Retrieval Augmented Generation with Synergizing Reasoning and Acting Prompt Engineering for Indonesian Open-Domain Question Answering

Large Language Models (LLMs) offer promising solutions for various downstream tasks. However, hallucination, limited reasoning capabilities, and high resource consumption remain significant challenges. Retrieval-Augmented Generation (RAG) addresses these issues by providing factual context from external sources. A notable application of RAG is in open-domain question answering, which can be implemented in multiple languages, including Bahasa Indonesia. This study investigates RAG in terms of both its retrieval and generation components. A grid search was conducted to compare sparse and dense retrieval methods. SahabatAI-Gemma-9B and its base model, Gemma-2-9B, are utilized as generative backbones. Additionally, the Synergizing Reasoning and Acting (ReAct) method is explored as a prompt engineering technique to enhance the reasoning capabilities of LLMs. The evaluation compares standard prompting with the ReAct approach. The experimental results show that BM25 is the most effective retriever among the tested methods. SahabatAI-Gemma-9B does not significantly outperform its base model. However, incorporating ReAct improves the recall of METEOR and BERTScore by approximately 2% and enables the generation of more relevant answers through structured reasoning. ReAct facilitates a step-by-step reasoning process, allowing the model to locate specific entities more effectively and to determine when to answer or express uncertainty, thus enhancing control over its responses.