Imagine trying to learn a new language with only a handful of phrases. That's the challenge faced by machine translation models when dealing with low-resource languages, where training data is scarce. A new research paper explores a clever technique called "in-context example selection via similarity search" to improve translation quality in these scenarios. Large language models (LLMs), like those powering Google Translate, have the remarkable ability to learn from examples provided right within the prompt. The researchers discovered that by carefully selecting these examples based on their similarity to the sentence being translated, the model's performance improves significantly. This method helps LLMs learn from the limited data more effectively by focusing on relevant examples. The study tested various LLMs and found consistent improvements, especially when translating into low-resource languages like Swahili and Wolof. While similarity search has shown limited benefits for high-resource languages like French and German, where large amounts of training data are already available, this research demonstrates its effectiveness for less common languages. The process is simple: before translating a sentence, similar examples are retrieved from a database, and these examples are then used as a guide for the LLM. The study also introduces a new evaluation metric called Language-Aware COMET (laCOMET) to address issues like empty translations or translations in the wrong language, common problems when dealing with low-resource settings. The research highlights the potential of similarity search in broadening access to machine translation for languages with limited digital resources. While more research is needed to fully understand and refine the method, it's a promising step toward breaking down language barriers and making information accessible to everyone.