A generative AI engineer is building a RAG-based internal knowledge assistant. During evaluation, the team finds that the vector search step retrieves 20 chunks that contain the relevant answer somewhere in the set, but the chunks actually passed to the LLM (top 3 by vector similarity) frequently miss the most relevant passage, causing incomplete answers. The embedding model and chunking strategy are already well-tuned. Which change to the retrieval architecture would most directly improve the relevance of the final chunks sent to the LLM?