<![CDATA[Abstrak. Pemilihan produk skincare sering menjadi masalah karena pengguna memiliki jenis kulit, kebutuhan, preferensi, dan batasan harga yang berbeda. Pencarian manual melalui marketplace atau media sosial sering tidak terstruktur, kurang personal, dan membutuhkan waktu lama, sedangkan layanan admin tidak selalu mampu merespons pertanyaan pelanggan secara cepat. Penelitian ini bertujuan merancang dan membangun chatbot rekomendasi produk skincare berbasis Retrieval-Augmented Generation (RAG) pada TiDB Serverless. Metode pengembangan sistem menggunakan Extreme Programming (XP) dengan tahapan planning, design, coding, dan testing. Sistem dikembangkan sebagai aplikasi web menggunakan Next.js, React, TypeScript, Drizzle ORM, TiDB Serverless, Ollama, model LLM Qwen2.5:7B, dan embedding BAAI/BGE-M3. Pendekatan hybrid retrieval diterapkan dengan menggabungkan vector search dan structured query agar sistem mampu memahami pertanyaan berbasis bahasa alami sekaligus menangani filter seperti harga. Hasil pengujian black box terhadap 15 skenario menunjukkan seluruh fungsi utama berhasil berjalan, meliputi rekomendasi produk, penolakan pertanyaan di luar topik, filter harga, login admin, pengelolaan produk, inventori, pesanan, chat logs, dan cache response. Sistem ini menunjukkan bahwa integrasi RAG, hybrid retrieval, dan TiDB Serverless dapat mendukung rekomendasi skincare secara lebih interaktif, kontekstual, efisien, dan berbasis data. Abstract. Selecting skincare products is often difficult because users have different skin types, needs, preferences, and price ranges. Manual searching through marketplaces or social media is often unstructured, less personalized, and time-consuming. This study aims to design and build a skincare product recommendation chatbot based on Retrieval-Augmented Generation (RAG) using TiDB Serverless. The system development method used Extreme Programming (XP), consisting of planning, design, coding, and testing stages. The system was developed as a web application using Next.js, React, TypeScript, Drizzle ORM, TiDB Serverless, Ollama, Qwen2.5:7B LLM, and BAAI/BGE-M3 embedding model. A hybrid retrieval approach was implemented by combining vector search and structured query so the system can understand natural language questions while handling filters such as price. Black box testing on 15 scenarios showed that all main functions worked successfully, including product recommendation, out-of-topic rejection, price filtering, admin login, product management, inventory, order management, chat logs, and response caching. The system supports skincare recommendation delivery in a more interactive, contextual, and data-driven manner.]]>
Copyrights © 2026
