Perbandingan RAG Konvensional dan Agentic RAG Terhadap Kualitas Jawaban pada Sistem Pencarian Informasi pada Dokumentasi IT

Abstract

Retrieval-Augmented Generation (RAG) is widely adopted in document-based question answering systems to improve factual grounding of large language models (LLMs). However, conventional RAG operates in a single-pass pipeline, limiting its effectiveness in handling complex queries that require multi-step reasoning. Agentic RAG extends this approach by introducing planning, internal evaluation, and tool-based interaction, enabling adaptive and iterative retrieval–generation processes. This study compares conventional RAG and Agentic RAG in the context of Indonesian IT documentation. Evaluation is conducted using RAGAS metrics (faithfulness, answer relevancy, and context precision) along with hallucination assessment. Two language models, llama3.1:8b and qwen3:8b, are employed to ensure that performance differences stem from the RAG architecture rather than model-specific factors. The experiment involves 50 test questions with varying levels of complexity. Results indicate that Agentic RAG consistently outperforms conventional RAG. Using llama3.1:8b, Agentic RAG achieves a faithfulness score of 0.78 and context precision of 0.81, compared to 0.70 and 0.72 in conventional RAG. Similar improvements are observed with qwen3:8b, accompanied by lower hallucination levels. These findings demonstrate that iterative mechanisms in Agentic RAG enhance answer alignment with source documents, particularly for complex IT documentation queries.