<![CDATA[The analysis of Indonesian stock markets is inherently complex due to market volatility, heterogeneous financial indicators, and the need to integrate both technical and fundamental information. Traditional analytical approaches often struggle to provide transparent and consistent investment recommendations, while Large Language Models (LLMs) may suffer from hallucination when applied to structured financial prediction tasks. This study proposes a Retrieval-Augmented Generation (RAG) framework based on the LLaMA model and FAISS to enhance the reliability of stock investment recommendations for LQ45-listed companies. Historical OHLCV data and fundamental financial reports are collected from Yahoo Finance and preprocessed into a structured knowledge base using embedding representations. Relevant information is retrieved using a top-k similarity search mechanism and integrated into a structured prompt engineering scheme to generate BUY, HOLD, or SELL recommendations. Experimental results demonstrate that the proposed approach achieves an F1-score of 0.76 for classification performance, with regression-based error metrics of MAE 299.38 and MAPE 10.06%. In addition, text-based evaluation using ROUGE yields a high score of 0.9906, indicating strong alignment between generated explanations and reference analyses. The findings suggest that the RAG framework significantly reduces hallucination by grounding LLM outputs in retrieved financial data, thereby improving the interpretability and robustness of AI-driven stock analysis. This research contributes to the field of informatics by providing a practical and extensible RAG-based framework for structured financial prediction tasks.]]>
