Implementasi Modul Chatbot pada Website K3 Menggunakan Metode Prototype

Abstract

Pencarian dokumen regulasi Kesehatan dan Keselamatan Kerja (K3) di PT XYZ masih dilakukan secara manual sehingga menyebabkan keterlambatan akses informasi dan risiko kesalahan interpretasi regulasi. Penelitian ini bertujuan membangun purwarupa website chatbot berbasis Retrieval Augmented Generation (RAG) menggunakan metode Prototyping, dengan Django yang diintegrasikan dengan LangChain, Chroma Vector Database, dan model Qwen3 melalui Ollama. Pengujian fungsionalitas menggunakan Black-Box Testing terhadap 14 skenario fitur mencapai keberhasilan 100%, pengujian kualitas jawaban menggunakan BERTScore pada 10 pertanyaan K3 menghasilkan rata-rata precision 0,7004, recall 0,7378, dan F1-score 0,7177 (kategori efektif), serta User Acceptance Testing (UAT) terhadap 6 responden pada 9 aspek fungsional utama menggunakan skala Likert menghasilkan rata-rata persentase kelayakan 93,33% (kategori sangat baik). Hasil ini menunjukkan sistem mampu menyediakan pencarian informasi K3 berbasis bahasa alami yang transparan, akurat, dan diterima sangat baik oleh pengguna.

Searching for Occupational Health and Safety (OHS/K3) regulatory documents at PT XYZ is still largely performed manually, causing delayed information access and a higher risk of regulatory misinterpretation. This study goals built a website-based chatbot Prototype using the Retrieval-Augmented Generation (RAG) approach, developed with Prototyping model, using Django integrated with LangChain, the Chroma Vector Database, and the Qwen3 model served through Ollama. Functional testing using Black-Box Testing across 14 feature scenarios achieved a 100% success rate, answer-quality testing using BERTScore on 10 OHS test questions produced an average Precision of 0,7004, Recall of 0,7378, and an F1-score of 0,7177 (classified as effective), and User Acceptance Testing (UAT) involving 6 respondents across 9 core functional aspects using a Likert scale produced an average feasibility score of 93,33% (classified as very good). These results show the system delivers transparent, accurate, natural language-based OHS information retrieval that is very well accepted by users.