RAG, You've Heard of It… But How Do You Apply It to Your Service?

RAG (Retrieval-Augmented Generation) is a technology that prevents Hallucination by externally retrieving and utilizing the latest information or domain-specific data that an LLM was not trained on. It automatically finds context relevant to a question and passes it to the LLM, making a general-purpose model behave like a specialized model.

A RAG system generally consists of three pipelines.

• Indexing: Processing and storing documents in a searchable form
• Generation: Retrieving information matching a question and generating an answer
• Evaluation: Measuring and improving the system's accuracy and quality

In a case study building an education operations system, RAG was introduced to integrate scattered information from a database, Google Drive, and Google Calendar. Initially, MCP (Model Context Protocol) was used, but due to issues such as environment constraints and unstable LLM integration, an attempt was made to switch to a method of directly implementing a RAG server.

The overall flow for successfully adopting RAG consists of 6 steps: assessing necessity, analyzing requirements, deciding on a framework, indexing, generation, and evaluation.