Hybrid Retrieval with RRF Raises RAG System Precision to 100% in Production

A software developer building a production RAG system called ContextQuery found that standard semantic search alone hit a retrieval precision ceiling of 72%, failing on exact keyword queries and short, specific inputs. To fix this, they combined semantic vector search using NVIDIA NIM embeddings with BM25 keyword-based retrieval, then merged the results using Reciprocal Rank Fusion (RRF). RRF works by scoring each retrieved chunk based on its rank across both retrievers, rewarding chunks that appear consistently in both result sets rather than topping just one. The approach required no additional machine learning models — only a mathematical formula applied on top of the existing retrieval infrastructure. After implementing hybrid retrieval with RRF, the developer reported achieving 100% retrieval precision on their evaluation runs.