Nowadays, software source code is mainly stored in software repositories and programming content websites. However, searching for code is troublesome, and, usually software engineers are forced to use conventional search engines which do not evaluate the usefulness of the results. The overall process proves to be highly time consuming and inefficient. To face these issues, software developers prefer recommendation systems that target query analysis and return relevant code examples. That undertaking, though, turned out to be extremely difficult due to syntactical differences between natural language and source code. Many of those systems present weaknesses regarding the form of the input queries (e.g. they do not receive input in natural language) and the quality and performance of their results. In addition, most of these systems utilize simplistic architectures and, as a result, do not capitalize the query and code semantics. A close analysis of the aforementioned problems resulted in the design and implementation of CODEtransformer, a system that improves upon the flaws of many code recommendation systems. Our system ensures data quality by mining code examples from popular GitHub repositories. These data are subject to preprocessing in order to maximize the extracted information. Afterwards, we train a state-of-the-art Neural Network which has the ability to accept natural language as input and perform semantic analysis on the code examples. We, then, construct a vector space consisting of code examples that ensures the best possible temporal response of each search. Ultimately, our system is not only evaluated by its performance compared to similar systems, but also through natural language queries, which are derived from Stack Overflow.