Retrievers

A retriever accepts an unstructured query and returns Document objects. It is broader than a vector store: it can wrap vector search, keyword search, an external API, or a custom fusion pipeline.

Tiny example

retriever = vector_store.as_retriever(
    search_type="mmr",
    search_kwargs={"k": 4, "fetch_k": 20},
)

documents = await retriever.ainvoke(
    "What is the refund window for damaged goods?"
)

LangChain retrievers are Runnables, so the familiar invoke, ainvoke, and batch interfaces apply. Vector stores themselves do not subclass Runnable.

Remember the contract

1. Input: one query.
2. Output: a ranked list of Document objects.
3. Metadata: enough provenance to cite and authorize every result.
4. Evaluation: answer-bearing evidence appears within the chosen k.

Failure note

Increasing k can hide poor recall while filling the prompt with irrelevant text. Apply permission filters before generation, log document IDs and scores, and evaluate retrieval separately from the final answer.

Related

Embeddings and vector stores · Callbacks and tracing · RAG evaluation