MMR
An information retrieval algorithm that reduces redundancy by balancing query relevance against result diversity. It is commonly applied in the vector retrieval stage of RAG systems to improve the coverage and breadth of returned results.
Timeline (last 90 days)
在RAG检索中，普通相似度搜索返回最接近的块；MMR则返回既接近又彼此不同的块，防止重复段落。
MMR通过一个λ参数平衡“相关性”与“多样性”，λ越大多样性权重越高。
A Reddit developer spent several weeks replacing the embedding model, adjusting chunk size from 512 to 1024, tuning chunk overlap, and experimenting with the MMR search algorithm, but saw almost no improvement in retrieval quality.
The MMR (Maximal Marginal Relevance) algorithm operates at the level of semantic similarity and cannot identify business constraints such as a document being temporally inconsistent with the query intent.
All Facts (4)
在RAG检索中，普通相似度搜索返回最接近的块；MMR则返回既接近又彼此不同的块，防止重复段落。
50%UnverifiedMMR通过一个λ参数平衡“相关性”与“多样性”，λ越大多样性权重越高。
50%UnverifiedThe MMR (Maximal Marginal Relevance) algorithm operates at the level of semantic similarity and cannot identify business constraints such as a document being temporally inconsistent with the query intent.
50%UnverifiedA Reddit developer spent several weeks replacing the embedding model, adjusting chunk size from 512 to 1024, tuning chunk overlap, and experimenting with the MMR search algorithm, but saw almost no improvement in retrieval quality.
50%