Multi-Query Retrieval

Multi-query retrieval is a RAG pattern that hedges against the single-query-phrasing failure mode of standard retrieval. Before hitting the index, the pipeline uses an LLM to generate several paraphrased or reframed versions of the user query — a literal rephrase, a broader abstraction, a narrower specialization, a keyword-only variant. Each version is run through the retriever independently, and the resulting candidate lists are unioned or merged with reciprocal-rank-fusion. The pattern improves recall when the single user phrasing happens to use vocabulary that is far from the corpus — jargon mismatch, indirect wording, questions phrased from a user's mental model rather than the authors'. The cost is one additional LLM call to generate queries and several retrievals per user request, which matters for latency-sensitive surfaces but is inexpensive for knowledge work where retrieval quality dominates.