Reciprocal Rank Fusion (RRF): How It Works and When to Use It

Reciprocal Rank Fusion (RRF) is a rank aggregation algorithm that merges results from multiple retrievers (typically BM25 and vector search) by summing reciprocal rank positions using the formula RRF_score(d) = Σ 1/(k + rank). Unlike score normalization methods, RRF operates on rank positions rather than raw scores, making it robust to mismatched score distributions between retrievers. The default k=60 was established in a 2009 SIGIR paper and has proven durable across benchmarks. The post covers the math behind RRF, a worked example showing why consistent mid-list ranking beats a single-list #1, comparisons with min-max and L2 normalization, and concrete implementation examples for OpenSearch, Elasticsearch, Azure AI Search, MongoDB Atlas, pgvector/PostgreSQL, and Weaviate. Production tuning guidance covers rank window size, weighted RRF variants, evaluation metrics (NDCG@10, Recall@k, MRR), and common pitfalls like deduplication failures and mismatched document IDs.