Transformer-based multistage architectures for code search

Code Search is one of the most common tasks for developers. The open-source software movement and the rise of social media have made this process easier thanks to the vast public software repositories available to everyone and the Q&A sites where individuals can resolve their doubts. However, in the case of poorly documented code that is difficult to search in a repository, or in the case of private enterprise frameworks that are not publicly available, so there is not a community on Q&A sites to answer questions, searching for code snippets to solve doubts or learn how to use an API becomes very complicated. In order to solve this problem, this thesis studies the use of natural language in code retrieval. In particular, it studies transformer-based models, such as Bidirectional Encoder Representations from Transformers (BERT), which are currently state of the art in natural language processing but present high latency in information retrieval tasks. That is why this project proposes a multi-stage architecture that seeks to maintain the performance of standard BERT-based models while reducing the high latency usually associated with the use of this type of framework. Experiments show that this architecture outperforms previous nonBERT-based models by +0.17 on the Top 1 (or Recall@1) metric and reduces latency with inference times 5% of those of standard BERT models