Personal Second Brain

Most recent

See 344 more →

RAG-Fusion - a New Take on Retrieval-Augmented Generation

2 min read

TL;DR

RAG-Fusion generates multiple semantically diverse query variants via an LLM and merges their ranked retrieval lists using Reciprocal Rank Fusion, producing more robust context than single-query RAG.

Summary

RAG-Fusion extends standard RAG by generating multiple semantically diverse query variants from the original user query using an LLM. Each variant retrieves its own ranked document list; the lists are then merged using Reciprocal Rank Fusion (RRF), which promotes documents that rank consistently across multiple perspectives. The fused ranking surfaces more comprehensive and robust context than single-query retrieval, reducing sensitivity to query phrasing and improving downstream QA accuracy.

Key contributions

  1. Proposes RAG-Fusion, combining LLM-based multi-query generation with Reciprocal Rank Fusion to improve retrieval robustness.
  2. Demonstrates that RRF is a parameter-free rank aggregation method that straightforwardly improves any existing retrieval pipeline.
  3. Shows that query diversification reduces brittleness of embedding-based retrieval without requiring thesauri or domain vocabulary.
  4. Establishes that multi-query retrieval outperforms single-query RAG on QA benchmarks with minimal additional latency.

When to cite

  1. When implementing multi-query retrieval strategies to reduce sensitivity to query phrasing in RAG systems.
  2. When applying Reciprocal Rank Fusion as a parameter-free method to aggregate ranked retrieval lists.
  3. When discussing query expansion techniques that use an LLM rather than domain-specific vocabulary resources.
  4. When evaluating the latency-quality tradeoff of multi-query retrieval approaches before deployment.

AI LLM RAG Prompt Engineering Retrieval-Augmented Generation for Large Language Models - A Survey

Graph View

Backlinks