As we have seen, retrieval methods are a critical component of RAG systems. They enable the identification and ranking of relevant content for queries, which is the first step in generating useful answers from an LLM. During your journey into RAG application development, you’re likely to encounter two dominant retrieval paradigms – dense retrieval and sparse retrieval. Because it is important to understand these concepts, this section will focus on their characteristics, trade-offs, and the benefits of combining them.

Dense retrieval

The dense retrieval method relies on embedding vectors to represent text in a continuous, high-dimensional space. Using embedding models, texts are encoded into fixed-length numerical vectors that are intended to capture semantic meaning. Queries are also encoded so that the similarity between them and the node vectors can be measured using geometric operations. In dense retrieval...