Transforming Thailand's Tourism With AI-Powered Assistance

The tourism industry faces significant challenges in providing accurate, real-time information to visitors. Before the NVIDIA-powered solution was implemented, tourists struggled with fragmented information spread across multiple websites, concerns about information authenticity, outdated content, and language barriers. These challenges often led to poor tourist experiences and inefficiencies in service delivery.

Siam AI and Wipro recognized the potential of NVIDIA's comprehensive AI solutions to address these challenges. The integration of NVIDIA's advanced technologies —particularly the NVIDIA NeMo™ framework and NVIDIA NIM™—enabled the development of a sophisticated tourism assistant platform called Sukjai that can understand and respond to diverse tourist inquiries effectively.

Siam AI is a prominent AI cloud computing service provider based in Thailand. Beyond its core cloud computing services, the company has developed SiamGPT, an advanced generative AI platform with specialized knowledge about Thailand and information particularly relevant to Thai people. SiamGPT features a modular design to accommodate expertise across various domains and supports multiple usage formats, including web applications, mobile applications, and API integration. The initial development phase has focused on creating an AI model specialized in Thai tourism. 

The training data for this model comprises tourism information provided by the Tourism Authority of Thailand (TAT). Siam AI has meticulously verified and enhanced the quality of this data while incorporating "Sukjai," TAT's mascot, as the model character for this LLM to further support TAT's promotional efforts. Additionally, Siam AI has supplemented the training with extensive unseen tourism data from various sources, resulting in a high-quality, reliable model suitable for practical applications.

The implementation of this NVIDIA-powered solution has delivered significant benefits:

• Unified Information Access: Tourists can now access accurate, real-time information through a single platform.
• Authentic Content: Direct integration with the Tourism Authority of Thailand ensures information reliability.
• Multilingual Support: Enhanced communication capabilities are available in both Thai and English.
• Rapid Deployment: Model deployment time is reduced from days to minutes through containerized solutions.
• Scalable Performance: Varying workloads are handled efficiently during peak tourist seasons.

Sukjai is an intelligent tourism assistant powered by a sophisticated three-tiered architecture that combines data processing, model orchestration, and query handling capabilities. The system's architecture is designed to provide accurate, contextual responses to tourist inquiries while maintaining scalability and performance.

The architecture consists of three main components:

1. Data Processing Layer: A foundation layer that begins with a web crawler that systematically collects tourism-related information. The collected data undergoes metadata extraction and is processed through a specialized chunking system that segments information into manageable, context-aware pieces. The processed data is then converted into vectors and stored in a vector database for efficient retrieval.
2. Data Creation and Fine-Tuning Pipeline: The system employs a sophisticated pipeline starting with data ingestion with NVIDIA NeMo Retriever, followed by data curation. Using NVIDIA NeMo Curator synthetic data generation pipelines, the Wipro team created high-quality question-answer pairs that resemble conversations in the Thai language. The Wipro Enterprise Generative AI Framework, or WeGA, is a context ingestion system that processes this curated data, which then undergoes labeling before being fine-tuned using NVIDIA NeMo framework. The final models are evaluated through NeMo Evaluator microservice to ensure optimal performance. Using the synthetic data generated with NeMo Curator, complemented by their proprietary data, the Wipro team achieved improved accuracy of the fine-tuned model.
3. Query Orchestrator: When a user interacts with the chat application, their query goes through multiple processing stages. The Query Orchestrator handles initial processing, generates pre-filters, and constructs context through vector search. It then utilizes NVIDIA NeMo Retriever and NIM microservices, including Llama 3.1, NeMo Retriever reranking, and NeMo Retriever embedding NIM to generate accurate and contextually relevant responses.

This architecture ensures that tourists receive real-time, accurate information while maintaining high performance and scalability through NVIDIA AI Enterprise. The integration of various components allows for seamless handling of multiple languages and complex queries, making it an effective solution for Thailand's diverse tourism needs.

The solution architecture combines NVIDIA's cutting-edge AI technologies to create a sophisticated tourism assistance platform. At its core, the platform utilizes the Llama 3.1 70B model deployed with NVIDIA NIM, which provides powerful natural language processing capabilities for handling diverse tourist inquiries. The model's deployment is optimized through the containerized format of NIM, significantly reducing setup times from days to minutes and streamlining operational efficiency.

For precise information retrieval and contextual understanding, the platform integrates two NeMo Retriever NIM based on Mistral models. These models work in tandem within a custom-developed retrieval-augmented generation (RAG) solution called the WeGA framework. This framework implements a unique chunking algorithm that efficiently segments and manages vast amounts of tourism-related data, ensuring the chatbot delivers highly relevant and contextually accurate responses to user queries.

Leveraging the NVIDIA NeMo framework, the underlying models powering the platform are precisely adapted to the specific requirements of the Thai tourism domain using both parameter-efficient fine-tuning (PEFT) and supervised fine-tuning (SFT) techniques. The entire system runs on a robust infrastructure comprising a Kubernetes cluster powered by NVIDIA H100 GPUs. This scalable architecture ensures consistent performance during peak tourist seasons and enables efficient workload management. This comprehensive technical stack enables continuous learning and adaptation to evolving tourism industry trends while maintaining high performance and accuracy in tourist assistance services.