Back
Key Takeaway
Building an AI window consultation chatbot that immediately utilizes unstructured documents
We built an AI chatbot that searches and utilizes unstructured financial business documents by applying a RAG architecture based on AIR Studio and Amazon Bedrock (Claude), and verified the feasibility of using generative AI in window consultation operations.
Finance (S Company)
Client :Finance (S Company)
Industry :Finance
Service Area :Data & AI
Applied Solution :AIR
1. Overview (Project Background)
This project was initiated with the goal of building an AI Chatbot service to utilize the vast unstructured documents used in credit union window operations more quickly and efficiently.
Various business manuals and unstructured documents in PDF and HWP formats existed across lending, deposit, mutual aid, and digital operations, but there were limitations in quickly finding and delivering necessary information in actual window operations.
Accordingly, we built an AI Chatbot environment where window members can query in natural language and receive answers quickly by applying RAG-based data loading and generative AI response structures.
2. Solution (Solution Approach)
To build the AI Chatbot, we designed a structure combining search and generative AI centered on AIR Studio.
AIR Studio-based Chatbot UI Provision
Implementation of a chatbot interface that allows immediate confirmation of necessary information through natural language queriesOpenSearch & Retriever Configuration
Configuration of OpenSearch utilizing RAG structure within AIR Studio, and application of Retrieval Chain for document search and extractionGenerative AI Integration
Configuration of LLM call structure through LCEL Chain within EC2 environment, and generation of answers based on search results using Claude model from Amazon Bedrock
3. Result (Achievements)
We established a foundation for effectively utilizing various business documents through the construction of an unstructured document-based RAG pipeline.
Configuration of preprocessing and chunking pipeline compatible with PDF and HWP documents
Conversion of text, image, and table data into Document format usable in RAG
Provision of API-based data upload and status confirmation features for operational convenience
Additionally, we applied multi-turn conversation functionality through an LLM response structure combining search documents and prompts, and verified the feasibility of utilizing generative AI in actual window consultation operations.
