Source: LangChain MultiQueryRetriever (langchain_classic)
LangChain's reference implementation of multi-query retrieval: an LLM writes 3 rephrasings, each is retrieved, and the unique union is returned.
Contents
- What: library API reference and source code
- Author/publisher: LangChain
- Date: read 2026-11-14 (current
master, packagelangchain_classic) - Original: API reference · source
Summary
MultiQueryRetriever wraps any retriever. from_llm(retriever, llm, prompt, include_original) builds a chain of prompt, LLM and a line-list parser; at query time it generates variants, runs the wrapped retriever once per variant and deduplicates the results. It now lives in the langchain_classic package, the home of pre-1.0 chains and retrievers.
Key claims
- The default prompt asks for 3 alternative versions of the question, one per line, explicitly to work around limits of distance-based similarity search.
- Results are merged as a unique union (deduplicated), not rank-fused.
include_original(defaultFalse) adds the user’s original query to the generated list.- Sync and async paths (
generate_queries/agenerate_queries).