ECTI Transaction on Application Research and Development

Development of a Context-Aware Customer Service Chatbot with Automated Tool Selection Using Large Language Models and LangChain

Prateep Meuntis, Sukumal Kitisin — Fri, 29 Aug 2025 00:00:00 +0700

Customer service organizations often face challenges in managing fragmented information scattered across internal systems and various types of documents. These issues are further compounded by the loss of procedural knowledge resulting from high employee turnover. This study presents the design and development of an automated chatbot system that integrates a large language model (LLM) with two categories of external tools: Retrieval-Augmented Generation (RAG) for extracting information from documents, and APIs for accessing external systems. The system is built on the LangChain framework, which enables agent orchestration and modular integration of tools. The RAG-based tools are designed to implement a hybrid search approach, combining keyword-based and semantic retrieval, with result ranking performed using Reciprocal Rank Fusion to identify the most relevant information. The chatbot is capable of automatically selecting the appropriate tools based on the context of each user query. Evaluation results show that the system achieved a tool call accuracy of 0.88 and an answer correctness rate of 0.80. Additionally, user and staff satisfaction rates were reported at 85.77% and 85.04%, respectively, highlighting the system’s effectiveness in addressing data retrieval complexity and facilitating knowledge integration in service environments.

Design and Implementation of a Web-based Information System for on-site Construction Workers’ Management

Jinnapith Theerachaipaisarn, Praewa Honda, Somchai Numprasertchai — Fri, 29 Aug 2025 00:00:00 +0700

One of the management challenges in the construction industry is managing personnel in large construction sites and and tracking the status of work. Modern construction management has applied technology as a tool to support this operation. This article presents the design and development of a web based information system for on-site construction workers’ management for a private construction company to help supervisors track personnel location and monitor the progress of assigned tasks. The proposed system is developed using the React framework and Node.js, connected to the PostgreSQL and PostGIS databases. Personnel location detection uses a Neo-6m device with an ESP32 as a medium to convert values and transfer data to the web server. The test results showed that the proposed system could work properly as designed. For the evaluation results from the 15 related users, it was found that the system can track personal location and monitor the status of work very efficiently and to the satisfaction of those involved at a very good level.

Forecasting Cash Withdrawals at Service Counters in Savings Cooperative Using Morning Transaction Data and Machine Learning

Pitiwat Thavornkulchai, Sethavidh Gertphol — Fri, 29 Aug 2025 00:00:00 +0700

Efficient cash reserve management is a critical factor in the operations of savings cooperatives, particularly for service counters that must ensure sufficient cash availability for members' transactions. This study proposes the development of machine learning models to forecast daily cash withdrawal for service counters. The approach utilizes machine learning techniques to classify withdrawal amounts into two categories: Not exceeding 3 million THB and more than 3 million THB, aiming to enhance cash management efficiency and reduce cash holding costs. The dataset comprises two years of historical counter withdrawal transactions from a savings cooperative in Thailand. The model employs morning withdrawal amounts and previous-day withdrawals as independent variables for prediction. This study evaluates the performance of Gradient Boosting, Logistic Regression, and Multi-Layer Perceptron (MLP) models and examines the impact of different time intervals on prediction accuracy. The results indicate that the MLP model achieved the highest accuracy of 84.80% when utilizing withdrawal data from 08:00 to 12:00, demonstrating its effectiveness in optimizing cash reserve forecasting.