Woowa Gongbang's New Colleague: Building a Chatbot Service with System Context (feat. RAG)

The design system platform Woowa Gongbang provides a consistent user experience through a hierarchical structure of Core, Clay, and Mold. However, as documentation, guides, and the codebase grew larger, users increasingly struggled to quickly find the information they needed.

Previously, the team provided MCP (Model Context Protocol), but it had limitations in accessibility for non-developers as well as constraints in data collection and operational control. To address this, a system was needed that went beyond simple search to understand design intent and component history.

To this end, the team built a RAG-based chatbot service using AWS Bedrock Knowledge Bases. The service consists of a web application and an API server, with chat history managed via DynamoDB.

Key implementation features include:

• Using Amazon OpenSearch Serverless (AOSS) as the vector store to implement sophisticated search with metadata-based filtering
• Configuring a RAG chain with LangChain and providing streaming responses based on SSE (Server-Sent Events)
• Optimizing the Retrieval process to find document fragments that are semantically related to the user's question