1. Overview (Project Background)

This project was initiated to establish a
secure LLM usage environment that minimizes the risk of technical information leakage and data learning issues that could arise as generative AI usage expands within the company.

As internal employees utilized public LLMs such as ChatGPT,
concerns were raised that corporate internal data could be leaked externally or used for model training,
and a security-focused approach to generative AI utilization was needed.

Additionally, beyond simple question-and-answer interactions,
through RAG (Retrieval-Augmented Generation) chatbot implementation based on in-house documents and embedding data,
we aimed to create a structure that automatically switches response methods depending on the availability of materials.

• When internal documents exist → RAG-based response

• When internal documents do not exist → LLM Only response

2. Solution (Resolution Approach)

Objective Definition

• Verification of data leakage prevention structure based on security solutions

• Performance and quality comparison and benchmarking of AWS-based LLM compared to GPT-4o

Key Verification Tasks

• Verification of architecture to prevent internal data from being used for external training

• Verification of response quality and accuracy using AWS LLM models

3. Result (Achievements)

Construction of RAG-based Data Processing Pipeline

• Establishment of preprocessing process that converts various types of documents into structures suitable for RAG

• Ensuring search accuracy by vector indexing preprocessed data in AWS OpenSearch

Document Parsing and Indexing Enhancement

• Document content parsing using LLM-based OCR

• Composition of parsed documents into a RAG-usable structure by loading them into VectorDB (OpenSearch)

Chat API Business Logic Implementation

• Intent classification performed upon user query input
  (In-house regulations / ESG / Others)

• Automatic selection of RAG pipeline or LLM Only response path based on classification results

Document Correction Function Verification

• Implementation of correction pipeline for typos and expression errors using LLM

• Verification of document quality improvement possibilities completed

Expected Benefits

RAG-based Chatbot Utilization

• Provision of in-house document RAG chatbot and Web RAG chatbot through AIR Studio

• Support for document management and configuration management functions by repository

• Establishment of chatbot verification system based on expected question-answer sets

Document Correction Automation

• Provision of Streamlit-based UI

• Automatic inspection and correction output of entire document content upon document upload