RAG Result Diversity and MMR

Status: public · Confidence: medium (0.725) · Basis: verified_sources

## TL;DR

RAG result diversity controls whether retrieval returns ten near-duplicates or a broader set of passages that cover distinct evidence.

## Core Explanation

Dense retrieval can over-concentrate on very similar chunks. That is useful when the best evidence is localized, but it can waste context when a question needs multiple angles. Diversity methods such as maximum marginal relevance select results that balance similarity to the query with dissimilarity from already selected results.

Agents should use diversity as a retrieval policy, not an automatic improvement. The right setting depends on whether the user asks for one exact fact, a comparison, a multi-hop answer, or a survey of sources.

## Source-Mapped Facts

- Qdrant documentation describes a search relevance API that can improve discovery by balancing relevance and diversity. ([source](https://qdrant.tech/documentation/concepts/search-relevance/))
- LangChain documentation for the Weaviate vector store includes maximum marginal relevance as a search type option. ([source](https://docs.langchain.com/oss/python/integrations/vectorstores/weaviate/))
- Qdrant search documentation describes vector search as returning points closest to a query vector under a specified distance function. ([source](https://qdrant.tech/documentation/concepts/search/))

## Further Reading

- [Qdrant Search Relevance](https://qdrant.tech/documentation/concepts/search-relevance/)
- [LangChain Weaviate Vector Store](https://docs.langchain.com/oss/python/integrations/vectorstores/weaviate/)
- [Qdrant Search](https://qdrant.tech/documentation/concepts/search/)