Get Diverse Results with MMR

By default, retrieve() returns the top-N most similar chunks — which can be nearly identical if they come from the same page. Use MMR (Maximal Marginal Relevance) to get results spread across different parts of the document.

Quick Setup via Config

retriever:
  search_type: "mmr"
  top_k: 5

from ragwire import RAGWire

rag = RAGWire("config.yaml")
results = rag.retrieve("Apple revenue breakdown")

Fine-Grained Control with mmr_search

from ragwire import mmr_search

results = mmr_search(
    rag.vectorstore,
    query="Apple revenue breakdown",
    k=5,
    fetch_k=20,      # fetch 20 candidates, pick the 5 most diverse
    lambda_mult=0.3, # lean towards diversity (0.0 = max diverse, 1.0 = max relevant)
)

for doc in results:
    print(doc.metadata.get("chunk_index"), doc.page_content[:100])

Parameter Guide

Parameter	Default	Effect
k	5	Number of results to return
fetch_k	20	Candidate pool size. Higher = better diversity coverage
lambda_mult	0.5	0.0 = maximize diversity, 1.0 = maximize relevance

Recommended starting values:

• fetch_k = 3–4x k (e.g. k=5, fetch_k=20)
• lambda_mult = 0.3–0.5 for a balanced diversity/relevance trade-off

When to Use MMR

Use MMR when:

• Your documents have long repetitive sections (boilerplate in contracts, repeated tables in financial reports)
• Similarity search keeps returning chunks from the same page
• You want results that cover different aspects of a topic

Stick with similarity when you need the absolute highest-relevance matches and repetition is not a concern.