Introducing Kakaopay Securities' Work Assistant AI Bot, 'ChoonCiri'

Kakaopay Securities' DevOps team developed the AI bot 'ChoonCiri', applying RAG (Retrieval-Augmented Generation) technology to make it easier to look up information scattered across the company and to resolve security issues.

They used Amazon Bedrock to integrate various LLMs, and built an environment that does not store input prompts or answers for security. For the Embedding model, they used Amazon Titan, and for the vector database, PGVector, implementing a search function based on data similarity.

The system is largely divided into two parts.

• Preprocessing part: Scans Confluence documents every hour and stores updated content in the vector DB.
• Service part: Receives user requests via Slack, GitHub, and Jira, and provides answers through the RAG process.

The main features currently provided are as follows.

• Wiki Question: Learns internal documents to search for and answer with accurate information
• Learn: Manually and immediately learns undocumented information
• General Question: Provides general LLM answers as a replacement for ChatGPT
• Summarize: Summarizes long conversation threads

Through this, the goal is to reduce the time new hires spend searching for information and to increase work efficiency.