With the emergence of large-scale language models (LLMs) such as GPT-3 and ChatGPT, Synology looked for ways to improve the efficiency of its customer service using AI. Instead of directly implementing these models, the company has developed a system based on the Recovery Augmented Generation (RAG) technique. This method allows the company to use its own database, ensuring that responses to customers are accurate, relevant, and comply with internal policies.
“By integrating RAG, we can offer accurate and useful answers to customer queries, while maintaining high standards of privacy and security. This allows us to resolve routine queries efficiently, freeing our support team to focus on more complex cases that require human attention,” explains Steven Liang, Generative AI Application Manager at Synology.
How the Synology AI system works
Synology’s RAG-based support system follows a four-step process designed to ensure data privacy and efficiency:
- Building a complete RAG database: The process begins with the creation of a database from technical service records from the last year, which go through a de-identification process to protect customer privacy. The data is then segmented and indexed for easy retrieval, with semantic embeddings that optimize searches.
Synology Increases Customer Service Efficiency 20x with AI-Powered Support Solution
- Customer needs analysis: Upon receiving a support request, the AI first anonymizes the data and then evaluates the customer’s intent. If the request requires specialized technical support or remote diagnosis, the system refers it to a human consultant. In other cases, AI determines the best response, either by extracting information or suggesting additional customer interaction steps.
- Contextually relevant data retrieval: Through a semantic search, the AI extracts relevant historical cases to construct an accurate answer. Synology is also improving the relevance of its results through a re-ranking model, which recalibrates results based on their similarity.
- Generation of policy-compliant responses, with oversight: AI generates responses based on customer intent, collected data, and Synology guidelines. Additional checks are performed to ensure that no sensitive information is included. If human review is required, the support team evaluates the AI-generated suggestion before sending it to the customer.
