How to Select the 5 Most Relevant Documents for AI Search

Improving the document retrieval step is critical for the performance of any Retrieval-Augmented Generation (RAG) system. While traditional approaches rely on embedding similarity or keyword search, these methods have limitations. More advanced techniques can be used to optimize for both recall (fetching all relevant documents) and precision (filtering out irrelevant ones). These methods include contextual retrieval which combines semantic and keyword search, using reranker models to improve document ordering, and employing an LLM to directly verify the relevance of each retrieved chunk.