Abstract:Abstract: The rapid growth of e-commerce in Indonesia has increased consumer interactions with digital platforms, particularly Lazada, Tokopedia, and Blibli, resulting in a large volume of customer reviews that reflect consumer…
onsumer experiences and perceptions but have not been optimally utilized in business decision-making. The main issue addressed in this study is how to process customer review data to generate meaningful information regarding consumer opinions. This research aims to apply web scraping techniques to collect customer review data and conduct sentiment analysis to identify trends in consumer opinions across the three e-commerce platforms. The dataset consists of 3,000 customer reviews, with 1,000 reviews collected from each platform, covering aspects such as shopping experience, service quality, delivery process, and customer satisfaction. The research methodology includes data collection through web scraping, text preprocessing for data cleaning and normalization, sentiment analysis using machine learning approaches, and visualization of sentiment results. The findings indicate differences in the distribution of positive, negative, and neutral sentiments across platforms, reflecting variations in consumer experiences and service strategies. These results demonstrate that sentiment analysis based on customer reviews can serve as strategic input to improve service quality, business performance, and marketing strategies in Indonesia’s e-commerce sector.
Keywords: customer reviews; digital services; e-commerce; sentiment analysis; web scarping
Abstrak: Pertumbuhan pesat e-commerce di Indonesia meningkatkan interaksi konsumen dengan platform digital, khususnya Lazada, Tokopedia, dan Blibli, yang menghasilkan ulasan pelanggan dalam jumlah besar sebagai cerminan pengalaman dan persepsi konsumen, namun belum dimanfaatkan secara optimal dalam pengambilan keputusan bisnis. Permasalahan utama penelitian ini adalah bagaimana mengolah data ulasan tersebut agar dapat memberikan informasi yang bermakna mengenai opini konsumen. Penelitian ini bertujuan menerapkan web scraping untuk mengumpulkan data ulasan pelanggan serta melakukan analisis sentimen guna mengidentifikasi tren opini konsumen pada ketiga platform e-commerce tersebut. Data yang digunakan berjumlah 3.000 ulasan pelanggan, dengan masing-masing platform diwakili oleh 1.000 ulasan yang mencakup pengalaman berbelanja, kualitas layanan, proses pengiriman, dan tingkat kepuasan pelanggan. Metode penelitian meliputi pengambilan data menggunakan web scraping, pra-pemrosesan teks untuk pembersihan dan normalisasi data, analisis sentimen dengan pendekatan pembelajaran mesin, serta visualisasi hasil sentimen. Hasil penelitian menunjukkan adanya perbedaan distribusi sentimen positif, negatif, dan netral pada setiap platform, yang mencerminkan variasi pengalaman konsumen dan strategi layanan. Temuan ini menunjukkan bahwa analisis sentimen berbasis ulasan pelanggan dapat menjadi masukan strategis untuk meningkatkan kualitas layanan, kinerja bisnis, dan strategi pemasaran e-commerce di Indonesia.
Kata kunci: customer reviews; digital services;e-commerce;sentiment analysis;web scarping
Abstract:Abstract: This research is motivated by the problem of building material inventory management at Jaqfar Building Store, which is still done manually and based on subjective estimates. This often results in inaccuracies in…
n determining stock levels, either in the form of overstock or understock, which hinders operational effectiveness. The purpose of this study is to apply the Multiple Linear Regression method to analyze the relationship between incoming stock (X1) and outgoing stock (X2) variables with the ending stock variable (Y) to produce an optimal inventory prediction model. The research methodology used includes collecting historical transaction data for building materials such as cement, ceramics, zinc, plywood, and iron. This web-based prediction system was developed using the PHP programming language and a MySQL database. The analysis results show that the resulting regression model can provide a mathematical picture of future inventory patterns based on historical data. Implementation of this system is expected to assist the management of Jaqfar Building Materials Store in making strategic decisions regarding purchasing and sales in a more measured and efficient manner.
Keyword: building materials; data mining; inventory; multiple linear regression
Abstrak: Penelitian ini dilatarbelakangi oleh permasalahan pengelolaan persediaan bahan bangunan di Toko Bangunan Jaqfar yang masih dilakukan secara manual dan berdasarkan perkiraan subjektif. Hal ini menyebabkan sering terjadinya ketidaktepatan dalam menentukan jumlah stok, baik berupa kelebihan barang (overstock) maupun kekurangan barang (understock) yang menghambat efektivitas operasional. Tujuan dari penelitian ini adalah menerapkan metode Multiple Linear Regression (Regresi Linear Berganda) untuk menganalisis hubungan antara variabel stok masuk (X1) dan stok keluar (X2) terhadap variabel stok akhir (Y) guna menghasilkan model prediksi persediaan yang optimal. Metodologi penelitian yang digunakan mencakup pengumpulan data historis transaksi bahan bangunan seperti semen, keramik, seng, triplek, dan besi. Sistem prediksi ini dikembangkan berbasis web menggunakan bahasa pemrograman PHP dan basis data MySQL. Hasil analisis menunjukkan bahwa model regresi yang dihasilkan mampu memberikan gambaran matematis mengenai pola persediaan di masa mendatang berdasarkan data historis. Implementasi sistem ini diharapkan dapat membantu manajemen Toko Bangunan Jaqfar dalam mengambil keputusan strategis terkait pembelian dan penjualan secara lebih terukur serta efisien.
Kata kunci: bahan bangunan; data mining; persediaan; regresi linear berganda
Abstract:YouTube has become a major platform for public discourse in Indonesia, yet large-scale sentiment analysis of its comments remains challenging due to dynamic content, informal language, and limited labeled data. This study…
y proposes a Selenium–IndoBERT pipeline for sentiment analysis of Indonesian YouTube comments using a pseudo-labeling approach. Data were collected from ten YouTube videos discussing the One Piece flag phenomenon, yielding 10,842 comments after preprocessing. Selenium was employed to extract comments from dynamic pages, while IndoBERT was fine-tuned on a small manually labeled dataset and used to generate pseudo-labels for unlabeled data. Model performance was evaluated using probabilistic metrics, including Coverage, Expected Calibration Error (ECE), and Brier Score. At a confidence threshold of 0.75, 78.5% of comments received pseudo-labels, with an ECE of 0.095 and a Brier Score of 0.174. Manual validation showed substantial agreement with human annotations (Fleiss’ kappa = 0.72). The results indicate that the proposed pipeline enables scalable and reliable sentiment analysis with minimal manual annotation.
Abstract:Abstract: The growing intensity of cyber attacks, marked by rapid, large-scale, automated, and adaptive execution, requires analytical methods that represent the diversity of network environments, including variations in…
target platforms such as IoT, traditional networks, and hybrid infrastructures. This study compares machine learning models for cyber attack classification under heterogeneous environmental conditions and formulates a conceptual optimization framework based on model performance. Four publicly available benchmark datasets were used, namely UNB CIC IoT 2023, UNB CIC IDS-2018, UNSW-NB15, and a Kaggle cyber security attacks dataset, comprising approximately 40,000 to over 3.6 million records and 25 to 80 features across IoT, conventional, and mixed network environments. Random Forest, XGBoost, Multilayer Perceptron, and Transformer were implemented within a unified pipeline involving preprocessing, feature selection, and Bayesian Optimization-based hyperparameter tuning. All models achieved F1-score and Cohen's Kappa above 96%, with XGBoost performing best (97.80%, 97.26%), followed by Random Forest (97.78%, 96.96%) and Transformer (97.44%, 96.82%), while MLP scored lowest (96.74%, 96.00%), a gap below one percentage point. Confusion matrix analysis revealed persistent misclassification in minority and overlapping attack classes, informing a proposed adaptive cyber attack simulation optimization framework.
Keywords: cyber attacks; optimization; machine learning; environmental variability.
Abstrak: Meningkatnya intensitas serangan siber yang berlangsung cepat, masif, otomatis, dan adaptif menuntut pendekatan analitis yang merepresentasikan keragaman lingkungan jaringan, termasuk perbedaan karakteristik platform sasaran seperti Internet of Things (IoT), jaringan konvensional, dan infrastruktur hibrida. Penelitian ini membandingkan model machine learning untuk klasifikasi serangan siber pada kondisi lingkungan heterogen, sekaligus menyusun kerangka optimasi konseptual berdasarkan performa model. Empat dataset benchmark publik digunakan, yaitu UNB CIC IoT 2023, UNB CIC IDS-2018, UNSW-NB15, serta dataset Kaggle cyber security attacks, dengan jumlah data berkisar 40.000 hingga lebih dari 3,6 juta rekaman dan 25 sampai 80 fitur, mewakili lingkungan IoT, konvensional, dan campuran. Random Forest, XGBoost, Multilayer Perceptron, dan Transformer diimplementasikan melalui pipeline terpadu mencakup pra-pemrosesan, seleksi fitur, dan optimasi hyperparameter berbasis Bayesian Optimization. Seluruh model mencapai F1-score dan Cohen's Kappa di atas 96%, dengan XGBoost menunjukkan performa terbaik (97,80%, 97,26%), diikuti Random Forest (97,78%, 96,96%) dan Transformer (97,44%, 96,82%), sementara MLP mencatat skor terendah (96,74%, 96,00%), dengan selisih kurang dari satu poin persentase. Analisis confusion matrix mengungkap misklasifikasi yang konsisten pada kelas minoritas dan serangan dengan karakteristik serupa, yang menjadi dasar kerangka optimasi simulasi serangan siber adaptif yang diusulkan.
Kata kunci: serangan siber; optimasi; machine learning; variabilitas lingkungan
Abstract:This study aims to apply the Analytic Network Process (ANP) method as a decision support tool in determining the eligibility of education grant recipients in North Sumatra Province. The background of this research arises…
from the large number of grant applicants compared to the available budget, as well as the absence of clear and objective evaluation standards. The ANP method was chosen because it allows the interdependence between assessment criteria such as institutional feasibility, performance and achievement, social and educational impact, and accountability and transparency to be analyzed comprehensively. Data were obtained through interviews, documentation, and observation at the North Sumatra Provincial Education Office. The results of the ANP model show that the criterion with the highest weight is accountability and transparency (0.44), followed by social and educational impact (0.31). Among the three alternatives, community-based education foundations (A2) obtained the highest total weight (0.30), indicating that they are the most eligible recipients of education grants. The implementation of the ANP-based decision support system produces valid and consistent ranking results (CR < 0.1), enabling faster, fairer, and more transparent decision-making. Therefore, the ANP method contributes significantly to improving governance, objectivity, and accountability in the distribution of education grants in North Sumatra Province.
Abstract:Abstract: The rapid degradation of mangrove ecosystems threatens coastal biodiversity, shoreline stability, and carbon sequestration capacity, particularly in areas experiencing intense human activity. However, community-based…
-based participatory mangrove monitoring remains limited due to the lack of accessible and user-friendly digital tools. This study aims to design an intuitive mobile application for mangrove tree detection and participatory ecological monitoring using a User-Centered Design (UCD) approach. The research was conducted iteratively through user needs analysis, prototype development, and usability evaluation involving local governments, conservation practitioners, and non-expert users. The proposed application integrates machine learning for automated mangrove recognition with geospatial visualization and real-time feedback to support field-based monitoring. Usability evaluation using the System Usability Scale (SUS) yielded an overall score of 82.3, categorized as excellent usability, indicating high user satisfaction and intuitive interaction. The results demonstrate that integrating UCD and machine learning enhances usability, user engagement, and the accuracy of mangrove documentation under real field conditions. Overall, this study presents a field-ready, user-centered mobile solution that bridges usability engineering and participatory mangrove monitoring as a replicable model for inclusive ecological application development.
Keywords: Carbon sequestration; mangrove monitoring; mobile application; user-centered design; usability evaluation
Abstrak: Degradasi ekosistem mangrove yang semakin cepat mengancam keanekaragaman hayati pesisir, stabilitas garis pantai, dan kapasitas sekuestrasi karbon, terutama di wilayah dengan aktivitas manusia yang intens. Namun, pemantauan mangrove secara partisipatif berbasis komunitas masih terbatas akibat kurangnya perangkat digital yang mudah diakses dan ramah pengguna. Penelitian ini bertujuan merancang aplikasi mobile yang intuitif untuk deteksi pohon mangrove dan pemantauan ekologi partisipatif dengan menggunakan pendekatan User-Centered Design (UCD). Penelitian dilakukan secara iteratif melalui analisis kebutuhan pengguna, pengembangan prototipe, dan evaluasi kegunaan dengan melibatkan pemerintah daerah, praktisi konservasi, serta pengguna non-ahli. Aplikasi yang diusulkan mengintegrasikan pembelajaran mesin untuk pengenalan mangrove secara otomatis dengan visualisasi geospasial dan umpan balik waktu nyata guna mendukung pemantauan di lapangan. Evaluasi kegunaan menggunakan System Usability Scale (SUS) menghasilkan skor keseluruhan sebesar 82,3 yang termasuk dalam kategori kegunaan sangat baik, menunjukkan tingkat kepuasan pengguna yang tinggi dan interaksi yang intuitif. Hasil penelitian menunjukkan bahwa integrasi UCD dan pembelajaran mesin meningkatkan kegunaan, keterlibatan pengguna, serta akurasi dokumentasi mangrove dalam kondisi lapangan. Secara keseluruhan, penelitian ini menyajikan solusi mobile berbasis UCD yang siap digunakan di lapangan dan menjembatani rekayasa kegunaan dengan pemantauan mangrove partisipatif sebagai model replikatif bagi pengembangan aplikasi ekologi yang inklusif.
Kata kunci: Carbon sequestration; mangrove monitoring; mobile application; user-centered design; usability evaluation
Abstract:Abstract: Mental health is an essential aspect of overall well-being, particularly for university students vulnerable to emotional strain. This study aims to identify clusters of student mental health trends using the K-Means…
Means clustering technique. The research involved 60 students from four academic programs at the Faculty of Science and Technology, selected using stratified and cluster sampling techniques. Data were collected using a modified Mental Health Inventory (MHI). The results revealed distinct commonalities among majors: the Statistics program was predominantly defined by the depressed cluster at 53.3%, while Mathematics followed at 40% within the same cluster. In contrast, Biology students predominantly fell under the neu-tral/stable cluster (66.7%), whilst Information Systems students exhibited an even distribution (33.3% per cluster) without a dominant trend. The clustering quality was evaluated using the Silhouette Coefficient, yielding a range of 0.39 to 0.60. Biology (0.60) and Statistics (0.54) exhibited a reasonable structure, but Information Systems (0.39) and Mathematics (0.34) demonstrated a deficient structure. In conclusion, K-Means effectively discerns mental health patterns, providing a data-driven basis for targeted psychological interventions in educational settings.
Keywords: biology; information systems; k-means; mathematics; mental health; silhouette coefficient; statistics
Abstrak: Kesehatan mental merupakan komponen vital dari kesejahteraan total, terutama bagi maha-siswa yang rentan terhadap stres emosional. Penelitian ini bertujuan untuk mengidentifikasi kelompok tren kesehatan mental mahasiswa melalui penerapan metode pengelompokan K-Means. Studi ini mencakup 60 mahasiswa dari empat program studi di Fakultas Sains dan Teknologi, yang dipilih melalui metode pengambilan sampel bertingkat dan kelompok. Data dikumpulkan dengan menggunakan Inventaris Kesehatan Mental (MHI) yang dimodifikasi. Temuan menunjukkan kesamaan yang jelas di antara jurusan: program studi Statistika terutama ditandai oleh kelompok depresi (53,3%), diikuti oleh Matematika dengan 40% dalam kelompok depresi. Sebaliknya, mahasiswa Biologi terutama termasuk dalam kelompok netral/stabil (66,7%), sedangkan mahasiswa Sistem Informasi memiliki distribusi yang merata (33,3% per kelompok) tanpa pola yang dominan. Kualitas pengelompokan dinilai dengan Koefisien Sil-houette, menghasilkan rentang 0,39 hingga 0,60. Biologi (0,60) dan Statistika (0,54) memiliki struktur sedang, sedangkan Sistem Informasi (0,39) dan Matematika (0,34) menunjukkan struktur yang buruk. Kesimpulannya, K-Means secara akurat mengidentifikasi tren kesehatan mental, menawarkan landasan berbasis data untuk terapi psikologis yang ditargetkan di ling-kungan pendidikan.
Kata kunci: biologi; kesehatan mental; K-Means; matematika; silhouette coefficient; sistem in-formasi; statistika
Abstract:Abstract: Understanding students’ emotional conditions is important for evaluating engagement and learning atmosphere in classroom environments. However, conventional evaluation methods are often subjective and difficult…
lt to apply in real time. Therefore, this study proposes a real-time multi-face emotion detection system designed for classroom learning environments. The system integrates a CNN-based Tiny Face Detector for multi-scale face localization with a convolutional neural network to classify seven facial emotions: angry, disgust, fear, happy, sad, surprise, and neutral. Experimental evaluation was conducted using classroom video data under varying lighting conditions, face orientations, partial occlusions, and different numbers of detected faces per frame. The proposed system achieves stable real-time performance with processing speeds ranging from 10–20 FPS, depending on face density. The results show higher recognition performance for expressive emotions, while subtle emotions remain more challenging. Overall classification accuracy reaches above 80% when emotion predictions are aggregated across multiple faces and time windows. These results indicate that the proposed system is suitable for objective analysis of emotional dynamics in classroom environments and supports the deployment of lightweight emotion-aware monitoring systems for educational applications.
Keywords: classroom monitoring; convolutional neural network; facial emotion recognition; multi-face detection; tiny face detector.
Abstrak: Pemahaman terhadap kondisi emosional mahasiswa penting untuk mengevaluasi keterlibatan dan suasana pembelajaran di kelas. Namun, metode evaluasi konvensional umumnya bersifat subjektif dan sulit diterapkan secara real-time. Oleh karena itu, penelitian ini mengusulkan sistem deteksi emosi multi-wajah secara real-time yang dirancang untuk lingkungan pembelajaran di kelas. Sistem mengintegrasikan Tiny Face Detector berbasis CNN untuk pelokalan wajah multi-skala dengan jaringan saraf konvolusional untuk mengklasifikasikan tujuh emosi wajah, yaitu marah, jijik, takut, senang, sedih, terkejut, dan netral. Evaluasi eksperimen dilakukan menggunakan data video kelas dengan variasi kondisi pencahayaan, orientasi wajah, oklusi parsial, serta jumlah wajah yang berbeda dalam satu frame. Sistem menunjukkan kinerja real-time yang stabil dengan kecepatan pemrosesan antara 10–20 FPS, bergantung pada kepadatan wajah. Hasil pengujian menunjukkan kinerja yang lebih baik pada emosi ekspresif, sementara emosi dengan ciri halus lebih menantang untuk dikenali. Akurasi klasifikasi keseluruhan mencapai di atas 80% ketika hasil emosi diagregasi berdasarkan banyak wajah dan interval waktu. Hasil ini menunjukkan bahwa sistem yang diusulkan berpotensi digunakan untuk analisis objektif dinamika emosi di kelas serta mendukung pemantauan lingkungan pembelajaran berbasis kecerdasan buatan.
Kata kunci: pengenalan emosi wajah; deteksi multi-wajah; Tiny Face Detector; jaringan saraf konvolusional; pemantauan kelas.
Abstract:Abstract: The implementation of dress code regulations in university environments is generally still carried out conventionally, requiring significant time and effort and potentially leading to subjective assessments. This…
is study develops an automatic student dress code compliance detection system using computer vision based on the YOLOv8 model. The dataset consists of 1,800 annotated images divided into eight clothing categories, split into 78% training (1,404 images), 14% validation (254 images), and 8% testing (143 images). All images underwent preprocessing and data augmentation before training the YOLOv8 model with an input size of 640×640 pixels for 50 epochs. During testing, the YOLOv8 model achieved an overall performance of Precision 0.844, Recall 0.773, F1-Score 0.802, and mAP@0.5 0.841, and was able to detect clothing objects with good accuracy and stable performance under various image conditions. The system was integrated with a Flask-based backend and a web-based frontend to enable real time detection and compliance classification, with a response time of less than 2 seconds, supporting automatic and consistent identification of student dress code compliance as “Compliant” or “Violation.”
Keywords: compliance detection; computer vision; dress code regulations; real time detection; YOLOv8.
Abstrak: Penerapan aturan berpakaian di lingkungan kampus umumnya masih dilakukan secara konvensional sehingga membutuhkan waktu dan tenaga yang relatif besar serta berpotensi menimbulkan subjektivitas penilaian. Penelitian ini bertujuan mengembangkan sistem pendeteksi kepatuhan berpakaian mahasiswa secara otomatis berbasis visi komputer menggunakan model YOLOv8. Dataset yang digunakan terdiri dari 1.800 citra beranotasi yang terbagi ke dalam 8 kategori pakaian, dengan pembagian data sebesar 78% data latih (1.404 citra), 14% data validasi (254 citra) dan 8% data uji (143 citra). Seluruh citra diproses melalui tahapan pre-processing dan data augmentation, kemudian digunakan untuk melatih model YOLOv8 dengan ukuran input 640×640 piksel selama 50 epoch. Pada tahap pengujian, model mencapai performa keseluruhan dengan Precision 0.844, Recall 0.773, F1-Score 0.802, dan mAP@0.5 0.841, serta mampu mendeteksi objek pakaian dengan akurasi baik dan performa stabil pada berbagai kondisi citra. Sistem kemudian diintegrasikan dengan backend berbasis Flask dan frontend web untuk mendukung proses deteksi waktu nyata dan klasifikasi kepatuhan, dengan waktu respons sistem kurang dari 2 detik, sehingga mampu mengidentifikasi status kepatuhan berpakaian mahasiswa ke dalam kategori “Aman” dan “Melanggar Aturan” secara otomatis dan konsisten.
Kata kunci: aturan berpakaian; deteksi waktu nyata; pendeteksi kepatuhan; visi komputer; YOLOv8.
Abstract:Abstract: This study discusses the optimization of RAG for a FAQ system in the field of information technology product security certification at BSSN. Although LLM generate reliable responses, they often lack up-to-date…
and domain-specific knowledge, which can be addressed through the RAG approach. This research aims to optimize a domain-specific RAG system by improving embedding performance, enhancing prompt robustness, and increasing retrieval accuracy. The research methods consist of three stages. The first stage involves fine-tuning the bge-m3 embedding model and evaluating its performance using MRR, Recall, and AUC. The second stage applies prompt engineering techniques, namely the SRSM and Autodefense, to mitigate direct-injection and escape-character prompt injection attacks. The third stage evaluates the proposed RAG system using Precision, Recall, and F1-Score metrics against four baseline models. The results of research show that the fine-tuned embedding model achieves higher performance than the original model, with MRR@1 and Recall@1 values of 0.80 and an AUC@100 of 0.7023. In addition, the proposed prompt engineering techniques demonstrate robustness against prompt injection attacks, while the overall RAG system attains a perfect Precision, Recall, and F1-Score of 1.00. In conclusion, the proposed approach effectively enhances retrieval accuracy, embedding quality, and system security, resulting in a more reliable RAG-based FAQ system for information technology product security certification.
Keywords: embedding fine-tuning; large language model; prompt engineering; prompt injection mitigation; retrieval-augmented generation
Abstrak: Studi ini membahas optimasi RAG untuk sistem FAQ di bidang sertifikasi keamanan produk teknologi informasi di BSSN. Meskipun LLM menghasilkan respons yang andal, mereka seringkali kurang memiliki pengetahuan terkini dan spesifik domain, yang dapat diatasi melalui pendekatan RAG. Penelitian ini bertujuan untuk mengoptimalkan sistem RAG spesifik domain dengan meningkatkan kinerja embedding, meningkatkan ketahanan prompt dan meningkatkan akurasi pengambilan. Metode penelitian terdiri dari tiga tahap. Tahap pertama melibatkan fine-tuning model embedding bge-m3 dan mengevaluasi kinerjanya menggunakan Mean Reciprocal Rank (MRR), Recall, dan AUC. Tahap kedua menerapkan teknik rekayasa prompt, yaitu Self- SRSM dan Autodefense, untuk mengurangi serangan direct-injection dan escape-character prompt injection. Tahap ketiga mengevaluasi sistem RAG yang diusulkan menggunakan metrik Presisi, Recall, dan F1-Score terhadap empat model dasar. Hasil penelitian menunjukkan bahwa model embedding yang disempurnakan mencapai kinerja yang lebih tinggi daripada model asli, dengan nilai MRR@1 dan Recall@1 sebesar 0,80 dan AUC@100 sebesar 0,7023. Selain itu, teknik rekayasa prompt yang diusulkan menunjukkan ketahanan terhadap serangan injeksi prompt, sementara sistem RAG secara keseluruhan mencapai Presisi, Recall, dan F1-Score sempurna sebesar 1,00. Kesimpulannya, pendekatan yang diusulkan secara efektif meningkatkan akurasi pengambilan, kualitas embedding dan keamanan sistem, menghasilkan sistem FAQ berbasis RAG yang lebih andal untuk sertifikasi keamanan produk teknologi informasi.
Kata kunci: penyempurnaan embedding; model bahasa besar; rekayasa prompt; mitigasi injeksi prompt; retrieval-augmented generation