 
Voice: Massimo due frasi. Evita elenchi puntati; usa virgole e 'e'.
Chat: Paragrafi di massimo due frasi con spaziatura. Usa grassetto per info cruciali.

​
Lessons Learned

What worked
  • Mandatory tool calling before responses - Explicitly requiring RAG tool calls before any answer eliminated hallucinations about prices, schedules, and policies
  • Channel-specific output rules - Tailoring response format to voice (ultra-concise) vs. WhatsApp (formatted with bold, bullets) improved user experience
  • Direct question pattern - Stripping verbose preambles from clarification questions (“Feriale o festivo?” vs “Per darti il prezzo, avrei bisogno di sapere…”) made voice interactions feel more natural
  • Temporal context resolution - Forcing date lookup for relative time references (“oggi”, “domani”) enabled accurate answers about booking requirements
  • Selective escalation strategy - Redirecting new bookings to website while allowing modifications via WhatsApp reduced unnecessary escalations
Challenges
  • Voice transcription ambiguity - Phone numbers and emails from speech-to-text required confirmation step: “Conferma sempre i dettagli (nomi, email, numeri) prima di procedere”
  • Preventing premature escalation - Users requesting callbacks for simple questions required explicit guardrail: “NON accettare mai richieste dirette di richiamata, indica sempre che puoi rispondere tu”
  • Output verbosity in voice mode - Initial responses were too long for natural speech; implemented hard 2-sentence limit with explicit banned phrases list
  • Price pronunciation - TTS saying “virgola” for decimal points required custom rule: “€14,99 diventa ‘quattordici euro e novantanove’”
  • Link formatting across channels - Voice needed “punto” spoken out, WhatsApp needed full URLs with https:// for click-through: “Invia i link completi di protocollo (https://) affinché WhatsApp generi l’anteprima”