Basic similarity search retrieves the most similar documents—but they may be redundant. Advanced retrievers address this: MMR balances similarity and diversity, self-query filters by metadata, parent-document retrieval chunks contextually. This post covers each strategy and when to use them.
Maximum Marginal Relevance (MMR)
MMR returns diverse results: similar to the query but different from each other.
Code Snippet
# MMR retrieval
results = vector_store.max_marginal_relevance_search(
"What is semantic search?",
k=5,
fetch_k=20, # Retrieve more candidates
lambda_mult=0.25 # Balance: 0=diversity, 1=similarity
)
Self-Query Retrieval
Translate user queries into metadata filters + similarity search.
Code Snippet
from langchain.retrievers.self_query.base import SelfQueryRetriever
retriever = SelfQueryRetriever.from_llm_and_db(
llm=llm,
db=vector_store,
document_content_description="Academic papers"
)
# Automatically filters by year=2024, then searches
results = retriever.get_relevant_documents("Papers from 2024 about transformers")
Conclusion
Advanced retrieval improves quality: MMR adds diversity, self-query adds filtering, parent-document retrieval preserves context. Choosing the right retriever depends on your data and queries. Next: evaluating RAG systems with LangSmith to measure faithfulness and relevance.