rag-fusion

Why we included this project

RAG-Fusion is for teams whose retrieval pipeline keeps missing relevant documents because users phrase queries differently than the corpus is indexed. The repo shows a working technique: have an LLM generate several query variations, run each against vector search, and merge the ranked results with Reciprocal Rank Fusion to recover documents a single phrasing would miss. What sets it apart is that it ships an evaluation harness built on NFCorpus and BEIR, with paired-bootstrap confidence intervals, and the accompanying writeup is honest about where fusion pays off and where it does not. The vector-only variant turns out to be roughly neutral in benefit, while the hybrid that adds BM25 and cross-encoder reranking shows measurable gains in both retrieval quality and generated answers. That gives teams a real basis for deciding whether the extra compute is worth it, instead of a one-size-fits-all promise.