HybridRetriever<T>
Models & Types · AiDotNet.RetrievalAugmentedGeneration.Retrievers
Hybrid retriever combining dense and sparse retrieval via Reciprocal Rank Fusion (RRF).
Inheritance: RetrieverBase<T> → HybridRetriever<T>
How It Works
Dense (cosine, ~0..1) and sparse (BM25, unbounded) scores live on incompatible scales, so a raw
weighted sum is dominated by whichever list has larger-magnitude scores. RRF instead fuses by RANK —
each list contributes weight / (k + rank) — which is scale-free and is the default hybrid
fusion in Elasticsearch/Weaviate/Qdrant and the LangChain/LlamaIndex ensemble retrievers
(Cormack et al., 2009). The weights still let callers favor one retriever; k (default 60) damps the
influence of low ranks.
Constructors
| Constructor | Summary |
|---|---|
HybridRetriever(IRetriever<T> denseRetriever, IRetriever<T> sparseRetriever, double denseWeight, double sparseWeight, int defaultTopK, int rrfK) |