Detail Cantuman

Perkembangan Large Language Model menghadapi tantangan halusinasi yang berpotensi merusak kepercayaan pelanggan dalam konteks bisnis dealer mobil. Penelitian ini mengimplementasikan arsitektur Light Retrieval Augmented Generation untuk mengembangkan chatbot dealer mobil berbasis Large Language Model, yang memanfaatkan basis data produk dan dokumentasi resmi secara real-time. LightRAG mengintegrasikan struktur graf pengetahuan dan paradigma two-level retrieval guna meningkatkan akurasi dan efisiensi. Evaluasi menggunakan BERTScore pada 18 sampel pertanyaan menunjukkan bahwa pendekatan Retrieval Augmented Generation konvensional (F1-score 78.7%) mengungguli LightRAG (F1-score 65.7%) dan Native-LLM (F1-score 67.5%) dalam presisi dan kelengkapan jawaban. Hasil mengindikasikan bahwa LightRAG, meski menjanjikan optimasi komputasi, memerlukan penyempurnaan lebih lanjut untuk menyeimbangkan precision dan recall. Implementasi pendekatan Retrieval Augmented Generation konvensional dengan model GPT-4o-mini dan text-embedding-3-small direkomendasikan sebagai solusi optimal untuk chatbot dealer mobil guna memastikan respons akurat dan komprehensif.

The development of Large Language Models faces the challenge of hallucinations that could potentially undermine customer trust in the context of car dealerships. This study implements the Light Retrieval Augmented Generation architecture to develop a car dealership chatbot based on Large Language Models, which utilizes product databases and official documentation in real time. LightRAG integrates a knowledge graph structure and a two-level retrieval paradigm to improve accuracy and efficiency. Evaluation using BERTScore on 18 sample questions shows that the conventional Retrieval Augmented Generation approach (F1-score 78.7%) outperforms LightRAG (F1-score 65.7%) and Native-LLM (F1-score 67.5%) in terms of answer precision and completeness. The results indicate that LightRAG, while promising computational optimization, requires further refinement to balance precision and recall. Implementing the conventional Retrieval Augmented Generation approach with the GPT-4o-mini model and text-embedding-3-small is recommended as the optimal solution for car dealer chatbots to ensure accurate and comprehensive responses.