Retrievers

A retriever is the interface between a query and a candidate set of relevant chunks. Every vector store exposes one through as_retriever, which returns a VectorStoreRetriever — a Runnable, so it composes with | like anything else in this reference.

retriever = vector_store.as_retriever(
    search_type="similarity",
    search_kwargs={"k": 4},
)
retriever.invoke("How many distribution centers does Nike have in the US?")

Search strategies

Strategy	What it fixes	Cost
similarity (default)	Plain nearest-neighbour by cosine distance.	Baseline.
mmr (maximum marginal relevance)	Near-duplicate results — the top-k similarity hits are often near-identical chunks; MMR re-ranks for relevance and diversity so results cover more ground.	Slightly more compute per query; usually worth it once your index has redundant chunks.
similarity_score_threshold	Returning irrelevant chunks just to fill k — this drops anything below a similarity cutoff, even if that means returning fewer than k.	Requires tuning a threshold per embedding model.
Multi-query expansion	A single phrasing missing relevant chunks — an LLM generates several reformulations of the query, retrieves for each, and merges results.	One extra LLM call per query.
Contextual compression	Retrieved chunks containing mostly irrelevant text — a compressor (often an LLM) trims each retrieved chunk down to the parts relevant to the query before it reaches the model.	One extra LLM call per retrieved chunk, or a cheaper extractive compressor.

MMR is the one worth reaching for by default once you notice retrieved chunks repeating themselves — set it with search_type="mmr" and tune fetch_k (how many candidates it diversifies over) alongside k.

See also

• Embeddings — what a retriever is searching over.
• RAG Pipeline — wiring a retriever into a full chain.
• Vector Stores Overview — per-store as_retriever differences.