ABSTRAK
Keluhan pelanggan pada media sosial dapat menjadi sumber informasi untuk mengetahui permasalahan pengguna PLN Mobile. Penelitian ini bertujuan merancang AI Agent berbasis Domain Knowledge Graph dan Graph Reasoning untuk mengklasifikasikan keluhan pelanggan PLN Mobile. Data penelitian berasal dari 20.000 tweet pengguna Twitter/X yang melalui proses filtering dan screening hingga diperoleh 55 data keluhan final dalam dua domain, yaitu Feature Design dan Registration, dengan 14 subkategori. Data dibagi menjadi 34 data knowledge, 6 data validation, dan 15 data test. Tahapan penelitian meliputi preprocessing, pemodelan ontologi, entity extraction, ontology mapping, pembangunan Domain Knowledge Graph pada Neo4j, serta implementasi Intent Agent, Knowledge Graph retrieval, dan Reasoning and Decision Agent. Knowledge Graph yang dibangun terdiri atas 277 node. Evaluasi dilakukan melalui empat skenario, yaitu A1 Single LLM, A2 Single LLM + Knowledge Graph, A3 Multi-Agent, dan A4 Multi-Agent + Knowledge Graph. A1 dan A3 memperoleh Accuracy dan Weighted F1-Score sebesar 86,67%, A2 sebesar 80,00% dan 77,44%, sedangkan A4 sebesar 73,33% dan 65,33%. Pada A4, tiga dari empat kesalahan berasal dari retrieval miss dan satu dari decision error. Hasil penelitian menunjukkan bahwa AI Agent berbasis Domain Knowledge Graph berhasil dirancang dan diterapkan.

ABSTRACT
Customer complaints on social media can provide useful information about problems experienced by PLN Mobile users. This study aims to design an AI Agent based on a Domain Knowledge Graph and Graph Reasoning for classifying PLN Mobile customer complaints. The research used 20,000 Twitter/X posts that underwent filtering and screening, resulting in 55 final complaint instances grouped into two domains, Feature Design and Registration, with 14 subcategories. The data were divided into 34 knowledge, 6 validation, and 15 test instances. The research stages included preprocessing, ontology modeling, entity extraction, ontology mapping, construction of a Domain Knowledge Graph in Neo4j, and implementation of an Intent Agent, Knowledge Graph retrieval, and Reasoning and Decision Agent. The resulting Knowledge Graph contained 277 nodes. Evaluation was conducted using four scenarios: A1 Single LLM, A2 Single LLM + Knowledge Graph, A3 Multi-Agent, and A4 Multi-Agent + Knowledge Graph. A1 and A3 achieved an Accuracy and Weighted F1-Score of 86.67%, A2 achieved 80.00% and 77.44%, while A4 achieved 73.33% and 65.33%. In A4, three of four errors were caused by retrieval misses and one by a decision error. The results show that the AI Agent integrated with a Domain Knowledge Graph was successfully designed and implemented.