Fine-tuned small language models (SLMs) have emerged as effective alternatives for closed-domain tasks, where large language models (LLMs) often lack sufficient parametric knowledge. This study presents a methodology for adapting a small language model to a closed-domain question answering (Q A) task. For each question, the model is trained to output an answer based on the most relevant context passage, among ten provided candidates, thus reproducing the logic of a Retrieval-Augmented Generation (RAG) framework. The fine-tuning data were derived from PetroKGraph, an existing knowledge graph built from Portuguese-language resources in the oil and gas (O G) domain. Experimental results show that the fine-tuned model achieves a 20 percentage points accuracy improvement over the base model on closed-domain questions. It also surpasses GPT-4o and GPT-4o Mini by 12 and 25 points, respectively. Moreover, its performance on general-domain tasks remains comparable to that of the base model, indicating that the specialized model effectively learned domain specific knowledge while maintaining general reasoning capabilities.

Retrieval-Augmented Generation (RAG) has emerged as an effective strategy to ground Large Language Models (LLMs) with reliable, real-time information. This paper investigates the trade-off between cost and performance by evaluating 13 LLMs within a RAG pipeline for the Question Answering (Q&A) task under noisy retrieval conditions. We assess four extractive and nine generative models—spanning both open- and closed-source ones of varying sizes—on a journalistic benchmark specifically designed for RAG. By systematically varying the level of noise injected into the retrieved context, we analyze not only which models perform best, but also their robustness to noisy input. Results show that large open-source generative models (approx. 70B parameters) achieve performance and noise tolerance on par with top-tier closed-source models. However, their computational demands limit their practicality in resource-constrained settings. In contrast, medium-sized open-source models (approx. 7B parameters) emerge as a compelling compromise, balancing efficiency, robustness, and accessibility.