Reciprocal Rank Fusion

Overview

Reciprocal Rank Fusion (RRF) is a simple yet effective method for combining ranked lists from different retrieval systems. It's become the standard approach for fusing results in hybrid RAG systems.

How RRF Works

RRF combines rankings by:

1. Taking ranked lists from multiple systems (e.g., BM25 and dense vectors)
2. Computing reciprocal rank score: 1/(k + rank)
3. Summing scores across systems for each item
4. Re-ranking by combined scores

Parameter k (typically 60) controls score smoothing.

In Production RAG (2026)

Standard implementation pattern:

1. Run dense and BM25 queries in parallel
2. Fuse ranked lists via RRF
3. Create diverse, high-confidence candidate set
4. Apply cross-encoder or ColBERT re-ranking over fused top-k (typically k=50-200)

Advantages

• Simple: No training required
• Effective: Performs well across domains
• Robust: Handles differences in score distributions
• Fast: Lightweight computation
• Proven: Wide industry adoption

Use Cases

• Hybrid search systems
• RAG retrieval pipelines
• Combining keyword and semantic search
• Multi-modal retrieval
• Ensemble retrieval methods

Alternative Methods

• Linear combination
• Weighted fusion
• Learning-to-rank approaches

RRF remains popular due to simplicity and effectiveness.

Implementation

Supported in most RAG frameworks:

• LangChain
• LlamaIndex
• Haystack
• Custom implementations