Pengembangan Aplikasi Chatbot Layanan Pengetahuan Hukum KUHP Dan KUHAP Berbasis Hybrid Retrieval Augmented Generation

Weinandra, Nicholas Marco (2026) Pengembangan Aplikasi Chatbot Layanan Pengetahuan Hukum KUHP Dan KUHAP Berbasis Hybrid Retrieval Augmented Generation. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5027221042-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Prinsip fiksi hukum mengasumsikan setiap individu dianggap mengetahui peraturan yang berlaku sejak diundangkan, namun akses masyarakat awam terhadap informasi hukum di Indonesia masih terbatas akibat kompleksitas bahasa dan struktur dokumen seperti Kitab Undang-Undang Hukum Pidana (KUHP) dan Kitab Undang-Undang Hukum Acara Pidana (KUHAP). Pendekatan berbasis kata kunci konvensional tidak mampu menangkap semantik pertanyaan pengguna secara menyeluruh, sementara penggunaan Large Language Model (LLM) secara langsung berisiko tinggi menimbulkan halusinasi dalam domain hukum. Penelitian ini mengembangkan aplikasi chatbot layanan pengetahuan hukum berbasis web menggunakan arsitektur Hybrid Retrieval-Augmented Generation (RAG) yang mengombinasikan pencarian semantik berbasis dense retrieval dan pencarian leksikal berbasis BM25 melalui skema Weighted Linear Score Fusion berparameter alpha. Proses pengembangan meliputi ekstraksi dan segmentasi dokumen hukum menggunakan OCR, fine-tuning model embedding SimCSE-IndoBERT-base pada dataset putusan Mahkamah Agung, implementasi hybrid retrieval, serta integrasi dengan LLM Gemini melalui antarmuka chatbot berbasis Next.js. Model SimCSE-IndoBERT-base hasil fine-tuning mencapai MRR@10 sebesar 0,62, meningkat 29,2% dibandingkan model dasarnya. Konfigurasi optimal ditetapkan pada α=0,6, dengan Hybrid RAG secara konsisten mengungguli Vanilla RAG di seluruh metrik dan peningkatan F1@1 sebesar 15%. Evaluasi kualitas jawaban menggunakan kerangka RAGAS menunjukkan Faithfulness sebesar 85,87% dan Hallucination Rate sebesar 14,13%, berbanding 76,81% pada konfigurasi LLM tanpa retrieval, membuktikan sistem menekan potensi halusinasi sebesar 62,68 poin persentase. Pengujian parameter temperature menunjukkan nilai 1,0 menghasilkan kualitas jawaban terbaik sekaligus mempertahankan consistency di angka 0,98. Pengujian usabilitas berbasis TAM terhadap responden mahasiswa hukum menghasilkan persentase kelayakan sebesar 89,6% dalam kategori Baik – Sangat Baik.
========================================================================================================================================
The principle of legal fiction assumes that every individual is deemed to know the law once it is enacted, yet public access to legal information in Indonesia remains limited due to the complexity of language and structure in documents such as the Criminal Code (KUHP) and the Criminal Procedure Code (KUHAP). Conventional keyword-based approaches are unable to fully capture the semantic intent of user queries, while the direct use of Large Language Models (LLMs) carries a high risk of hallucination in the legal domain. This study develops a web-based legal knowledge chatbot using a Hybrid Retrieval-Augmented Generation (RAG) architecture that combines dense semantic search with BM25-based lexical search through a Weighted Linear Score Fusion scheme parameterized by alpha. The development process encompasses legal document extraction and segmentation via OCR, fine-tuning of a SimCSE-IndoBERT-base embedding model on Supreme Court ruling datasets, implementation of hybrid retrieval, and integration with the Gemini LLM through a Next.js-based chatbot interface. The fine-tuned SimCSE-IndoBERT-base model achieves an MRR@10 of 0.62, a 29.2% improvement over its base counterpart. The optimal configuration is established at α=0.6 and K=3, with Hybrid RAG consistently outperforming Vanilla RAG across all metrics and a 15% improvement in F1@1. Answer quality evaluation using the RAGAS framework shows a Faithfulness score of 85.87% and a Hallucination Rate of 14.13%, compared to 76.81% in the LLM-only configuration, demonstrating that the system reduces hallucination potential by 62.68 percentage points. Temperature parameter testing shows that a value of 1.0 produces the best answer quality while maintaining a consistency score of 0.98. Usability testing based on the Technology Acceptance Model (TAM) with law student respondents yields a feasibility score of 89.6%, categorized as Feasible - Highly Feasible.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Retrieval Augmented Generation, Chatbot Hukum, Hybrid Retrieval, KUHP, KUHAP =============================================================== Retrieval Augmented Generation, Law Chatbot, Hybrid Retrieval, KUHP, KUHAP
Subjects:	T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Nicholas Marco Weinandra
Date Deposited:	22 Jul 2026 07:23
Last Modified:	22 Jul 2026 07:23
URI:	http://repository.its.ac.id/id/eprint/136284

Actions (login required)

View Item