Uma ferramenta baseada em Retrieval-Augmented Generation para extração de insights em comentários de vídeos do YouTube

Abstract

This work presents the development of CommentLens, a browser extension that employs the Retrieval-Augmented Generation (RAG) architecture to help users understand the opinions expressed in the comments of YouTube videos. The problem addressed is the information overload caused by the number of unstructured comments associated with popular videos, whose manual reading becomes hardly feasible at large volumes. As a solution, the tool collects a video’s public comments through the platform’s official API, ranks the collected comments according to the retrieval strategy employed, and uses the selected items as context for a large language model to generate a concise answer, accompanied by the comments indicated as its sources. The system implements two retrieval strategies (a lexical one and a semantic one, based on embeddings and cosine similarity) and runs them in a paired fashion over the same input, recording the interactions in a relational database for later comparison. Communication with external services is mediated by a backend that protects the access credentials. As a result, a functional tool was obtained, with automated tests covering the core logic of the backend, along with the infrastructure required for a later comparative analysis between the retrieval strategies. The empirical evaluation of answer quality constitutes future work.