Reeman Korea Builds a Natural Language-Based Product Search System Using Amazon Bedrock Knowledge Bases

Reeman Korea sought to solve the search failure problem caused by existing keyword matching methods failing to reflect users' search habits centered on ingredients or formulations. The core issue was improving the problem where users who didn't know the exact product name would search for something like "Centella cream" and end up abandoning their purchase.

Through the AWS EBA (Experience-Based Acceleration) program, a team of 5 completed an MVP in 5 weeks. They designed a RAG (Retrieval-Augmented Generation) architecture based on Amazon Bedrock Knowledge Bases to build a natural language search environment.

The key technical decisions were as follows:

• Hybrid Search: Combined semantic search and keyword search to improve accuracy
• No-Chunking embedding strategy: Adopted to prevent data loss
• LLM classifier layer: Added for query expansion and metadata filter extraction
• Anthropic Claude 3.5 Sonnet: Used for natural language response generation

As a result of building the system, search accuracy improved from 72% to 95%, and the average response time was reduced to within 2.5 seconds.